Robert E Goodin - Kai Spiekermann - An Epistemic Theory of Democracy-Oxford University Press, USA (2018)

OUP CORRECTED PROOF – FINAL, 13/03/18, SPi
A N E P I S T E M IC T H E O RY O F D E M O C R AC Y
An Epistemic Theory
of Democracy
R O B E RT E . G O O D I N
and
KA I SP I E K E R M A N N
1
3
Great Clarendon Street, Oxford, OX2 6DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© Robert E. Goodin and Kai Spiekermann 2018
The moral rights of the authors have been asserted
First Edition published in 2018
Impression: 1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2017963789
ISBN 978–0–19–882345–2
Printed and bound by
CPI Group (UK) Ltd, Croydon, CR0 4YY
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
Preface
This is a book in democratic theory, not applied mathematics. Much valuable

technical work has been done on the Condorcet Jury Theorem and related
results. We have benefited enormously from that work in writing our own
book. But this book itself is not a contribution to any highly technical literature.
Although we hope that some of its insights (particularly the more conceptual
ones in Part I of the book) might be of interest to more technically minded
readers, the principal intended audience for this book is the non-technical
reader who is interested in what the practical political upshot of the Condorcet
Jury Theorem might be for the theory and practice of democracy.
In this book, we offer some conjectures but no proofs. A few of the results
reported are analytically derived. But the vast majority are computationally gener-
ated, utilizing a Monte Carlo simulation procedure, described in Appendix A2.
Through those, we strive to get a sense of ‘how these functions behave’ across
a range of scenarios likely to mirror those commonly found in the real political
world. Establishing what would occur across a realistically likely range of cases
is much more important, for the practical political purposes of this book, than
establishing what is necessarily the case across the entire range of possible
cases. For the same sort of reasons, we typically present our results as diagrams
rather than tables. For our purposes, it is more important to convey a general
sense of what is going on across the relevant range rather than to fixate on any
particular point on the curve.
For the purposes of our computational exercises, we typically need to plug in
numbers for some of the crucial variables to enable us to estimate the values
of others. The numbers we plug in are, to some extent, plucked out of thin air.
We offer no grounds for thinking that they are empirically the true values, neces-
sarily. Nonetheless, we hope readers will share our sense that they are plausible
enough values for results based on them to be of genuine political interest.
Books resting on mathematics, however lightly, might naturally be expected
to be somewhat plodding. No doubt some non-technical readers will find
our discussion in some places rather hard going, and alas unavoidably so. But
overall the spirit of the book is meant instead to be ‘playful’. Or, perhaps more
precisely, it is ‘exploratory’. Our aim, more than anything, is simply to ‘see what
happens’ when you vary the many interrelated conditions that might affect the
overall epistemic performance of modern democratic government.
The bulk of this book was drafted well before the unsettling political events
of 2016. Brexit and Trump are indeed worrying results from the point of view
of epistemic democrats. But as our epilogue shows, there are perfectly good
vi Preface
ways of making sense of those results in terms of the Condorcet Jury Theorem,
because there are good reasons to think some of its key assumptions were
violated in those cases. If its assumptions fail to obtain, that does not mean the
Condorcet Jury Theorem and analyses based on it are false. It merely means that
those analyses do not (always) track the real world, and it is important to see
why. The mathematics are as they are, nonetheless, and it is well worth seeing
what they imply for more ordinary democratic politics, even if those politics
will be wracked from time to time by such extraordinary cataclysms as Brexit
and Trump.
Acknowledgements
We have been working on these themes off and on, jointly and separately, for
a dozen years or more. Naturally, we have incurred a great many debts over
that time.
The first is to Christian List. Bob Goodin coauthored a first paper on these
topics with him while Christian was a doctoral student visiting ANU as our
first Harsanyi Fellow. Christian went on to supervise Kai Spiekermann’s own
doctoral dissertation, and to coauthor several related papers with him in turn.
Our thinking on these topics has been sharpened by conversations with him
over many years. Kai Spiekermann’s has been sharpened in similar fashion by
conversations and collaborations with Franz Dietrich, their sometime col-
league at LSE.
We should record, more generally, our gratitude to colleagues at the institu-
tions at which we worked during the gestation of this book. For Bob Goodin
that includes: the School of Philosophy at ANU; the Bioethics Department at
the National Institutes of Health, Bethesda, Maryland; and the Government
Department at the University of Essex. For Kai Spiekermann that includes: the
Philosophy Department at the University of Warwick; the Department of
Government at LSE; and the School of Philosophy at ANU.
Earlier versions of these materials were presented at conferences, workshops
and seminars at: Australian National University, University of Copenhagen,
Wissenschaftskolleg Greifswald, Harvard Law School, London School of
Economics, Trinity College Dublin, University of Maryland, Università degli
Studi Milan, New York University, Princeton University, l’Université Paris-
Sorbonne, University of Turku, the Swedish Collegium for Advanced Study
in Uppsala, Washington University St. Louis, and the APSA meetings in
Washington, DC. We are grateful to those audiences for their helpful comments
and suggestions. We are particularly grateful for advice from Martin Marchman
Andersen, Antoella Besussi, Giulia Bistagnino, Geoff Brennan, Randy Calvert,
Stef Collins, Garrett Cullity, Franz Dietrich, John Dryzek, Lina Eriksson, Dave
Estlund, Greta Favara, Barbara Fried, Archon Fung, Jerry Gaus, David Gauthier,
Charles Girard, Alvin Goldman, Bernie Grofman, Russell Hardin, Clarissa
Hayward, Jeff Howard, Adam Kern, Tony King, Saul Levmore, Skip Lupia,
Klemens Kappel, Dimitri Landa, Jenny Mansbridge, Iain McLean, David
Miller, Nick Miller, Mick Moran, Dennis Mueller, Cara Nine, Bertell Ollman,
Joe Oppenheimer, Fabienne Peter, Philip Pettit, Ryan Pevnick, John Quiggin,
Andrew Rehfeld, Mathias Risse, Don Saari, Theresa Scavenius, Norman Schofield,
Katri Seiberg, Piotr Swistak, Ana Tanasoca, Larry Temkin, Mariam Thalos, Jeremy
Waldron, and Jurgen De Wispelaere.
viii Acknowledgements
At Oxford University Press, we are grateful to Dominic Byatt, first for arran-
ging a pair of insightful referees for the book and subsequently for so efficiently
seeing the book into print. It is always a pleasure to work with him.
This book borrows in places on previously published articles. We are grateful
to their publishers for permission to reuse some of that material, typically in
somewhat different ways, in this book.
Robert E. Goodin and Kai Spiekermann, ‘Epistemic aspects of representative
government’, European Political Science Review, 4 (no. 3) (Nov. 2012), 303–25.
Robert E. Goodin and Kai Spiekermann, ‘Epistemic solidarity as a political
strategy’, Epistème, 12 (no. 4) (Dec. 2015), 439–57.
Kai Spiekermann and Robert E. Goodin, ‘Courts of many minds’, British
Journal of Political Science, 42 (no. 3) (July 2012), 555–72.
Contents
Analytic Table of Contents xi

List of Figures xix
List of Tables xxiii
1. Introduction 1
PA RT I . T H E C O N D O R C E T J U RY T H E O R E M
2. The Classic Framework 17
3. Extensions 23
4. Limitations 37
5. Independence Revisited 67
PA RT I I . E P I S T E M IC E N HA N C E M E N T
6. Improve Individual Competence 85
7. Diversity 96
8. Division of Epistemic Labour 110
9. Discussion and Deliberation 132
PA RT I I I . P O L I T IC A L P R AC T IC E S
10. Respecting Tradition 149
11. Following Leaders 164
12. Taking Cues 178
13. Pluralism: Differing Values and Priorities 195
14. Factionalism: Differing Interests 208
PA RT I V. S T RU C T U R E S O F G OV E R N M E N T
15. Epistocracy or Democracy 225
16. Direct versus Representative Democracy 244
17. Institutional Hindrances to Epistemic Success 260
18. Institutional Aids to Epistemic Success 288
x Contents
PA RT V. C O N C LU SIO N S
19. The Relation between Truth and Politics, Once Again 303
20. Headline Findings, Central Implications 312
21. Epilogue: What about Trump and Brexit? 322
A P P E N D IC E S
A1 Key to Notations 369
A2 Estimating Group Competence by Monte Carlo Simulation 371
References 373
Index 425
Analytic Table of Contents
List of Figures xix

List of Tables xxiii
1. Introduction 1
1.1 Epistemic Competence: A Minimal Requirement
of Good Government 1
1.2 Ask Around 3
1.3 Pooling Information and Judgements Based on It 4
1.4 The Condorcet Jury Theorem, in Brief 5
1.5 Extending the Condorcet Jury Theorem 7
1.6 Forestalling Familiar Objections 8
1.6.1 Competence 9
1.6.2 Independence 10
1.6.3 The Truth Value of Values 11
1.7 The Structure of the Book 12
PA RT I . T H E C O N D O R C E T J U RY T H E O R E M
2. The Classic Framework 17
2.1 The Setup 17
2.2 The Assumptions 17
2.2.1 The Competence Assumption 17
2.2.2 The Independence Assumption 18
2.2.3 The Sincerity Assumption 19
2.3 The Theorem 19
2.4 The Calculation 20
2.5 How Quickly Group Competence Converges to Perfection 21
3. Extensions 23
3.1 Weakening the Competence Assumption 23
3.1.1 Mean Competence among Heterogeneous Voters 23
3.1.2 Topic-specific Competence 25
3.2 Extending the CJT to More than Two Alternatives 26
3.3 Using Different Decision Rules 31
3.3.1 A Bayesian Parallel 32
3.3.2 Other Decision Rules 33
4. Limitations 37
4.1 Facts and Values 38
4.1.1 Moral Realism 38
4.1.2 Moral Conventionalism 40
xii Analytic Table of Contents
4.1.3 Moral Separability 40

4.1.4 Moral Majoritarianism 41
4.2 There Might Be No (Single) Truth among the Alternatives 42
4.2.1 The Question Is Not Truth Apt 43
4.2.2 The Right Answer Is Not on the Agenda 43
4.2.3 There Are Multiple Truths 44
4.3 Failures of Sincerity: Strategic Voting 45
4.3.1 Beliefs versus Preferences 46
4.3.2 Game Theoretic Complexities 47
4.3.3 Sincerity as a Default 49
4.3.4 Deliberation Induces Sincerity 50
4.4 Failures of Competence 50
4.5 Failures of Independence 54
4.5.1 Worries over Independence 54
4.5.2 The Relation between Competence and Independence 55
4.5.3 Living with Dependence 60
4.6 Limitations Arising from the Choice Situation 62
4.6.1 The Choice Situation Is Systematically Misleading 62
4.6.2 The Options Are Ill Formulated 63
5. Independence Revisited 67
5.1 Understanding the Independence Assumption 67
5.1.1 Independence Is Not Absence of Interaction 68
5.1.2 Independence Is Not Unconditional 69
5.1.3 Independence Conditional on the State of the World 69
5.2 Dealing with Dependence by Further Conditionalizing 70
5.2.1 Independence Conditional on the Available Evidence 71
5.2.2 Independence Conditional on an Opinion Leader’s
Interpretation of the Evidence 73
5.2.3 Independence Conditional on All Common Causes 74
5.3 The Best Responder Corollary 76
5.4 Epistemic Implications of the Best Responder Corollary 79
5.4.1 A Single Common Cause 79
5.4.2 The Evidence-Limited Case 80
5.4.3 Multiple Common Causes Cases 80
5.5 Some People Are in Better Decision Situations than Others 81
PA RT I I . E P I S T E M IC E N HA N C E M E N T
6. Improve Individual Competence 85
6.1 Don’t Worry 86
6.1.1 Voter Ignorance 86
6.1.2 Knowledge ≠ Competence 91
6.2 Priorities in Improving Voter Competence 92
6.3 What More Can Be Done to Improve Individual Competence? 94
Analytic Table of Contents xiii
7. Diversity 96
7.1 Clones and Common Causes 98
7.2 Negatively Correlated Votes 100
7.2.1 Epistemic Benefits of Negatively Correlated Votes 100
7.2.2 Are Negatively Correlated Votes Plausible? 103
7.2.3 Epistemic Benefits of Diverse Cognitive Models of the World 104
7.3 Engineering Diversity 107
8. Division of Epistemic Labour 110
8.1 Localized Search 111
8.1.1 Incomplete Agendas 112
8.1.2 Diversified (or Many Random) Search Parties as a Solution 114
8.1.3 Transition Costs 116
8.1.4 Recognizing the Best When We See It 119
8.2 Narrowing the Focus 120
8.2.1 Letting Individual Competence Vary with the
Number of Options 121
8.2.2 Considering Options a Few at a Time 123
8.2.3 Subgroups Propose, Whole Groups Dispose 124
8.2.4 Experts Propose, Whole Groups Dispose 126
8.3 Devolving Control over Some Dimensions 128
9. Discussion and Deliberation 132
9.1 The Ideal and Practice of Deliberation 133
9.2 The Many Benefits of Deliberation 134
9.2.1 Deliberation to Increase Individual Competence 135
9.2.2 Deliberation to Reduce Dependence 138
9.2.3 Deliberation Induces Sincerity 140
9.2.4 Deliberation to Improve Best Responder Performance 141
9.2.5 Deliberation Can Change the Decision Problem 142
9.3 The Deliberation Effect 144
PA RT I I I . P O L I T IC A L P R AC T IC E S
10. Respecting Tradition 149
10.1 Traditionalism in Practice: Precedent in the Courts 150
10.2 The Epistemic Costs of Complete Deference
to Previous Decisions 151
10.3 Modelling Partial Deference 153
10.4 Solutions 156
10.4.1 Hiding Precedents 156
10.4.2 Resisting Precedent: Stubborn Judges 157
10.4.3 Discerning Traditionalists: Picking Informative Precedents 159
10.4.4 Lots of Precedents to Choose Among 161
10.5 Beyond Traditionalism in the Courts 162
xiv Analytic Table of Contents
11. Following Leaders 164

11.1 A Single Opinion Leader 165
11.2 Multiple Correlated Opinion Leaders, but Some
Independent Voters 168
11.2.1 Positively Correlated Opinion Leaders 169
11.2.2 Polarization: Negatively Correlated Opinion Leaders 170
11.3 Everyone Partially Follows Uncorrelated Opinion Leaders 172
11.3.1 Opinion Leaders of Purely Random Competence 173
11.3.2 Competent Opinion Leaders 174
11.4 Many Multiply Mediated Opinion Leaders 175
12. Taking Cues 178
12.1 Cue-taking and Low-information Rationality 178
12.2 How Effective is Cue-taking? 180
12.2.1 Experimental Evidence 180
12.2.2 Evidence from Large-Scale Surveys 181
12.3 Cautionary Tales 183
12.3.1 Some Cues Might Be Unreliable 183
12.3.2 Some Cues Might Not Be Chosen for Their Reliability 184
12.4 The Epistemic Effects of Cue-Taking: Two Models 185
12.5 Calculating the Potential Epistemic Effects of Cue-taking 188
12.5.1 Baseline Calculation 189
12.5.2 Sensitivity to the Number and Reliability of
Independent Cues 190
12.5.3 Relative Insensitivity to the Rate of Cue Use 191
12.5.4 Cue-taking with Varying Individual Voter Competence 191
12.5.5 Cueing Incompetents Only 193
13. Pluralism: Differing Values and Priorities 195
13.1 Differing Values 197
13.1.1 Baseline Scenario 197
13.1.2 Six Variations 198
13.1.3 The Democratic Upshot 205
13.2 Differing Priorities 205
13.3 Democratic Competition over Values and Priorities 206
14. Factionalism: Differing Interests 208
14.1 A Factional Interpretation of the CJT 209
14.1.1 Uniform Voter Competence 209
14.1.2 Unequal Factional Competence 211
14.2 Epistemic Solidarity and Block Voting 213
14.3 Who’s with Us? 216
14.3.1 Differential Abstention from Epistemic Solidarity 217
14.3.2 Differential Group Selection Competence 217
14.3.3 Strategic Leadership and Coordination 220
Analytic Table of Contents xv
PA RT I V. S T RU C T U R E S O F G OV E R N M E N T
15. Epistocracy or Democracy 225
15.1 Beating the Smartest Guy in Town 227
15.1.1 Beating the Smartest Single Guy 227
15.1.2 Beating the Smartest Clique of Guys 228
15.2 Modelling the Epistemic Effects of Expanding the Electorate 229
15.2.1 Each Individual’s Competence Level Is Known 229
15.2.2 Only Average Individual Competence Is Known 230
15.2.3 Enfranchising Batches of Voters with Heterogeneous
Individual Competence 230
15.2.4 Enfranchising Voters with Heterogeneous Knowledge Bases 231
15.3 The Epistemic Logic of Enfranchising the Less Competent 231
15.4 Competence-Weighted Voting Rules 233
15.5 Epistemic Considerations beyond Competence 235
15.5.1 Other Ways Smaller Groups Might Outperform Larger Ones 235
15.5.2 Other Ways Larger Groups Might Outperform Smaller Ones 237
15.5.3 The Political Upshot 238
15.6 The Differential Benefits of Learning from Experience 239
15.6.1 The Classical Argument of Participatory Democrats 239
15.6.2 Improving Already Competent Voters 240
15.6.3 Rendering Initially Incompetent Voters Competent 242
15.6.4 The Political Upshot of Learning from Experience 242
16. Direct versus Representative Democracy 244
16.1 How Can a Smaller Group of Representatives Be Better
than a Larger Group of Voters? 245
16.1.1 Incompetent Masses Choosing Competent Representatives 246
16.1.2 Competent Voters Choosing Even-More-Competent
Representatives: The Selection Effect 248
16.1.3 Epistemic Benefits of Smaller Groups: The Deliberation Effect 251
16.2 Delegate versus Trustee Representatives 254
16.2.1 Delegate-style Representation: The Epistemic Costs of
Bunching Voters into Constituencies 255
16.2.2 Trustee-style Representation: The Deliberation Effect Again 256
16.2.3 Mixed Assemblies with Both Delegates and Trustees 257
17. Institutional Hindrances to Epistemic Success 260
17.1 Strong Leaders 261
17.1.1 Party Leaders Dictating Party Policy 262
17.1.2 Mitigating Factors: Many Independent Leaders 263
17.2 Small Upper Chambers or Committees 264
17.2.1 Legislative Committees as Epistemic Bottlenecks 265
17.2.2 Smaller Upper Houses as Epistemic Bottlenecks 266
17.2.3 Mitigating Factors: The Selection and Deliberation Effects 269
17.3 Party Whips and Small, Pivotal Parties in Coalitions 272
17.3.1 Party Whips in the Legislature 272
xvi Analytic Table of Contents
17.3.2 Coalition Government with Small, Pivotal Parties 275

17.4 Presidential Vetoes and Supermajority Rules 276
17.4.1 Supermajority Rules 277
17.4.2 Presidential Vetoes 281
17.4.3 Political Cooling-off Periods 283
18. Institutional Aids to Epistemic Success 288
18.1 Mechanisms to Make Decision Situations More
Truth Conducive 288
18.1.1 Finding New, Better Alternatives 288
18.1.2 Weeding Out Bad, Confusing Alternatives 290
18.1.3 Improve the Evidence Base 293
18.2 Mechanisms to Increase Independence 294
18.2.1 Restricted Franchise, Secret Ballot 295
18.2.2 Public Funding of Elections and Public Broadcasting 296
18.2.3 Proliferate Independent Opinion Leaders 297
18.3 Mechanisms to Increase Competence 298
18.3.1 Increasing Individual Competence 298
18.3.2 Increasing Collective Competence 298
18.4 Mechanisms to Increase Sincerity 299
PA RT V. C O N C LU SIO N S
19. The Relation between Truth and Politics, Once Again 303
19.1 The Limitations of Truth-seeking in Politics 303
19.2 Is Pursuit of the Truth Dangerous in Politics? 305
19.3 Are There Matters That Should Not Be Put to a Vote? 307
19.4 Who Should Decide What Is True? 308
20. Headline Findings, Central Implications 312
20.1 Headline Findings 314
20.1.1 As Good As It Gets, As Bad As It (Probably) Gets 314
20.1.2 All’s Well So Long As There Are Sufficient, Numerous,
Competent, Independent Influences at Work Somewhere 315
20.1.3 There Are Ways of Coping with Incompetent Voters 316
20.1.4 The Case for Large Groups 316
20.1.5 Smaller Groups to Deliberate and Winnow the Options 317
20.1.6 The Decision Situation Is Crucial 318
20.2 Central Implications for Political Practice 319
20.2.1 Avoid Epistemic Deference 319
20.2.2 Pluralism Is Good 319
20.2.3 More High-Quality Evidence Is Good 320
20.2.4 Small-Scale Deliberative Conclaves to Advise the
Electorate Are Good 320
20.3 Getting It Right Matters 321
Analytic Table of Contents xvii
21. Epilogue: What about Trump and Brexit? 322

21.1 The Political Lies of 2016 325
21.1.1 Brexit Lies 325
21.1.2 Trump Lies 327
21.1.3 How Lies Undercut the CJT 330
21.2 In the US Anyway, the Big Liar Actually Lost 331
21.3 Sending a Strategic Signal 332
21.4 Differing Priorities 335
21.5 Opinion Leaders Lied, and Voters Believed Them 339
21.6 Affective Explanations 341
21.6.1 Expressing Emotions 342
21.6.2 Expressing Identity 343
21.6.3 Having Fun 345
21.7 Epistemic Insouciance 347
21.7.1 Voters Were Indifferent on the Topics of the Lies 349
21.7.2 True Fictions 350
21.7.3 Actions, Not Words, Are What Matter 351
21.8 Everyone on Facebook Agrees with Me 354
21.9 Epistemic Malevolence 358
21.10 Epistemic Agnosticism 365
21.11 Conclusion: Epistemic Democracy under Threat 367
A P P E N D IC E S
A1 Key to Notations 369
A2 Estimating Group Competence by Monte Carlo Simulation 371
References 373
Index 425
List of Figures
2.1 Group competence Pn for different levels of pc. 22

3.1 Probability that the majority of voters, each pc = 0.51 competent (and
voting with equiprobability for all incorrect alternatives), will
vote for the correct alternative, with 2, 3, 4, and 5 alternatives. 30
3.2 Probability that the plurality of voters, each pc = (1/k) + 0.01 competent
and equiprobable to vote for false alternatives, will vote for the correct
alternative. 31
3.3 Comparing group competence for three alternatives, voters with
pc = 0.34 and equiprobable votes for the other two alternatives,
with decision taken by Borda Count, Condorcet Pairwise Criterion,
and Plurality Vote. 36
4.1 Probability that the majority of voters is correct, if voters are individually
pc=0.45 likely to be correct, for electorates of varying sizes. 51
4.2 Probability that the majority of 200 voters is correct, depending on
whether voters vote independently (cases I and II) or are certain to
follow an opinion leader (cases III and IV) who is 0.40 likely to be
correct or who is 0.60 likely to be correct, for electorates of varying
levels of homogeneous individual competence. 59
5.1 Independence conditional on the State of the world. 69
5.2 Direct violation of independence conditional on the State of
the world. 70
5.3 Indirect violation of independence conditional on the State of the
world. 71
5.4 Independence conditional on the available Evidence about the State
of the world. 72
5.5 Independence conditional on the Opinion Leader. 74
5.6 Violation of Independence due to multiple common causes. 75
5.7 Voters are influenced by both Evidence and a common cause. 76
5.8 Convergence of probability of the majority among n voters being
correct, for various values of the probability that the decision situation
is truth-conducive (ω), shown with homogeneous pBR = 0.55 . 79
5.9 Experts versus laymen facing different decision situations. 82
6.1 Different distributions of individual competence and the resulting
group competences. 93
7.1 One common cause and many common causes. 99
7.2 Several common causes and direct access to Evidence. 100
xx List of Figures
8.1 Rugged policy landscape. 113

8.2 Search on 1,000 patches, probability of finding correct patch
depending on numbers of search parties. 116
8.3 Rugged policy landscape and inertia-induced equilibria. 118
8.4 Population and specialist votes in comparison. 130
10.1 Differing points at which deference might set in. 152
10.2 Probability of correct majority decision from a court with nine
members, each pc*= 0.55 likely to be individually correct, with
weight w = 1. 156
10.3 Probability of correct majority decision from a court with nine
members, each pc*= 0.55 likely to be individually correct. 158
10.4 Discerning traditionalists. 160
10.5 Probability of correct majority decision from a heterogeneous
nine-member court (each judge pc*= 0.55 likely to be individually
correct), where four judges vote on the basis of their private signal
and the rest decide by weighting on the basis of their own private
signal w = 1 and only take into account informative votes. 161
11.1 Multiple votes influenced by the same Opinion Leader, either
without (a) or with (b) direct influence of Evidence on Votes. 166
11.2 Probability of correct majority decision among voters with
individual competence pc* = 0.55, given a single opinion leader of
competence pOL = 0.55 followed with probability π. 167
11.3 Probability of correct majority decision among voters with
individual competence pc* = 0.55, given a single opinion leader of
competence pOL = 0.4 followed with probability π. 168
11.4 Perfectly positively correlated Opinion Leaders and some
independent Votes. 170
11.5 Negatively correlated Opinion Leaders and independent Votes. 171
11.6 Many multiply mediated Opinion Leaders, direct links between
Evidence and LOLs and Evidence and Votes are omitted. 176
12.1 One Cue as the only access to the Evidence for all voters. 186
12.2 Several Cues as well as direct access to the Evidence. 187
12.3 Probability of majority voting for correct alternative, with varying
levels of individual competence pc* and probability of being guided
by any given cue π, for 990 voters, nine cues, and probability of any
given cue being correct of pK = 0.70. 192
13.1 Baseline scenario with two groups of equal size and all pcV = 0.55. 198
13.2 60% of voters with value V1, 40% with V2 and with all pcV = 0.55. 199
13.3 Two groups of equal size, but voters with V1 have pcV1 = 0.8 while
voters with V2 have pcV2 = 0.55. 200
List of Figures xxi
13.4 60% of voters subscribing to value V1 with pcV1 = 0.8 and 40% of
voters subscribing to value V2 with pcV2 = 0.55. 201
13.5 60% of voters subscribing to value V1 with pcV1 = 0.55 and 40% of
voters subscribing to value V2 with pcV2 = 0.8. 202
13.6 Five equally large groups, but two groups support X1. 203
13.7 Five equally large groups, four of which systematically err in the
same direction. 204
14.1 Probability of victory for the majority faction from majority voting,
as population size increases. 210
14.2 Approximate expected vote distribution, E = 200,000, M = 1,000,000;
pcE = 0.7, pcM = 0.51. 215
14.3 Probability of Mass majorities as a function of group selection
competence. 218
14.4 Probability of Mass majorities as a function of the Elite group selection
competence, Mass group selection competence fixed. 219
15.1 Isocompetence curve showing points at which a group with
nREST and pcREST has the same epistemic performance as a
group with nSMART = 100 and pcSMART = 0.7. 228
15.2 Group competence as a function of n, given that the first fifty voters
have pc= 0.6, while all others have pc = 0.52. 232
15.3 Group competence as a function of n, given that the first fifty
voters have pc = 0.6, while all others have pc = 0.52., shown with
unweighted (equal) votes and with Grofman and Shapley’s weighted
voting rule. 235
15.4 Learning by experience among twenty voters with an initial
pcEXPERT = 0.6 and 1,000 voters with an initial pcLAY = 0.501,
with the competence of each increasing by 1% in each round. 241
15.5 Learning by experience with twenty voters with an initial
pcEXPERT = 0.6 and 1,000 voters with an initial pcLAY = 0.49,
with the competence of each increasing by 1% in each round. 243
16.1 The epistemic competence of an assembly mixing delegates and
trustees, with the number of delegates among ninety-nine
representatives on the x-axis and group competence on the y-axis. 258
List of Tables
4.1 Interaction of the Competence and Independence Assumptions. 56

6.1 US presidential election outcomes if all voters voted the same
way as informed voters of the same demographic. 90
7.1 Independent and negatively correlated votes. 102
7.2 Voters with partial information. 106
8.1 Group competence as a function of declining individual
competence with many alternatives. 122
8.2 Voting on multiple alternatives, either with three experts selecting
their top two options and 303 voters voting in the run-off, or a
direct plurality vote among population with voter competence as
in Table 8.1. 127
8.3 Alternatives and voting competence. 129
11.1 Probability of correct majority decision among 1,000 voters with
individual competence pc* = 0.55 split evenly among multiple opinion
leaders (pOL = 0.50), each voter following his respective opinion
leader with probability π. 173
11.2 Probability of correct majority decision among 1,000 voters with
individual competence pc* = 0.55 split evenly among multiple
opinion leaders (pOL = 0.55), each voter following his respective
opinion leader with probability π. 174
11.3 Probability that a majority of voters will be correct if they follow,
to varying degrees, three opinion leaders each with competence
pOL = 0.55, uninfluenced voters being competent with pc* = 0.55. 175
11.4 Probability of correct majority decision among 990 voters with
opinion leaders (pOL = 0.55 when not following another opinion
leader), each voter following his respective Local Opinion Leader
(and each Local Opinion Leader his respective Big Opinion Leader)
with probability π. 177
14.1 Competence threshold Mass voters have to exceed to make true
Mass interest more likely to win than Elite interest, for various
values of pcE and E/M. 212
16.1 Estimated necessary individual competence of representatives
to make their collective decision epistemically equal to that of
the electorate or no more than 1 percentage point worse (assuming
voters are individually pc = 0.51 competent). 252
xxiv List of Tables
17.1 The collective competence of assemblies divided into similarly

sized parties with party leaders dictating party policy (where the
probability of the party leader being correct is pOL = 0.550). 263
17.2 Probability of correct majority decision among 1,000 MPs with
party leaders (pOL = 0.55), each voter following his respective
party leader with probability π = 0.5. 264
17.3 Probability of a majority vote for the correct outcome in committees
and whole legislatures, of varying sizes (pc = 0.55). 266
17.4 The probability of a correct decision from Congress (assuming
each legislator is pc = 0.55 competent). 269
17.5 The collective competence of assemblies divided into strictly
whipped parties with party policy chosen by majority vote of
that party’s MPs (for a 603-member legislature, each MP
having individual competence of pc = 0.55). 275
17.6 The collective competence of assemblies divided into strictly
whipped parties with party policy chosen by majority vote of that
party’s MPs, with a pivotal small party (for a 603-member legislature,
each MP having individual competence of pc = 0.55). 276
21.1 Fact-checking Trump’s lies. 328
A2.1 Sources of numerical data in tables and figures. 371
In general, a law which has not been voted unanimously involves subjecting
men to an opinion which is not their own, or to a decision they believe
contrary to their interest. It follows that a very great probability of the
truth of this decision is the only reasonable and just grounds according
to which one can demand such submission.
—Condorcet
Introduction
1.1 EPISTEMIC COMPETENCE: A MINIMAL

REQUIREMENT OF GO OD GOVERNMENT
Many factors feed into our assessment of the quality of government.1 But
sheer competence—governments knowing what they are doing, knowing how
to achieve what they attempt—surely must be high on that list. Though sheer
competence is not the only thing on that list,2 without at least minimal compe-
tence good government is impossible.
The perception of incompetence, much more than anything to do with the
war in Vietnam, is what led to George McGovern’s landslide defeat by Richard
Nixon in 1972.3 The sheer incompetence of the White House’s handling of the
bungled Watergate burglary—much more than concerns that the president was
a ‘crook’—is what led to the unravelling of Nixon’s presidency two years later.4
What forced Margaret Thatcher’s resignation as British Prime Minister was her
stubborn insistence upon persevering with a poll tax that was not only inequit-
able, but, more fundamentally, simply unworkable.5 When survey researchers
1 See for example the array of variables in the Quality of Government dataset <http://www.
qog.pol.gu.se/>.
2 ‘Running for president on a promise to be competent and honest is thin gruel’, Greider (1988)
presciently commented at the start of Michael Dukakis’s doomed presidential campaign. Borosage
(2016) recalled the remark when observing Hillary Clinton’s 2016 campaign.
3 Popkin et al. 1976, pp. 793–5, 799–803. Nominating as his vice-presidential running mate
someone with a history of mental illness and electric shock therapy for it, without having discovered
that fact in background checking before the nomination, represented a catastrophic failure of due
diligence from which McGovern’s campaign never recovered.
4 Starting with the botched Watergate break-in itself and continuing through a series of
ham-fisted attempts to cover up White House involvement in it, culminating in the clumsy
erasure of a crucial 18 minutes on the Oval Office tapes, the whole episode was a comedy of
never-ending errors. When Nixon proclaimed in his televised speech of 17 November 1973, ‘I am
not a crook’, almost as many Americans agreed as disagreed that ‘President Nixon is a man of high
integrity’; and even by the time he resigned eight months later, 35% of Americans still agreed
(with 53% by that time disagreeing) (Ladd 1998, p. 32).
5 ‘When the tax finally ceased to be collected in 1993, it emerged that some £2 billion to
£2.5 billion of poll tax remained unpaid . . . Short of being dynamited, houses and flats cannot
2 An Epistemic Theory of Democracy
ask people about their ‘trust in government’, they find that people’s responses
are as much driven by their trust in its competence as their trust in its values.6
Nor are those merely brute political facts about the judgments that voters
pass on their leaders. They actually constitute good grounds for throwing the
rascals out—or anyway they would do, if there were any reason to think the next
lot would be more competent. There are genuine cases of rogues and scoundrels
in government, to be sure. But much more common are bumbling incompe-
tents, who simply are not up to the job—it is, after all, a pretty tough one.
The blunders of our governments derive from many sources.7 Evil intent is
occasionally one of them. More often, however, government errors are owing
more to a lack of due care and attention—misfeasance rather than malfeasance.
Part and parcel of that is public officials simply not getting their facts right—a
failure to reason properly from true facts to the logical conclusions. We should
never underestimate the impact of sheer ignorance—crucial facts that were
missing, dots that were not connected—in accounting for why public policies
sometimes go so badly wrong.
From what generic flaws in decision-making do these blunders arise?
Doubtless they are many and varied. But central among them is a lack of openness
to inputs of outsiders. Certainly that rings true of our opening examples.
Richard Nixon was notoriously secretive and paranoid, assiduously shielded
from outsiders by a so-called Berlin Wall consisting of his two principal
aides, Bob Haldeman and John Ehrlichman.8 Margaret Thatcher was notori-
ously stubborn, as well as privately insecure; her motto was, ‘The lady’s not
for turning’, and she continually batted away unwelcome advice. Had George
McGovern’s team canvassed more widely when conducting their background
checks, they would not have made the electorally fatal mistake of nominating
a running mate with a medical history that many voters thought rendered
him unfit for office.
The phenomenon is a perfectly general one. There is a lot of knowledge that
is widely dispersed across a given society, and political decision-making and
public policymaking would be improved, and errors avoided, if that dispersed
knowledge were taken more systematically into account. That is the back-
ground assumption guiding our work. The overall aim of this book is to help to
identify the best ways of doing that.
simply disappear. Determined individuals can and did . . . [with] an estimated 700,000 adults dis-
appearing from – or never appearing on – the electoral register’ (King and Crewe 2013, pp. 59–60).
6 Rothstein and Stolle 2008, pp. 452–4, esp. Table 2, Model 9 (their ‘effectiveness’ is our
‘competence’). See similarly Dahlberg and Holmberg 2014.
7 King and Crewe 2013. Similar themes are explored by Bovens and Hart (1995).
8 These men constituted what we will discuss under the heading of ‘epistemic bottlenecks’ in
Chapter 17 of the book.
Introduction 3
1.2 ASK AROUND
When you want to find out the facts, the best thing to do is usually just to ‘ask
around’. Seemingly the most natural thing to do is to ‘ask the experts’. But
notice, significantly, that even that injunction is in the plural. When one doctor
advises you to undergo surgery, it is always a good idea to get a second opinion
if time permits. And, if it’s a serious operation, maybe a third and a fourth. And
if they give conflicting advice, maybe more than that.
It is not as if there are no objective facts about the world, and it is all just a
matter of opinion. There are facts. But all of us, even putative experts, have only
imperfect access to them.9 That is why we seek advice on the facts from a diver-
sity of sources and adopt schemes to institutionalize that practice.10 That is why
scientific experts themselves convene ‘consensus conferences’ to work out what
facts all of them pretty much agree on, on the basis of evidence from their own
labs, and what purported facts are still in dispute and require further research.11
Why confine our enquiries to experts, though? Knowledge is widely dis-
tributed across society.12 Some people (putative ‘experts’) know a lot about a
little. Many know a little about a lot. That is the ‘many minds’ or ‘wisdom of
the multitude’ insight upon which fashionable contemporary techniques
of ‘crowdsourcing’ are based.13
Jokey factoids abound. The best estimate of the number of jellybeans in a jar,
or the weight of a fatted ox at a county fair, can be derived from pooling the
estimates of a great many people entering a competition to guess that number.14
In other more serious contexts as well, pooling the information of large
numbers of people through ‘prediction markets’ proves more reliable than
expert judgements. That is true when it comes to predicting the sales of a
Hewlett Packard printer or the outcomes of an Eli Lilly drug trial, for just
two examples.15 And as followers of politics will know, the Iowa Electronic
Markets reliably outperform opinion polls in predicting the outcomes of a
wide range of elections.16
9 Tetlock 2005. 10 Lane 1999. 11 NIH 2013. 12 Hayek 1945.

13 Surowiecki 2004; Wolfers and Zitzewitz 2004; Hanson 2013. The ‘many minds’ phrase is
from Sunstein (2006a; 2009) and Vermeule (2009b). The ‘wisdom of the multitude’ is Aristotle’s
phrase in Book 3, Chapter 11 of the Politics, recalled by Waldron (1995; 1999a, ch. 5; cf. Schwartzberg
2015). In arguing for the superiority of legislatures over courts or executives as decision makers,
Waldron (2000; 2016, pp. 130–4) himself makes much of the fact that there are simply greater
numbers of people involved in the decision-making in the first than the second two bodies.
14 Surowiecki 2004, pp. xi–xiii, 5. Sunstein 2006c, p. 24. The studies to which they refer are,
respectively: Treynor 1987; Galton 1907c.
15 Chen and Plott 2002; Servan-Schreiber 2012.
16 That remained true even in the 2016 US presidential election, when the final average of the
polls gave Trump only a 28.6% chance of winning (FiveThirtyEight 2016) but the Iowa Electronic
Markets had the two candidates at virtually even money at the close of betting (IEM 2016). See
more generally: Forsythe et al. 1992; Forsythe et al. 1999; Arrow et al. 2008. For details of its
operation see the IEM website <http://tippie.uiowa.edu/iem>.
Crowdsourcing is an increasingly popular tool of government and public

policymaking.17 Writing in the journal Science in 2008, a veritable who’s who of
social scientists—including three Nobel Laureates in Economics—advocate
regulatory reforms that would widen the use of such techniques. They bookend
(preface and conclude) their plea by saying:
There is mounting evidence that [prediction] markets can help to produce
forecasts of event outcomes with a lower prediction error than conventional
forecasting methods. . . .
These markets have great potential for improving social welfare in many domains.18
What’s the trick? It’s simple: the truth is constant and singular, while error is
multiple and random. Randomly distributed errors cancel one another out,
leaving the truth as the one strong signal that comes clearly through.19
1.3 PO OLING INFORMATION AND

JUD GEMENTS BASED ON IT
Not all knowledge is propositional knowledge. ‘Know how’ is as important as

‘know that’, when it comes to good government.20 But knowledge of good tech-
niques and procedures can be profitably shared as well. That process might
perforce involve more judgement and less mechanical aggregation of reports of
beliefs pro and con. Still, good practice too can be learned, and we can learn it
from one another.
In this book, however, we will by and large be focusing upon propositional
knowledge more narrowly, and ways of pooling judgements about it more
mechanically. Among decision theorists, the most familiar way for doing so is
through the use of Bayes’ theorem, updating your own prior estimate of the
probability of a proposition being true in light of reports from more-or-less
trusted others that it is true. We will discuss briefly the convergence between
that approach and ours in Section 3.3.1. But for the most part we shall here take
another tack—one that is at once more intuitive and more political.
17 Lehdonvirta and Bright 2015. On his first day in office President Obama (2009) signed a
memorandum dictating that ‘Executive departments and agencies should offer Americans
increased opportunities to . . . provide their Government with the benefits of their collective
expertise and information’. On her experience as Chief Technical Officer in the White House
administering that policy, see Noveck (2015). On the attempt to crowdsource a constitution for
Iceland see Landemore (2015).
18 Arrow et al. 2008, pp. 877, 878.
19 Of course, if error is not random then the trick does not work. We will say more about why
error might not be random, in politics, in Section 4.5 and Chapter 5. Perhaps another ingredient
contributing to the epistemic success of prediction markets, in particular, is the clever incentive
structure combined with the power of crowdsourcing (Hanson 2013; Sunstein 2006b).
20 Ryle 1949, ch. 2.
Introduction 5
That older and more natural way to pool opinions is through a vote, taking
the decision of the majority (or in a many-option contest, of the plurality) to be
veridical.21 In Joshua Cohen’s early and influential description of this approach:
An epistemic interpretation of voting has three main elements:
(1) an independent standard of correct decisions—that is, an account of justice or
of the common good that is independent of current consensus and the outcome of
votes;
(2) a cognitive account of voting—that is, the view that voting expresses beliefs
about what the correct policies are according to the independent standard, not
personal preferences for policies; and
(3) an account of decision-making as a process of the adjustment of beliefs, adjust-
ments that are undertaken in part in light of the evidence about the correct answer
that is provided by the beliefs of others.22
In what follows we will have occasion (in Chapters 3 and 4 particularly) to
amend Cohen’s own account in almost every particular.23 Still, that description
provides a good first approximation to what is involved in the family of
approaches that we shall be discussing.
1.4 THE COND ORCET JURY THEOREM, IN BRIEF
The history of philosophy contains many allusions to the ‘wisdom of the

multitude’, starting with Aristotle’s.24 But the most precise way of fleshing
out that notion is one that goes back to the early days of modern probability
theory—Condorcet’s famous jury theorem (‘CJT’ for short).25 We describe the
theorem more formally in Chapter 2, and extensions of it and objections to it
in the three chapters following that.
Roughly and in brief, the CJT says two things. First, the majority vote among
a group of (independent, competent, sincere) voters, each of whom is more
likely to be right than not, is itself more likely to be right than are individual
voters separately. Second, as the number of such voters approaches infinity, the
probability that the majority among them is correct approaches one.
21 That, as distinct from sharing information with one another about the reasons lying behind
their votes (Edelman 2002)—although we will discuss that too, in Chapter 9.
22 Cohen 1986, p. 34.
23 Specifically, in Section 4.1.4 (and again in Chapters 13 and 14) we will deny Cohen’s claim
that Condorcet’s jury theorem (the CJT) can only be used to track ‘justice or . . . the common good’
and cannot relate to ‘personal preferences’. And throughout, we deny that the CJT necessarily
involves any ‘adjustment of beliefs’ on the part of individuals (that would be a Bayesian frame-
work, which, as Section 3.3.1 shows, is similar to but not the same as the CJT).
24 Waldron 1995; 1999a, ch. 5. Arguably, however, Aristotle himself meant something other
than what the epistemic interpretation of that phrase suggests (Schwartzberg 2016).
25 Condorcet 1785; 1785/1976; 1785/1989; 1785/1994.
The process of aggregating votes is purely mechanical, to be sure. It elides

other important elements of judgement, which surely matter in all sorts of ways
(we will consider the role that deliberation and discussion have to play in
Chapter 9). But if what we are seeking are true statements about the world,
then soliciting the independent views of lots of informants on the propositions
in question and simply totting up their responses can be a very good way of
discovering those truths.26
The mechanism by which the CJT works is once again quite simple. Over the
course of many repetitions of some stochastic process, the relative frequency with
which any given outcome occurs will approximate ever more closely the probabil-
ity of its occurring ex ante.27 Imagine an urn filled with a very large number of
balls, 52 per cent of which are red and 48 per cent of which are black. In your first
ten draws you might well get four reds and six blacks. But after a hundred draws28
the proportions would become much nearer to the true percentages in the urn,
and after a thousand draws they would become nearer yet again. By extension, if
each voter votes independently of every other, and each is 52 per cent likely to vote
for the correct outcome, then among hundreds (and still more thousands or
millions) of such voters the correct outcome is very likely to garner something
very close to 52 per cent of the votes. In any case, the correct outcome is very likely
indeed to get above the 50 per cent + 1 threshold required to win—and it is
increasingly likely to do so the more such independent voters there are.
A word about history.29 Nicolas Caritat, Marquis de Condorcet, was himself
a world historical figure. In the editorial introduction to the volume of his
26 Perhaps Rousseau’s horribly garbled discussion of why ‘the general will never errs’ in
The Social Contract was referring to something like the CJT (Grofman and Feld 1988;
cf. Philonenko 1984). In outlining the difference between the will of all and the general will,
Rousseau (1762, bk. 2, ch. 3, para. 2) says:
the latter looks only to the common interest, the former looks to private interest, and is
nothing but a sum of particular wills; but if, from these same wills, one takes away the pluses
and the minuses which cancel each other out, what is left as the sum of the differences is the
general will.
But what mathematical sense can that possibly make? How can the ‘cancelling out’ involved in the
latter operation be any different from the addition (inevitably also involving pluses and minuses) in
the former operation? How can the latter sum, which is supposed to be so distinct from the former,
be any different at all? Perhaps what Rousseau was gesturing towards (as his subsequent discussion
of the operation of factions in that same chapter suggests) was the workings of something like the
jury theorem published by Condorcet twenty-three years later, and the way in which that theorem
involves error (which is particular to each of the individuals and randomly distributed among
them) cancelling, and leaving truth (which is common to all) as the remainder. But who would
possibly have guessed that, from Rousseau’s own garbled formulation?
27 On one definition, probability just is that frequency. But in saying this we do not necessarily
mean to endorse a frequentist definition of probability: what we say in the text will be the empirical
consequence of probability defined in many other ways as well.
28 Assume that the balls are put back into the urn after each draw. Or, alternatively, assume that
the number of balls in the urn is much greater than the total number of draws.
29 For a good short account, see McLean and Hewitt (1994). For a more extended treatment see
Williams (2004).
Introduction 7
political writings in the Cambridge ‘blue books’ series, he is dubbed ‘the last of
the great French Enlightenment philosophes’—and, indeed, he was regarded as
such ‘in his [own] lifetime’.30 A protégé of d’Alembert and Turgot and a favourite
of Voltaire, Condorcet debated with Borda in the Académie Française and
served as permanent secretary of the Académie des Sciences until it was
suspended in the aftermath of the Revolution. He co-founded with Tom Paine
the first Republican club in France. He presided over the Assemblée and its
committee that drafted both a constitution for the new Republic and the
‘Déclaration des droits de l’homme et du citoyen’. In the end he was denounced
by Robespierre and hounded to death during the Terror. Franklin and Jefferson
had been frequent participants in Condorcet’s wife’s famous salon during their
time as American emissaries in Paris, and Condorcet commended his beloved
daughter to the care of them and their families in his last testament written
while in hiding.31 Jefferson included Condorcet’s Essay on the Constitution and
the Functions of Provincial Assemblies in a crate of books he sent to James
Madison, just as the Madison was drafting the Virginia Plan for the American
Constitution. And so on.
But our interest here is in Condorcet’s work, not in the person or any direct
influence of his on subsequent world events. That direct influence, it seems, was
slight.32 There is no evidence that Madison even opened Condorcet’s book as it
passed through his hands (although he had clearly read Condorcet’s Letters
from a Freeman of New Haven to a Citizen of Virginia).33 In short, Condorcet
and his jury theorem were largely lost to history (certainly anyway to political
theory), until it was resurrected in the middle years of the twentieth century.34
Yet, as this book shall show, it is a powerful tool for building an epistemic theory
of democracy.
1.5 EXTENDING THE COND ORCET JURY THEOREM
The Condorcet Jury Theorem and jury theorems related to it are the key analytic
devices driving this book. Voting (or counting heads more generally) lies at
their heart and will therefore always be a key part of the story.
30 Lukes and Urbinati 2012, p. xv. Urbinati 2006, p. 178. 31 Condorcet 1794/1994, p. 290.
32 Although it should be noted that his sketch on ‘progress’, written while in hiding, provided
one of the principal foils for Malthus’s 1798 Essay on the Principle of Population—the full title of
which continues as it Affects the Future Improvement of Society, with Remarks on the Speculations
of Mr. Godwin, M. Condorcet, and Other Writers (Malthus 1798/1992, p. vii).
33 McLean and Urken 1992, pp. 453–5.
34 In passing, and disapprovingly, by Black (1958, pp. 159–65); more enthusiastically and influ-
entially by Barry (1964, 9–14; 1965, Appendix A, pp. 292–3). Within economics it lay dormant for
even longer (Piketty 1999).
We will hardly confine ourselves to the analysis of voting mechanisms alone,

however. On the basis of the CJT, we will construct a broad epistemic analysis
of democratic government as a whole. Deliberation as well as voting will be
part of that (in Chapter 9). So too will be agenda setting and civic education (in
Chapters 8 and 6). The role of leadership, of tradition, of factions, and parties
will all be analysed in these terms (respectively, in Chapters 11, 10, 14, and 17).
The structure of government, and the way key players ought to approach their
roles within it, will also be subject to that scrutiny (in Chapter 16).
We will be building on the foundation provided by the Condorcet Jury
Theorem, not simply applying it. Much of the work of the book lies in elaborating
extensions and variations on the basic CJT. There, what we will be doing is
developing and applying an umbrella framework that sustains several related
jury theorems.
Those extensions and elaborations on the basic CJT are designed, in the first
instance, to help overcome some of the familiar objections to the applicability
of the CJT to political life in the real world (we say more about those shortly).
But we offer those extensions and elaborations of the CJT not so much defen-
sively as positively—as a way of helping to extend the applicability of the CJT to
political life in the real world in certain genuinely important respects.
For example, casual observers might imagine that the CJT requires more
competence among voters than it really does (as we show in Section 3.1.1). They
might wrongly assume that it applies only to binary choice situations, when
actually it can be extended to apply to many-option cases (Section 3.2). They
might wrongly assume that it requires voters to be totally independent of one
another, whereas it actually presupposes interdependence among them in one
crucial respect (Section 5.1) and it can, without any change in the substantive
results whatsoever, tolerate at least a limited amount of interdependence among
them in other respects (Section 4.5.3).
Perhaps the most important extension of the CJT grows out of our discus-
sion of independence in Sections 5.3 and 5.4. There, we set out a related jury
theorem according to which the epistemic success of any group is limited by
the truth-conduciveness of the decision situation before it (the veracity of
available evidence, for example). The truth-conduciveness of the decision situ-
ation fixes the best that even the ‘best responder’ to that situation can do in the
epistemic circumstances at hand.
1.6 FORESTALLING FAMILIAR OBJECTIONS
Now let us briefly foreshadow three of the most common objections to the
applicability of the Condorcet Jury Theorem, and sketch how we will deal with
them in the chapters to come.
Introduction 9
1.6.1 Competence
The Condorcet Jury Theorem crucially presupposes that individual voters are
minimally competent, which is to say, ‘better than random’. As long as that is
true of individual voters (and they vote sincerely and independently of one
another), the CJT tells us that the majority among a large electorate is very
highly likely to be right. If individual voters are worse than random, however,
then the CJT offers an equally firm warning that the majority among a large
number of such voters is very highly likely to be wrong.
So which is more likely? A long line of commentators, starting with Plato
and including Condorcet himself, fear that the ignorance of the general public
would drive mass democracy in the latter direction.35 Their fears have been
stoked by the findings of contemporary social research that people in general
are woefully ignorant of basic political facts, when quizzed about them.36 Those
findings are warmly embraced by activists who are looking for an excuse to
wind back the state.37 And in light of those findings, sober social theorists hesi-
tate to put all of their epistemological eggs in the CJT basket.38
In the chapters that follow, we will argue that voter ignorance is not as deep
as people’s responses to those quiz-style survey questions might suggest. Voters
can employ ‘cues’ and informational shortcuts to help them vote in the right
way from their own perspective (the way they would have voted if they had
been fully informed) without having the sort of detailed factual knowledge
being tested in those quizzes (Chapters 6, 12). People can inform themselves
through deliberations and discussions (Chapter 9), and by taking the advice of
more knowledgeable acquaintances (Section 11.4). They can pool knowledge
with others whose judgements they trust should be similar to their own
(Sections 14.2 and 14.3). And, as John Stuart Mill never tired of emphasizing,
politically empowering people leads to their becoming more politically
informed as well (Section 15.6).39 In all of those ways, voters can and arguably
do get more task-specific knowledge that is likely to make them competent
enough for CJT purposes.
Remember, too: ‘competent enough for CJT purposes’ means merely ‘better
than random’. If people have absolutely no information or inclinations what-
soever, their votes will be random. It is easy to see how people, on average,
might be a bit better than random—and that is all the CJT requires (Section 3.1.1).
Maybe they are interested in the matter, or maybe they incidentally acquired
information bearing on the matter in the course of their other activities, for
example. It is hard to see how people, in general, would be worse than random.
35 Condorcet 1785/1976, pp. 49–50.

36 See Section 6.1.1 for a discussion of those findings.
37 Hayek 1960, p. 110; Caplan 2007; J. Brennan 2011b; 2014; 2016.
38 See esp. Estlund 2008, pp. 228–30. 39 Mill 1861/1977, ch. 8, pp. 467–9.
Of course, any given individual might be wrong on any given occasion; indeed,
among a large group of people, some people might even be wrong on virtually
all occasions. But how can a large number of individuals be systematically
worse than random, except by all of them being subject to the same set of
external influences that lead them all to err in the same direction at one and
the same time?
1.6.2 Independence
That points to the second standard critique of the Condorcet Jury Theorem.
It assumes that people’s votes are independent of one another’s. Historical debates
are replete with arguments for limiting the franchise to those of ‘independent
will’, and various classes of people have historically been denied the vote because
they were thought to fail that test (among them slaves, servants, apprentices,
wives, bankrupts, and the propertyless).40
Often those older arguments were couched in terms of ‘virtue’. But in terms
of the CJT it is purely a matter of epistemic prudence. Given a large number of
minimally competent voters each of whom is sincere and independent of one
another, the CJT assures us that the majority is very likely to be correct. But if
everyone votes the way the boss tells them to vote, then the probability that the
majority vote among them is correct is no higher than the probability that
the boss is correct, however many votes have been cast.
We no longer have slaves nor very many servants, and we now regard wives
and people without property as more independent than we used to do. But
there are many other ways in which the independence of voters might be
compromised. People read the same newspapers (or nowadays blogs), listen
to the same speeches, they share many of the same experiences, they talk to
one another. There is no way that their votes are statistically independent of
one another’s.
But that is simply the wrong standard, for jury theorem purposes, as we
explain at length in Chapter 5. For a start, the votes of competent voters track
objectively true facts about the state of the world (that is just what it means to
be ‘competent’ in the CJT). There will inevitably be a statistical correlation
among their votes, for that reason. Furthermore, voters do not have direct,
unmediated access to the true state of the world. They only have evidence about it.
For that reason, too, there will be correlation among the votes of competent
people voting on the basis of the best evidence that is available to them all. The
best the majority of even a very large number of highly competent voters can
40 Kouser 1984. For arguments to that effect see: the exchanges between Ireton, Cromwell, and
Petty at the Putney Debates (Woodhouse 1938, pp. 82–3); Blackstone 1783, bk. 1, ch. 2, sec. 5;
Jefferson 1785/1964, query 19.
Introduction 11
then do is, naturally, limited by the quality of the best evidence available to them;
and the probability that such a majority is right is effectively upper-bounded by
the probability that that evidence is informative. (That is the ‘Best Responder’
corollary introduced in Section 5.3.) This result is useful on two accounts: it
leads to a more realistic assumption about independence, and also to a more
realistic jury theorem, sensitive to the quality of the evidence.
There are other epistemically less fortuitous ways in which people’s votes
might fail to be independent of one another’s. Voters might blindly follow the
same opinion leader, or they might harbour the same prejudices, or they might
just share the same psychology and be subject to the same heuristics and biases.41
Those ways would all compromise the epistemic performance of even a very
large group of voters—but not necessarily fatally so. Variants of the CJT still
work, albeit a little more slowly, even with interdependence among votes of that
sort, just so long as these misleading factors are not too strong (Section 4.5.3).
They still work if the effects of the various sources of interdependence cancel
one another out (Section 11.2.2)—or even if there are just many independent
and minimally informative common influences at work (Sections 11.3, 12.4,
12.5). And restricted versions of the CJT still work even if none of that is true
(Section 5.4).
1.6.3 The Truth Value of Values
Perhaps the most fundamental objection to the application of the Condorcet

Jury Theorem to politics is that it pertains only to matters of fact, whereas
politics crucially involves value judgements. Value judgements cannot be
true or false, correct or incorrect, in the same way that factual judgements
can be. Hence, the CJT simply does not apply to them—or to politics more
generally, insofar as politics is shot through with value judgements. Or so the
objection goes.42
There are various ways in which value judgements might have truth values,
which we will sketch shortly. But first it is worth emphasizing that our princi-
pal focus in this book is not on those. It is instead on common or garden truths
of the sorts we bump into in science and when negotiating our way in the
world. Our discussion of senses in which moral claims might be ‘true’ is meant
merely to show that those may be true too, albeit perhaps in rather different
41 Estlund (1993, p. 99, n. 44; 2008, p. 16) echoes earlier worries on this score (Condorcet
1785/1976, pp. 49, 62; Mill 1872/1974, bk. 3, ch. 18, sec. 3, p. 539). See similarly: Waldron 1989, p. 1323;
Sunstein 2006c, pp. 34–6.
42 Black 1958, p. 163; Miller 1992, p. 56; Copp 1993; Estlund 1993. Some would say that
‘prudence’, and ‘judgement’ as the virtue involved in tracking it, is different in kind yet again
(Beiner 1983); but here we simply take that to be a mixture of those two other kinds, facts and
values.
ways than those other truths are true. To reiterate, however: those ‘moral truths’
are certainly not the only, or even the principal, ones we will be discussing in
this book.
The objection in view confidently asserts that value judgements have no truth
value. But that assertion itself may be simply incorrect. We discuss various ways
that might be so in Section 4.1. For a start, there is disagreement among meta-
ethicists on precisely this question. Various strands of cognitivists, especially
moral realists, would insist that moral claims can indeed be true or false. And
even those who say that morality is ‘invented’ (or more politely, ‘socially con-
structed’) would have to agree that there is surely some fact of the matter about
‘what the conventions around here are’.
In short, our response to this objection will be the same as that of two prom-
inent earlier epistemic democrats, with whom we disagree in many other
respects. Estlund and Landemore write:
By ‘correct or right decision’ here, or ‘the truth’, can be meant an array of things,
from objective truth of the matter (about facts or morality) to a more intersubjective,
culturally-dependent, and temporary construct (about more socially constructed
facts or moral questions). What epistemic democrats emphasize . . . is merely the
Habermasian (and commonsensical enough) point that we wouldn’t be exchan-
ging reasons in the first place if we did not believe that there was something to
figure out, whether we call this something the truth, the right, or the correct, just
or socially useful answer.43
1.7 THE STRUCTURE OF THE B O OK
This book is divided into four nearly equal parts.

In Part I we first introduce the Condorcet Jury Theorem more fully and more
formally (always striving, nonetheless, to keep the discussion easily accessible).
We then go on to discuss some extensions of the classic CJT, to discuss its limits
(which we argue are less limiting than often supposed), and to discuss in more
detail the crucial Independence Assumption that is often seen as the Achilles’
heel of the CJT.
In Part II we discuss various ways in which epistemic performance might be
enhanced. One is at the individual level, improving the individual competence
of voters. Others operate at a more systemic level. Those include: introducing
more diversity among the decision-making group; introducing a division of
epistemic labour within it; and encouraging discussion and deliberation across it.
Many of the contributions of discussion and deliberation operate outside the
strict framework of the CJT, but are clearly related to it: improving the decision
43 Estlund and Landemore 2018.

Introduction 13
situation, expanding the evidence base, expanding the agenda for decision,
and such like.
Part III examines the epistemic effects of various familiar practices through
the lens of the CJT. Both respecting tradition and following leaders can often
carry considerable epistemic costs, although we show that there are ways in
which those costs can be largely avoided even with those practices. Taking cues
can be a good epistemic strategy for voters who are not themselves competent
and is an epistemically tolerable strategy even for those who are. We conclude
this part of the book with a discussion of how differing values and priorities
within the community might be accommodated within a CJT-like framework.
A version of the CJT goes through, even under circumstances of pluralism,
regarding v alues and priorities. And an epistemic form of factionalism can help
the epistemically disadvantaged avoid having their distinctive interests electorally
disadvantaged.
Part IV uses the CJT logic to assess various structures of government, from
an epistemic point of view. There we show that epistemic considerations do not
unequivocally favour an epistocracy rather than a democracy, and that in fact
there are important epistemic considerations weighing in favour of a broad
franchise. We go on to show reasons for thinking that, counter-intuitively,
representative democracy might actually be epistemically preferable to direct
democracy and that it is epistemically no problem for many of those represen-
tatives to act as trustees, as long as just a few of them see themselves as delegates
bound by the instructions of their constituents. We go on to show how some
institutional designs can impose epistemic bottlenecks that hinder collective
competence, while other institutional design features can serve to increase it.
Part V summarizes the results and revisits the larger philosophical issues
surrounding democracy and truth-tracking. We ask whether there are truths in
politics, what they might look like, and whether it is possible to identify them
by asking many voters. Some have suggested that going after ‘The Truth’ is
counter-productive, perhaps even dangerous. While there are some situations
in which democratically searching for The Truth is unhelpful, legitimate gov-
ernment requires that decisions be epistemically meritorious in general.
Finally, in the Epilogue, we assess the implications of the 2016 US presidential
elections and the ‘Brexit’ vote in the UK for epistemic theories of democracy—
and indeed for democratic theory in general.
Part I
The Condorcet Jury Theorem
In this part of the book we will introduce more fully and formally the basic
device that will be driving our discussions throughout the rest of the book—the
Condorcet Jury Theorem.
We begin in Chapter 2 by describing the classic CJT framework, the rela-
tively strong assumptions underlying it, and the theorem that can be proven on
the basis of them. We also provide illustrations to suggest that the same general
tendency can be strongly observed well short of the limiting case to which the
formal theorem itself pertains.
We then proceed in Chapter 3 to show how that classic CJT framework can
be extended in various directions. There we show that the CJT’s Competence
Assumption can be weakened to allow for voters with heterogeneous compe-
tence levels and that a parallel theorem can be proven for plurality decisions
with more than two alternatives. We also show, in more illustrative fashion,
that the epistemic advantages of democratically pooling the judgements of a
large number of voters is not peculiar to majority or plurality rule alone but
instead extends to a wide range of other procedures for aggregating votes.
Chapter 4 is dedicated to a discussion of why the aggregation of votes can
sometimes fail to track the truth because one or another of the assumptions
required by the CJT is violated. One ostensible reason is that politics is about
values and there might not be any truth (or anyway not any single truth) for
votes to track about those or other things of political importance. Other reasons
are that voters might not be competent, sincere, or independent in the ways the
CJT requires. Naturally, some of those problems prove more problematic than
others. But in each case we find reasons for thinking that those limitations may
not be as limiting as they are sometimes supposed to be.
Much the most important reason the CJT might fail is generally thought to
be the demandingness of the Independence requirement. We discuss that issue
at length in Chapter 5. There we distinguish several forms of independence and
discuss just how much of a risk failure in respect of each of them poses for the
basic CJT results. Building on those discussions, we offer two new models in
the spirit of the CJT. One explains why even a large group of competent voters
may be less than virtually certain to be correct (because Independence is

violated by their all relying on the same body of the best available evidence).
The other suggests how even groups of voters whose votes are determined by the
same common causes may come to reliably correct conclusions, if those com-
mon causes are numerous, reliable, and independent of one another. A suite of
such modified jury theorems emerge out of the Best Responder Corollary we
introduce in Chapter 5.
Everyone who has heard of the CJT result and wants to avoid getting embroiled
in those discussions in connection with their own particular project typically has
some stock excuse—typically one or another of those discussed in Chapters 4
and 5—for not so engaging. The upshot of our discussion in those chapters is
that most of those excuses simply will not suffice. At the very least those excuses
need to be stated much more precisely to do their intended work.
The Classic Framework
In this chapter we introduce, briefly, the classic version of the Condorcet

Jury Theorem. Formal proofs are found elsewhere.1 In this chapter we sim-
ply describe the classic setup, what assumptions are required, and what
results follow if those assumptions are met. Extensions of the CJT beyond the
classic framework, to other settings and using different and weaker assump-
tions, will be discussed in Chapters 3 and 5.
2.1 THE SETUP
The classic CJT framework involves a group composed of n voters making a choice
between two alternatives by means of majority rule.2 For convenience, n is usually
assumed to be odd so as to avoid ties. If n is even, we assume that ties are broken
by a fair coin toss. Exactly one of the two alternatives is correct. We call the variable
that denotes the correct answer the ‘state of the world’ (or the ‘state’ for short).
2.2 THE ASSUMPTIONS
The CJT rests on three assumptions as follow.
2.2.1 The Competence Assumption
Before voting, each voter forms a belief concerning which of the two alterna-
tives is correct (and all voters plump for one or the other alternative: none
1 See e.g. Ladha 1992, pp. 632–3. A round-up surveying proofs in the vicinity of the CJT is
provided by Grofman et al. 1983.
2 Although the classic CJT is invariably expressed in terms of ‘majority rule’, notice that in the
two-option case of the classic CJT framework ‘majority rule’ and ‘plurality rule’ are extensionally
equivalent. This will become important in Section 3.2 below.
remains ‘undecided’). Informally, a voter’s competence is the probability that

that belief matches the true state of the world. In other words, competence
measures how likely the voter is to identify the correct alternative. A bit more
formally expressed, the Competence Assumption says:
Competence: Each voter’s belief about the correct alternative is true with prob-
ability pc > 0.5 (and this holds for both states and is the same for all voters3).
This assumption ensures that voters tend to recognize which is the correct
alternative among the two before them. Importantly, our notion of correctness
is flexible enough to accommodate any external standard, as long as the stand-
ard that is applied fixes exactly one alternative as the right one. In particular, a
correct alternative can also be the ‘better’ alternative according to some inde-
pendent standard of betterness (such as ‘more nearly correct’4), rather than the
‘true’ alternative in a more demanding sense of truth.
Notice that the classical CJT framework stipulates that every voter has
exactly the same probability of assessing correctly which is the correct alterna-
tive. That assumption can be relaxed in ways we shall discuss in Section 3.1.
In the classic CJT’s two-option case, the probability pc (which we shall
call the ‘voter’s individual competence’) must be above 0.5. More generally,
what the Competence Assumption requires is that voters are ‘better than
random’ at choosing the correct option. Someone choosing randomly in a two-
option case would be right half the time; the Competence Assumption requires
that they be better than that.
2.2.2 The Independence Assumption
The Independence Assumption says:

Independence: The beliefs of all voters are statistically independent, given
the true state of the world regarding the correct alternative.
In the classical CJT framework Independence is most commonly discussed in
pairwise terms.5 But Independence can also be violated by there being certain
3 I.e., each voter is assumed to be pc > 0.5 likely to say the alternative is correct if it is indeed
correct, and with the same probability pc > 0.5 likely to say it is incorrect if it is indeed incorrect.
In other words, the probability for identifying the correct alternative is the same for both states
and all voters. This rider should be understood as being attached to all subsequent references to
assumptions about a voter’s competence, even if (for brevity) we often omit to state it explicitly.
4 In the case of the ‘guess the weight of the ox’ example mentioned in Chapter 1, the correct
answer is 1,207 lbs (Galton 1907c). Someone who guessed 1,200 lbs was ‘more nearly correct’
than someone who guessed 1,500 lbs. And if participants in the contest had been asked to
choose between two alternatives (1,200, 1,400), the correct (exactly correct) choice out of that set
is 1,200 lbs—even though the exactly correct truth about the weight of the ox is 1,207 lbs.
5 E.g. Estlund (2008, p. 225): ‘the probability of one voter, say Joe, getting the right answer is
exactly equal to the probability of Joe getting the right answer given that Jane did’.
The Classic Framework 19
patterns of votes such that, even though the votes are all pairwise independent,
the votes together are not independent.6
The effect of the Independence Assumption within the classical CJT frame-
work is to ensure that each voter is providing new, independent information
when reporting his belief, rather than echoing someone else’s judgement7 or
expressing a judgement that was produced by the same underlying causal fac-
tors that led to anyone else’s. We will discuss this assumption further, along
with ways it could and should be relaxed, in Section 4.5 and Chapter 5 below.
2.2.3 The Sincerity Assumption
The Sincerity Assumption says:

Sincerity: All voters vote for the alternative they believe to be the correct
alternative.
In Condorcet’s way of putting it, ‘We shall suppose . . . that all express their
opinion in good faith’.8 One effect of the Sincerity Assumption is to rule out the
various forms of ‘strategic voting’ as discussed in Section 4.3. It also allows us
to apply Competence and Independence to votes.9
2.3 THE THEOREM
Call the probability that a majority in a group of size n votes correctly Pn. Given
the assumptions above, two results can be proven.
Non-asymptotic Result:
For every n and a constant pc such that ½ < pc < 1 : Pn + 2 > Pn .
That is to say, the majority of a larger group of voters is more likely to be correct
than the majority of a smaller group of voters.10 Note that this formulation
subsumes the result, sometimes stated separately, that the probability that the
majority vote of the group (with n > 2) is correct is greater than the probability
of each individual voter being correct (Pn > pc).11
6 Kaniovski 2010.
7 Condorcet’s (1785/1976, p. 47) original formulation was restricted to that alone: ‘we shall
suppose that none of the voters influences the votes of others’.
8 Condorcet 1785/1976, p. 47.
9 Sincerity and Independence entails the independence of votes; Sincerity and Competence
the competence of votes. Whenever we apply Independence and Competence to votes, Sincerity
is assumed.
10 Note that an increase in group size of 1 does not necessarily yield higher group competence,
as the move from an odd to an even-sized group can produce ties.
11 Since after all Pn = pc in the case of n = 1.
The second result that can be proven is this:

Asymptotic Result:
limn→∞ Pn = 1.
That is to say, the probability of a majority of jurors being correct converges to 1

as the number of jurors goes to infinity.
2.4 THE CALCULATION
Assuming the three conditions specified in Section 2.2 are satisfied, group
competence can be calculated by using the binomial distribution. The prob-
ability Pn that an odd number12 of n voters, each of whom is individually pc
likely to vote correctly, will, in making their decision by majority rule, choose
the correct outcome is:
n n
Pn = ∑   ( pc ) (1 − pc ) .
i n −i
n +1 i
i=
2
 
For large values of n we can estimate this value by normal approximation

 n 
 2 − npc 
Pn ≈ 1 − Φ  ,
 npc (1 − pc ) 
 
where Φ is the cumulative distribution function of the normal distribution.
In the applications of the CJT in other parts of the book we sometimes need
to calculate iso-competence. For suitably large groups, those calculations can
be done using the so-called Grofman–Dummkopf–Witkopf theorem.13 The
question to which that theorem provides an answer is this: suppose we have
two groups of different sizes; how individually competent would each member
12 The formula applies only to odd n. However, when ties are broken with a random coin toss
then, for an even n, the group competence is equal to the group competence for group size n – 1.
For an intuitive explanation, note that tossing a fair coin to decide in case of a tie is probabilistic-
ally equivalent to removing one random individual from the tied group and then adopting the
majority vote of the resulting smaller group. Since the resulting smaller group has the competence
of a group with size n – 1, the even group is, with random tie break, just as competent as the next
smaller odd group.
13 Grofman et al. 1983. The theorem relies on the application of the normal approximation and
can therefore only offer a good approximation for the condition of equal competence if both
groups are large enough to make the normal approximation a good approximation.
The Classic Framework 21
of one group have to be, in order for that group’s collective competence to equal
the collective competence of some other group whose own members each have
a different individual competence (assuming the two groups do not differ in
any respect except individual competence and group size)?14
2.5 HOW QUICKLY GROUP COMPETENCE

CONVERGES TO PERFECTION
To say, as the CJT does, that the probability that a majority of voters is correct
approaches certainty ‘at the limit’ as the number of voters approaches infinity is
not necessarily particularly informative. After all, sometimes what is true at the
limit radically ceases to be true not very far from that limit.15 Even the largest
electorates are not infinite. So the formal proof of the CJT needs to be supple-
mented, for application to real-world settings, with some information about
how the function in question behaves well short of the limit.
A note on numbers: How competent voters might be in the real world is an
empirical question that lies largely outside the scope of this book. When offering
numerical examples, our aim is not to provide any precise estimates. Instead,
our aim is to give a more qualitative feel for how the mathematics underlying
the CJT behave in something akin to real-world political settings. The numerical
examples we give are thus intended to be purely illustrative.
Figure 2.1 presents graphic evidence of just how quickly the probability that
the majority is correct rises toward certainty, as either (or both) individual
competence and/or group size increases. For ease of comparison, the figure
shows small population sizes (n ≤ 100) on a more expanded scale on the left,
and larger population sizes on a more condensed scale on the right.
14 Let there be two groups, the first of size n with individual competence pc1, the second
with size n + y and individual competence pc2, such that pc1 > pc2. The Grofman–Dummkopf–
Witkopf theorem states that (assuming both groups are appropriately large) the two groups are
collectively approximately equally competent if:
n ( pc1 − pc2 ) ( pc1 + pc2 − 1 )
y= .
4(1 − pc1 ) pc1 ( −1 / 2 + pc2 )
2
Solving for pc1 yields

1 1 n − 4npc 2 + 4npc22 + y − 4pc2 y + 4pc22 y
pc1 = + .
2 2 n + y − 4pc2 y + 4pc22 y
15 See for example our discussion the ‘impartial culture’ analysis in Section 3.3.2. If every
preference ordering is exactly equally likely, then the probability of voting cycles increases as the
number of voters increases; but if one preference ordering (e.g. the ‘correct’ one as in the CJT
model) is even slightly more likely than some other, then the opposite is the case (the probability
of cycles decreases the more voters there are).

1.0 pc = 0.7 1.0
pc = 0.6
Probability of correct majority winner Pn
0.9 0.9
pc = 0.55
0.8 0.8
0.7 0.7
0.6 0.6
pc = 0.51
pc = 0.505
0.5 0.5
0 20 40 60 80 100 100 5,000 10,000
Population size n
Figure 2.1 Group competence Pn for different levels of pc .
As we see in Figure 2.1, even with a relatively small group of 100 individuals,

each of whom is only 55 per cent likely to vote correctly, the majority vote of the
group is almost 85 per cent likely to be correct. If there are 1,000 such persons
in the group, it is virtually certain that the majority vote indicates the correct
alternative. The convergence happens even faster with higher individual compe-
tence. However, even very small levels of individual competence (e.g. pc = 0.505)
eventually lead to high group competence as the population size grows further.
Among a group of 10,000 people (the size of a small town) who are individually
pc = 0.505 competent, the majority is nearly 85 per cent likely to be correct,
for example.
Extensions
The classic Condorcet Jury Theorem, as presented in Chapter 2, operates under

some pretty restrictive assumptions. The task of this chapter is to see how and to
what extent those can be relaxed, while obtaining CJT-like results under those
other settings. The extensions in question involve: weakening the Competence
Assumption (Section 3.1); introducing more alternatives than just two (Section
3.2); and using other decision rules besides majority rule (Section 3.3).
In subsequent chapters we will, largely for expository convenience, phrase
our discussions in the language of the classic CJT framework which assumes
majority voting over two options and identically competent voters who are
independent of one another.1 That is the most familiar form of the CJT, after all;
and its simplifying assumptions do indeed serve to help simplify both the ana-
lysis and the exposition. But in most instances, qualitatively similar results
would follow if the strict assumptions of the classical CJT were weakened in the
ways outlined over the course of this chapter and the next ones.
3.1 WEAKENING THE COMPETENCE ASSUMPTION
The classic CJT framework assumes that every voter has exactly the same
competence as every other voter, and that that competence is better than random.
In the real world, however, it would be wildly implausible to think that everyone
is exactly as competent as everyone else on absolutely every topic.
3.1.1 Mean Competence among Heterogeneous Voters
Condorcet himself seems to have been oddly attached to that assumption of

identical competence, or to something very much like it, in the rhetoric with
1 ‘Independent, conditional on the state of the world’, as we explain in Section 5.1.3, is the

correct way of stating the assumption underlying the classic CJT.
which he surrounds his mathematics.2 Many complain, with good cause, that
that is a highly unrealistic assumption; and some seriously senior scholars
think that the CJT can be dismissed out of hand on those grounds alone.3 But
in terms of the mathematics, the assumption of equal competence is purely a
convenience rather than being in any way a mathematical necessity. We will
now show that the CJT still works, even when relaxing that assumption and
allowing competence to vary among voters in the electorate.
In allowing voter competence to be heterogeneous rather than identical, let
us for simplicity begin by assuming that the distribution of voter competences
is symmetric and that the mean is fixed as the group grows in size. Then it can
be proven that both the Asymptotic and the Non-Asymptotic conclusions still
hold, substituting the mean of voters’ competences pc for pc in the classic CJT
framework.4
Here is the underlying intuition. Ex hypothesi, in the cases under discussion
voter competence varies symmetrically around the mean. In a symmetrical distri-
bution, the mean is also the median. Hence, for every voter who, by some level of
probability ε, is more likely than the mean to vote for the correct outcome (com-
petence pc + ε ) there is some other voter who, by the same level of probability ε, is
less likely to do so to the same degree (competence pc − ε ). When averaging over
these two voters, they are each pc likely to vote for the correct alternative. Therefore
all voter pairs are expected to vote for the correct alternative with probability pc ,
so that they are (on pairwise average) expected to behave just like a group of voters
with homogeneous competence pc . The more voters there are with individual
competence symmetrically distributed around a constant mean of pc >0.5 , the
more likely it is that the majority vote among them is correct, and that value
approaches certainty as the number of such voters approaches infinity.
The obvious criticism of this relaxation of competence homogeneity is, of
course, that voters typically do not come with competences that are so neatly
paired symmetrically around the mean. Fortunately, the Asymptotic Result is
also preserved for any distribution of voter competences as long as the mean
pc is well defined and, as the number of voters grows, converges to a value
greater than 0.5. In the limit and with the other assumptions of the CJT hold-
ing, if pc > 0.5 then the probability that the majority of the group is correct
approaches certainty as the number of voters approaches infinity.5
2 ‘Numerous assemblies would . . . be appropriate to a country in which . . . there was a great

equality between minds, as to the soundness of their judgments and the truth of the principles
according to which they governed their conduct’, Condorcet (1785/1976, p. 50) writes in the Essai.
3 Converse (2000, 349), e.g., writes that ‘The Condorcet model is not the most telling model
[because] it assumes . . . that individuals contributing to the group judgment are “modestly and
equally well informed”. This does not seem a promising gambit for diagnosing the electorate,
given the staggering heterogeneity of informedness across it.’
4 Grofman et al. 1983, p. 268. For further details see: Grofman et al. 1982.
5 The assumption about the mean is weakened here. It is not assumed that the mean is constant,
it suffices that while the group grows the mean converges so that pc = limn→∞( p1 +…+ pn ) / n exists
Extensions 25
The Non-Asymptotic Result of the classic CJT, however, does not necessarily
obtain with heterogeneous and asymmetrically distributed individual voter
competences.6 It is sometimes possible for smaller groups to be more compe-
tent than larger groups; it is even possible for group competence to be lower
than mean individual competence.7 What is true is this: group competence will
always be higher than mean individual competence, just so long as the individ-
ual competence of each voter i is pci >0.5 and the group (assumed to be odd in
size) has more than two members.8 But that result is something well short of
the classic CJT’s Non-Asymptotic conclusion that larger groups always have
higher group competence than smaller groups.
3.1.2 Topic-specific Competence
Not only does individual voter competence vary across individuals. It also
varies across topics and issue areas.9
This variation in competence across topics and issue areas can take two forms.
One is ad hominem. Someone who is a specialist in nuclear physics, and highly
competent in assessing what is and is not true in that field, may be considerably
less competent at judging where the truth lies in disputes of a philological nature.
That is to say, any given individual’s competence is not necessarily constant across
all topics; rather, it ought to be seen as being indexed to the topic.10
There is a second way in which competence can vary across issues and
topics—to be discussed further in the next two chapters11—that has less to do
with the expertise of any given individual and more to do with the nature of the
issue or topic. There are some issues or topics that are intrinsically hard in their
and exceeds 0.5. Dietrich (2008) and Owen et al. (1989) provide proofs. See also Boland (1989),
p. 183, building on Hoeffding (1956). The assumption that the mean is bounded away from
0.5 avoids scenarios where the mean converges so rapidly to 0.5 that group competence falls in
group size. See Paroush (1998).
6 For instance, the Non-Asymptotic Result will not obtain if the most competent individual has
competence pc and the following holds:
1
pc1 pci
> ∏ in=2 .
(
1 − pc1
) (
1 − pci )
See Ben-Yashar and Paroush (2000, p. 192) and Nurmi (2002, p. 53), based on Nitzan and Paroush
(1982) and Shapley and Grofman (1984).
7 Here is one example offered by Grofman et al. (1982, p. 687) where that is the case: suppose
there are five voters, whose pc respectively are (1, 1, 0.2, 0.2, 0.2); pc = 0.52 but Pn of a correct decision
being reached by majority rule is only 0.488.
8 Ben-Yashar and Paroush (2000).
9 As noticed, within the CJT literature, by Miller (1986, p. 182).
10 What naturally seems to follow from that fact is that we should construct topic-specific
electorates or topic-specific competence weightings for each elector, to maximize the group’s epi-
stemic performance. On those ideas, however, see our deflationary remarks in Chapter 15, especially
Section 15.4 on the practical irrelevance of competence weighting.
11 Specifically, in Sections 4.5.1, 5.2.1, and 5.3.
own right; they are difficult for anyone to judge correctly. Maybe the evidence
is thin or misleading. Maybe the problem is just terribly complicated.12 With
respect to those sorts of cases, everyone (even putative experts) is less likely
to be correct than everyone is with respect to questions about simple and
straightforward facts to which everyone has easy access.13
There is obviously nothing in the CJT mechanism that requires individuals to
be equally competent across all issues or topics. That is just another ‘assumption
for convenience’ in the classic framework. The CJT logic applies equally well,
issue by issue and topic by topic, to any issue or topic where mean individual
competence is better than random. It applies less strongly—in the sense that
group competence is lower—the nearer mean individual competence is to
random, of course. And it applies more strongly the greater mean individual
competence is above random. So group competence will vary, issue to issue and
topic to topic, as a non-linear function of individual competence issue to issue
and topic to topic. Still, the Asymptotic CJT result will hold with respect to any
topic on wherever mean individual competence is better than random.
The topic-by-topic and issue-by-issue approach does, however, require us to
tread carefully in one respect. The votes must now be independent conditional
on the state and the issue or topic.14 We will discuss a revised independence
assumption along those lines, and the reasoning behind it, in Chapter 5.
3.2 EXTENDING THE CJT TO MORE THAN

T WO ALTERNATIVES
In its classic framework the CJT assumes there are only two alternatives. Some
see that as a serious limitation on the applicability of the CJT, since after all
most real-world choice situations involve more than just two options.15
12 Peter (2016) argues that in such cases decisions should perhaps be made democratically, not
because the majority is particularly likely to be correct either, but merely because no one is. She
argues for a prima facie case for an egalitarian distribution of decision-making power (assuming
a decision has to be made) when there is no one who could reliably pronounce on what the correct
decision would be. Notice, however, that that is not an epistemic argument for democracy (not
even a ‘negative’ epistemic one); her argument assumes that epistemic considerations are not in
play, and that some other non-epistemic principles of a proceduralist or egalitarian sort then
come into play to clinch the case for democratic rule.
13 See Mill’s (1872/1974, bk. 3, ch. 18, sec. 3, p. 539) warning of:
the fallacy of reasoning from a wide average, to cases necessarily differing greatly from any
average. It may be true that taking all . . . together with one another, the opinion of any one of the
judges would be oftener right than wrong; but the argument forgets that in all but the more
simple cases . . . the proposition might . . . be reversed.
14 Dietrich and Spiekermann (2013a,b).
15 Estlund (1997, p. 189) supposes that that makes the CJT’s ‘approach to the epistemic value of
democratic procedures . . . less than trustworthy’. Estlund (2008, pp. 15, 226–8) subsequently
reiterates the same unease, even acknowledging the results in List and Goodin (2001). See similarly:
Extensions 27
This section shows that such fears are groundless, and that the same basic CJT
reasoning can extend to choices among more than two alternatives.
Before doing so, however, we should say that it was not wrong to wonder.
After all, not all results that hold for settings with two alternatives necessarily
hold for three or more alternatives as well. For instance, one of the most famous
results in modern political science—the logic according to which two compet-
ing political parties converge towards one another in policy space—holds for
the case of two parties, but only for the case of two parties. If you introduce any
more parties, parties spread themselves much more evenly across policy space
(as Downs himself acknowledged right from the start).16 So it is well worth
considering whether the CJT might likewise be a very special result deriving
from some peculiarity of the classical two-alternative setup, or whether it can
be generalized to more than two alternatives.17
The short answer is that it can be so generalized. We will here take the plurality
rule to be the natural extension of the majority rule to cases involving more than
two alternatives.18 Under the plurality rule, the winner is whichever alternative
gets more votes than any other alternative. That is not the only way of extending
the majority rule beyond the two-alternative case, to be sure; and Condorcet
himself devoted the second half of his Essai to developing another.19 We show in
Section 3.3 that all the other standard ways of extending the majority rule to
more than two alternatives have the same (or even a little better) truth-tracking
properties as the plurality rule. But the plurality rule is the simplest extension,
and it is therefore the one upon which we will focus for the purposes of this book.
Under the plurality rule with more than two alternatives, the threshold of
the Competence Assumption can be lowered. In the classical CJT framework
concerning cases involving two alternatives, the Competence Assumption is
Farrelly 2012, pp. 14–15; Swift 2014, p. 223. Of course, if we decide k > 2 option cases by a series of
pairwise comparisons among those options, ‘Condorcet voting cycles’ might arise and no option
emerge as the clear winner (Condorcet 1785; Arrow 1963). But much the more natural way of
extending from majority rule in the k = 2 case to the k > 2 case is through plurality rule, as we
suggest in this section. And with the plurality rule no voting cycles can emerge.
16 Downs 1957, ch. 8.
17 Of course you can always force any choice into a two-alternative form (‘φ’ or ‘not-φ’) or into
a series of pairwise choices (Condorcet 1785/1976, p. 51; Shapley and Grofman 1984, p. 337). Still,
we may prefer to leave the choice in its more natural many-alternative form rather than reducing
it to one or more pairwise choices. As Riker (1982, p. 60) says:
Unfortunately, there is no fair way to ensure that there will be exactly two alternatives.
Usually the political world offers many options, which, for simple majority decision, must be
reduced to two. But usually . . . the way the reduction occurs determines which two will
be decided between. There are many methods to reduce the many to two; but, as has long
been obvious to politicians, none of these methods is particularly fair . . . because all methods
can be rigged.
18 The two rules are extensionally equivalent in the two-alternative case of course—but not so
in the case of more than two alternatives.
19 Condorcet 1785.
ordinarily stated as pc > 0.5.20 The natural extension of that to the k-alternative
case is to say that, there, the Competence Assumption requires pc > 1/k. And
that is correct, with one caveat: for the k-alternative case, the Competence
Assumption must also require that each voter has a higher probability of voting
for the correct alternative than for any of the incorrect ones.21 Of course if each
voter in a three-alternative case has pc > 0.5, so much the better; we can be all
the more confident that the alternative chosen by the group is the correct one.
But the jury theorem for three alternatives, for example, holds even if each
voter has pc > 0.34, just so long as each voter has probabilities lower than that of
voting for each of the incorrect alternatives.
If more than half of the voters actually back the same alternative in a vote
among many alternatives (and the other assumptions of the CJT are satisfied),
then we can be all the more confident that that alternative is the correct one. But
demanding an absolute majority is unnecessarily strong. Pluralities suffice.
Even if the winning alternative garners only 34 per cent of the votes in a three-
alternative contest, that can be strong evidence that that alternative is correct in
a large population of competent and independent voters. We call the probability
of a correct plurality vote PnPV.
A bit more formally stated, the theorem generalizing the CJT to cases involv-
ing more than two alternatives assumes that there are n individuals voting by
the plurality rule on k > 2 alternatives x1, x2, . . . , xk, where precisely one of those
alternatives is the correct alternative xc .22 Each voter has a k-tuple of probabilities
〈 p1 , p2 , …, pk 〉 of believing that the respective alternatives x1, x2, . . . , xk are
correct, with pc in this tuple being the probability of voting for the correct
alternative xc (and these probabilities sum up to 1 and are the same for each
voter). The Competence Assumption needs to be revised to:
Multi-Alternative Competence: Each voter believes the correct alternative c
is true with probability pc , and this probability exceeds his probability pe , with
e ≠ c, of believing that each of the incorrect alternatives is true.23
20 As we have said in Section 2.2.1, for both states: i.e. each voter is assumed to be pc > 0.5 likely
to say the alternative is correct if it is indeed correct, and pc > 0.5 likely to say it is incorrect if it is
indeed incorrect.
21 This requirement is of course automatically satisfied in the case of two alternatives where
pc > 0.5. The reason this further stipulation is required for the k > 2 option case is to protect against
scenarios like this: suppose in a three-alternative case each voter assigns p1 = 0.35 to the correct
alternative, p2 = 0.5 to one of the incorrect alternatives, and the remaining p3 = 0.15 to the other
incorrect alternative; then even though pc > (1/k), the plurality vote will nonetheless converge on
the incorrect second alternative.
Note that saying the voter is ‘better than random’ in the k-alternative case is equivalent to
saying pc > pe for every wrong alternative e. ‘Random’ would assign a probability of exactly 1/k
to each alternative. Being ‘better than random’ means assigning a higher probability than that to
the correct alternative and a lower probability to each incorrect alternative.
22 Note, however, that among the incorrect alternatives, some may be ‘more nearly correct’
than others.
23 Note that by assuming that all voters vote for the correct alternative with probability pc , we
implicitly assume that the competence is the same for all possible states, i.e. the competence is the
same no matter which of the k alternatives is the correct one.
Extensions 29
Note that the effect of this assumption is to ensure that individual voter
competence is better than random because pc > (1/k).
The Independence and Sincerity Assumptions remain as before. If all those
assumptions hold, it can then be shown that:
Plurality Vote Reliability Result for k > 2 alternatives: The correct alternative is
more likely than any of the incorrect alternatives to be the plurality winner.
Plurality Vote Asymptotic Result for k > 2 alternatives: As the number of
voters approaches infinity, the probability of the plurality winner being
the correct alternative PnPV converges to 1, so that limn→∞ PnPV = 1.
The proof of this generalization of the CJT is provided elsewhere.24 The basic
intuition underlying it is similar to that in the dichotomous case: the results are
driven, again, by the law of large numbers. For example, if the votes in a three-
alternative contest are statistically independent, and each voter individually
has a 40 per cent probability of voting for the correct alternative and a 30 per cent
probability of voting for each of two incorrect ones, then among a very large
number of such voters the proportions of the votes will be very nearly 40 per cent
for the correct alternative to 30 per cent for each of the two incorrect alternatives.
That would make the correct alternative the plurality winner.25
The generalized version of the CJT for k > 2 alternatives, like the classic CJT,
is expressed as a limit theorem. As with the classic CJT, therefore, one might
reasonably wonder just how quickly the process proceeds, given that the real
world is characterized by far fewer than an infinity of voters.26 In Chapter 2 we
provided Figure 2.1 to give reassurance on that score with respect to the classic
CJT. Here we provide similar figures to give similar reassurance with respect to
the generalized CJT for k > 2 alternatives.
In Figure 3.1, we hold the probability that each voter will vote for the correct
alternative constant at pc = 0.51 and let the number of alternatives k vary. For the
purposes of Figure 3.1 we assume voters are equally likely to vote for each of
the incorrect alternatives, so the probability of each voter voting for each of the
incorrect options is 0.49/(k-1) As we see from Figure 3.1, increasing the num-
ber of alternatives in this way while holding individual voter competence con-
stant makes the probability of the majority being correct converge to certainty
even faster than in the two-option case.
As Figure 3.1 shows, it would certainly be convenient if individual voter
competence were pc > 0.5 even in the k > 2 case, as the classic CJT framework
assumes it to be in the k = 2 case. In that setting, the correct alternative has a
constant expected share of the vote, while the expected share of each of the
24 List and Goodin 2001, Appendix I, pp. 295–7.

25 Of course if we force the trichotomous decision into a dichotomy of the form ‘(correct
option) vs (either of the incorrect options)’, then the correct option will lose by 60% to 40%
(Sunstein 2006c, 37). But that is simply to say that we should epistemically avoid decision procedures
that have us voting over disjuncts in that way. See Section 4.6.2.
26 These calculations were prompted by a challenge from Estlund (2005, p. 610).

1.0
Probability of correct plurality winner PnPV

0.9
0.8 (0.51, 0.1125, 0.1125, 0.1125, 0.1125)
(0.51, 0.1633, 0.1633, 0.1633)

0.7
(0.51, 0.245, 0.245)
(0.51, 0.49)
0.6
0.5
0 20 40 60 80 100
Population size n
Figure 3.1 Probability that the majority of voters, each pc = 0.51 competent (and voting
with equiprobability for all incorrect alternatives), will vote for the correct alternative,
with 2, 3, 4, and 5 alternatives.
incorrect alternatives diminishes as the number of alternatives increases. But it

may well be the case that in choosing among multiple alternatives voters are
less than 0.5 likely to identify the (most nearly) correct among the alternatives
on offer. After all, it is simply more confusing to have so many more options; it
is no surprise that the competence of voters in choosing among all of them
should suffer.27
Here is one way to think about it (we will develop a somewhat more sophis-
ticated approach in Section 8.2.1). In reckoning voter competence, assume that
each voter is 1 percentage point better than random at identifying the correct
alternative.28 In the two-alternative case that means each voter has pc = 0.51, in
the four-option case pc = 0.26, in the ten-option case pc = 0.11.29
27 As Jackman and Sniderman (2002) speculate and Lau et al. (2014, p. 254) demonstrate in
their analysis of 69 elections across 33 countries.
28 That is to say, each voter starts from the Principle of Insufficient Reason and assumes as a
baseline that the probability of each of k options being correct is 1/k. Then he exercises judgement
in such a way as to add 0.01 to that assessment of the probability of the alternative that is correct
and subtract accordingly from the probabilities of all incorrect alternatives. As we show in
Section 4.6.2, this analysis effectively constitutes an answer to the ‘disjunction’ challenge
(Sunstein 2006c, p. 37; Estlund 2008, pp. 228–30, 232–6; Vermeule 2009b, p. 7).
29 Adding a flat one percentage point to each obviously constitutes a larger proportional increase
in individual competence the more alternatives there are. We could instead add (0.01)(1/k), making
the value added a decreasing function of the number of options, as we do in Section 8.2.1.
Qualitatively the same results would emerge, although obviously collective competence over large
numbers of options would then increase much more slowly with increasing numbers of voters than
in Figure 3.2.
Extensions 31
1.0
Probability of correct plurality winner PnPV
0.8
0.6
(0.51, 0.49)
0.4
(0.3433, 0.3283, 0.3283)
(0.26, 0.2466, 0.2466, 0.2466)

0.2
(0.21, 0.1975, 0.1975, 0.1975, 0.1975)
1 x 0.06, 19 x 0.0495
0.0
0 2,000 4,000 6,000 8,000 10,000
Population size n
Figure 3.2 Probability that the plurality of voters, each pc = (1 / k ) + 0.01 competent
and equiprobable to vote for false alternatives, will vote for the correct alternative.
We display such scenarios in Figure 3.2. There we assume that the probability

that each voter will vote for the correct alternative is pc = (1 / k ) + 0.01 and
the probability that a voter will vote for each of the incorrect alternatives is
pe = ( 0.99 − [1 / k ]) / ( k − 1) . In that setup, voter competence decreases as the
number of alternatives increases. As a result, the curves in Figure 3.2 generally
converge more slowly towards 1 the more alternatives there are. (However,
a great many alternatives can lead to a somewhat faster convergence to 1, as in
the case of twenty alternatives in Figure 3.2.) The important messages to take
away from Figure 3.2 are these: (a) no matter how many alternatives there are,
the convergence to 1 will indeed occur with increases in group size; and (b) even
with electorates just the size of a small town (10,000 voters) the probability of
the plurality vote being correct is extraordinarily high.
The interesting thing about the generalized CJT for k > 2 alternatives, then,
is that we can relax that classic CJT Competence Assumption a great deal, and
qualitatively the same pattern found in the classic CJT framework still obtains.
Specifically, we can drop the classic framework’s requirement of pc > 0.5 to require
only that pc > pe for every wrong alternative e.
3.3 USING DIFFERENT DECISION RULES
Throughout the rest of this book, we will be focusing upon aggregating people’s
judgements using the majority rule in binary choice situations. But it is worth
noting in passing that broadly the same desirable epistemic properties are
displayed by a variety of other democratic decision rules for aggregating the

judgements of multiple voters over more than two options.
3.3.1 A Bayesian Parallel
We shall demonstrate this with respect to a range of more standard alternatives

to majority or plurality rule shortly. But let us begin by observing that the
same is true of Bayes’ theorem itself, perhaps the most classic rationality prescrip-
tion within probability theory. Using Bayes’ formula, we start with a ‘prior
probability’ (our own judgement of the probability that some proposition is
true) and we update our probability assessment in light of incoming evidence,
making use of our own assessment of ‘conditional probabilities’ (roughly, the
probability that we will observe the incoming evidence for one alternative,
given that that alternative is actually correct or given that that alternative
is actually false).
Now, Bayes and Condorcet answer slightly different questions.30 With
Bayes, we are asking, ‘What is my subjective probability that φ is true, given
my private signal, the evidence garnered from all other voters, and my prior?’
In the Condorcetian framework, we are asking, ‘What is the probability that
the majority of a group of sincere voters will “track the truth”, that is, vote for φ
if it is actually true and vote against φ if it is actually false?’
Bayes and Condorcet also operate at different levels. Bayes’ theorem adjusts
the beliefs of individuals (in the present application, in light of the reported
beliefs of other individuals). Condorcet’s formula, in contrast, tells us how
much we can trust the outcome of group-level decision processes to be right,
even if our own private beliefs in the matter remain unchanged.31
Still, the two frameworks are convergent as long as the voters truthfully
reveal their private signals. In particular, if one starts with a symmetric prior
of 0.5, and the competence of all voters is homogeneous and known, then
Bayesian updating recommends what Condorcet recommends: follow the
majority view.32
30 Although historically Condorcet was himself inspired by his friend Laplace’s work in effect
extending Bayes’ theorem, and Condorcet saw what he was doing in the Essai as continuous with
that (Gillespie 1972). Laplace and Condorcet were close protégés, historically: Condorcet, imme-
diately upon becoming its permanent secretary, facilitated Laplace’s election to the Academy of
Science; and, when on the run from the Terror, the name Condorcet gave to the keepers of the inn
where he was arrested was ‘Pierre Simon’, the given names of Laplace (Gillespie 1972).
31 It is no ‘paradox’ (pace Wollheim 1962) that someone should think that not-φ is probably the
correct alternative if the majority of independent and equally competent others have sincerely
voted for it, even though her own private signal (and hence vote) was for φ (Swift 2014, p. 206).
32 The parallel is developed more fully in Goodin 2003, ch. 7. Young (1988, pp. 1237–8), Austen-
Smith and Banks (1996), and Goodin and List (2006) all reformulate the CJT in Bayesian terms,
each for somewhat different purposes.
Extensions 33
3.3.2 Other Decision Rules
Looking beyond Bayes, there is a raft of alternative decision rules by which we

might aggregate the judgements of several individuals into group decisions
where there are more than two alternatives. All of the sensible rules are exten-
sionally equivalent, of course, in the case of only two alternatives.33 With more
than two alternatives, however, the plurality rule ignores information about
voters’ views on all but their top-ranked alternative. That is the source of criti-
cism of the plurality rule by Borda, Condorcet following him, and many others
in train.34 They advocate other decision rules that they claim are superior—not
just fairer, but also more accurate and more revealing—insofar as they take
into account information about voters’ views on all alternatives, not just their
top-ranked alternative.
When looking at the ‘complete profile’ of voters’ views in this way (rather
than focusing purely on their top-ranked alternatives alone) interesting
mathematical curiosities emerge. Not least of them is the possibility of intran-
sitivities giving rise to a ‘voting cycle’, such that no alternative is uniquely chosen
by the group under certain voting rules. Condorcet himself became obsessed
with such cycles, halfway through the Essai. Generations of modern social
choice theorists and democratic theorists under their sway have joined him in
that obsession.35
Now, cycles are not the subject of this book. But it is just worth noting in
passing that cycles are much less frequently observed in the real world than in
the mathematical imagination.36 Cycles would indeed be more common in the
context of an ‘impartial culture’, where each possible preference ordering is
equally likely to be submitted by each voter.37 Indeed, in that context, the prob-
ability of a voting cycle actually increases with the number of voters (and also
with the number of alternatives under consideration).
Whereas the ‘impartial culture’ model assumes that each voter is exactly
equally likely to vote each way, the CJT assumes (we think realistically) some-
thing very different—namely, that each voter is more (even if only a little
more) likely to vote for one alternative (viz., the correct alternative) than any
other. And if that CJT assumption is substituted for the ‘impartial culture’
assumption, then cycles are less likely—and decreasingly so, as the number of
voters increases.38
33 With only two alternatives, there is only one pair for the Condorcet pairwise comparison to
compare and only one pair of options for the Borda procedure to rank.
34 Borda 1784/1994. Condorcet 1785/1976, pp. 51 ff; 1785/1994, pp. 120–38. Black 1958.
35 See e.g. Arrow 1963; Riker 1982; Dahl 1989, pp. 142, 145.
36 Feld and Grofman 1992. Felsenthal, Maoz, and Rapoport 1993. Mackie 2003. List 2005b.
Cf. Niemi and Weisberg 1972, pp. 181–270.
37 The term is Gehrlein’s (1983), although the approach is much older (Niemi and Weisberg 1968).
38 How much deviation from equiprobability is necessary to avoid the standard (‘impartial
culture’) result on the probability of cycles? Not much at all. It turns out that ‘the impartial culture
In any case, as we say, our interest in this book lies not with voting cycles but
with the epistemic performance of broadly democratic decision rules. The most
familiar substitutes for the majority rule in more-than-two-option cases are:
Plurality rule: Choose the alternative that is ranked first by the largest
number of voters.
Condorcet pairwise criterion: Choose the alternative that defeats all others in
pairwise elections using the majority rule.
Borda count: Give each of the k alternatives a score of 1 to k based on the
alternative’s ranking in a voter’s preference ordering; that is, the alternative
ranked first receives k points, the second one k − l, . . . , the lowest-ranked
alternative one point. The alternative with the highest number of points is
declared the winner.
Those are the decision rules whose epistemic performance we shall here
be comparing.
We have a particular interest in the plurality rule, of course, because (as we
have indicated in Section 3.2 above) a result very much akin to the CJT can be
proven in respect of that decision procedure for many-option cases. Besides,
it is the simplest rule and its results are the easiest to calculate. We will, for
those reasons, assume the plurality rule as the decision procedure to be used
whenever many-option choices come under discussion in the rest of the book.
The plurality rule is not necessarily the very best truth-tracker among all
these possible decision rules for more than two options. But, as we shall show,
all of those commonly recommended democratic decision procedures—the
plurality rule included—are very good truth-trackers, at least given a reasonably
large number of people each of whom is individually more likely to vote for the
correct option than any of the incorrect ones.
In the calculations that follow, here and throughout the rest of the book, in
the event of cycles or ties we use a random decision to break the deadlock.39
This is not necessarily the standard procedure. But given that we are only inter-
ested in choosing one correct alternative, and since it is useful to make the
results comparable to the results we reported earlier that relied on a random
tie-breaking procedure, this modelling choice is sensible. In any case, politically
assumption can be seen as an extreme limiting case the slightest deviation from which is already
sufficient to circumvent the standard cycling result, provided the electorate is sufficiently large’.
Furthermore, ‘given suitable systematic, however slight, deviations from an impartial culture, the
probability that there will be a cycle under pairwise majority voting vanishes as the size of
the electorate increases’ (List and Goodin 2001, p. 302). Only under highly restrictive (and hence
presumably highly unlikely) assumptions about how probabilities might vary from those of an
‘impartial culture’ will the probability of cycles continue to increase with the number of voters, as
in the standard ‘impartial culture’ result (see further Tangian 2000).
39 Among the decision procedures under discussion, of course, cycles can arise only with the
Condorcet pairwise comparison procedure.
Extensions 35
it is much more likely that a tie or a cycle will lead to some sort of decision—either
because random facts sway the result one way or the other, or because any
outcome is better than nothing so that the parties involved find a way to
break the stalemate.
The question that concerns us here is, as we say, simply how good these three
decision rules are at tracking the truth. To assess that in similar fashion to the
way we have been assessing the plurality rule in the CJT framework, we need
some way of moving from assumptions about the probabilities that each voter
has of choosing each alternative <p1, p2, . . . , pk> to inferences about the prefer-
ence profiles over all the alternatives necessary to apply the Borda count or the
Condorcet pairwise comparison.
There are many possible ways of doing that. But here is one plausible way:
(1) let the probabilities p1, p2, . . . , pk dictate the probability with which each
of the options will be the first-choice option in each voter’s preference
ordering;
(2) once the first preference is fixed, let the relative probabilities associated
with each of the remaining options determine the probability of each
of those options appearing as the second-choice option in the same
preference ordering; and so on.
Having thus generated probabilities of each voter holding each of various pos-
sible preference orderings, we then proceed to calculate the probabilities with
which each option would win under each of the alternative decision rules
under discussion.40
As we see from Figure 3.3, the Borda count and the Condorcet pairwise
criterion somewhat outperform the plurality rule in the scenarios considered.
That is unsurprising, given that the latter two procedures use all the informa-
tion available in the preference profile, while the plurality rule only takes into
account the first-ranked alternative of each preference ordering.
The larger lesson to take away from Figure 3.3 is that all of the alternative
decision rules perform epistemically pretty nearly as well as one another.41
They are all ‘epistemically eligible’, so to speak. Put another way, pretty much
whatever democratic aggregation rule we use (among those that have been
commonly discussed, at least), democracy is a good truth-tracker.
Even though other decision rules do slightly outperform the plurality rule
epistemically, there may be some compelling reasons to go with the plurality
rule nonetheless. One has to do with its sheer transparency. Thus, for example,
40 For more details of this procedure see List and Goodin 2001, Appendix 2. Notice, however,
that unlike in List and Goodin, we do not count ties and cycles as decision failures. Rather, we
break the stalemate by random decision between the tied alternatives.
41 As Young (1988, p. 1239, emphasis in original) writes, elaborating on an insight from Bernard
Grofman: ‘If there are a large number of voters and pc is not very close to 0.5, then the probability
is very high that the truly best [alternative] will be selected by any choice rule. . . .’

0.8
0.7
Probability of correct winner
0.6
0.5
Borda count
0.4 Condorcet winner
Plurality vote
0.3
0 1,000 2,000 3,000 4,000 5,000
Population size n
Figure 3.3 Comparing group competence for three alternatives, voters with pc = 0.34
and equiprobable votes for the other two alternatives, with decision taken by Borda
Count, Condorcet Pairwise Criterion, and Plurality Vote.
we commonly hear claims such as this: ‘A voting method should be relatively

simple and transparent, both for voters and for those calculating the winner. . . .
Simplicity helps explain why plurality voting is so widespread. . . .’42 Borda
begins his classic paper by saying, ‘There is a widespread feeling, which I have
never heard disputed, that in a ballot vote, the plurality of votes always shows
the will of the voters.’43 Its transparency may explain why.
Another even more important factor, however, might be sheer technical
implementability. Condorcet himself acknowledged that his exhaustive pairwise
comparison procedure would be a demanding task, even whilst urging that it
be adopted in the Draft Constitution of the new French Republic.44
42 Levin and Nalebuff 1995, p. 19.

43 Borda 1784/1994, p. 114. Borda goes on to explain why he thinks that is an error, of course.
44 He writes, ‘it is both awkward and time-consuming to form an initial judgment about the merits
of the candidates and difficult to rank a large number of candidates taken two by two and to use this
to deduce a general result would be an immense and lengthy task’ (Condorcet 1793/1994, p. 218).
In an earlier tract, Condorcet (1789/1994, p. 176) gives a sample calculation to show ‘this method
is very time-consuming. To compare just 20 candidates two by two, we must examine the votes on
190 propositions, and for 40 candidates, on 780’; and of course even then the procedure may prove
inconclusive (in modern parlance, there may be no ‘Condorcet winner’). Dodgson’s scheme for
how to adapt the Condorcet procedure to choose ‘rationally’ in the latter circumstances is (unless
the number of voters or candidates is small or structured in a very particular way) probably com-
putationally ‘inherently intractable’: ‘a candidate’s mandate might have expired before it was ever
recognized’ as the winner (Bartholdi, Tovey, and Trick 1989, pp. 158, 161).
Limitations
The Condorcet Jury Theorem, both in its classical formulation and in all
the extensions of it we have discussed, is conditional in form. It tells us that
aggregating many independent votes is highly likely to track the truth—but
only on condition that certain assumptions are satisfied. If they are not, the
aggregated vote may well fail to track the truth (or may even do the opposite).
So let us next consider ways in which the CJT might fail. At the same time, let
us try to assess just how worrying (or not) these potential sources of failure
really are.
Many writers wish to deny the relevance of the CJT, sometimes because it
sits uneasily with some other aspect of their larger theories, sometimes simply
because it seems ‘too good to be true’.1 They usually gesture briskly towards one
or two easy ways of dismissing the CJT, and then move on. The larger aim of
this chapter and the next is to show that, while the CJT genuinely is vulnerable
to many sources of failure, it is nowhere nearly as vulnerable to them as such
brusque dismissals would tend to suggest.
Section 4.1 considers the objection that political choices crucially involve
value judgements which, by their nature, cannot be correct or incorrect.
Section 4.2 considers the possibility that there might be no truth to be
tracked. Section 4.3 considers the possibility that voters might strategically
misrepresent their beliefs rather than reporting them sincerely. Section 4.4
considers the possibility that voter competence might be worse than random.
Section 4.5 considers at length what many regard as the CJT’s most important
premise—the assumption that votes are statistically independent, conditional
on the state of the world—and ways in which that might be violated. Finally,
Section 4.6 looks at the problems arising when the choice situation is systemat-
ically misleading. The debates from these last two sections continue through
to Chapter 5.
1 Estlund 1993, pp. 92–4; 1997, p. 188; 2008, pp. 15 ff. and ch. 12; 2012. Estlund et al. 1989.
Gaus 1997, pp. 149–50. Martin 1993, pp. 142–4, 370–1. Miller 1992, p. 56.
4.1 FACTS AND VALUES
This book is entitled An Epistemic Theory of Democracy. Epistemology concerns

knowledge, and most knowledge is of facts. But according to what has come to
be known as Hume’s Law, facts are to be sharply distinguished from values.2
And of course disputes over values lie at the heart of politics.
The conjunction of those two propositions—that the CJT is concerned with
discovering facts, whereas so much of politics concerns values—is commonly
taken as conclusive grounds for dismissing the political relevance of the theorem.
David Miller, echoing Duncan Black, writes:
Although occasionally a political community may have to decide on some question
to which it is plausible to suppose a correct answer exists (say some scientific ques-
tion in circumstances where there is complete consensus on the ends which the
decision should serve), it is much more likely that the issue will concern compet-
ing claims which cannot all be met simultaneously in circumstances where no
resolution of the competition can be deemed objectively right.3
To those objections there can be four perfectly plausible lines of response. In
what follows we will discuss them in descending order of potency. We will not
argue for any particular one of these responses in preference to any of the
others. Instead, we are inclined to think that any of them (or any combination
of them) would suffice to establish the relevance of the CJT for politics.4
4.1.1 Moral Realism
The first and boldest approach would be to take a moral realist line,5 asserting
that there are indeed facts about what is morally of value, and these facts extend
to the realm of politics. On that line of analysis, political value judgements can
be true or false, and there are facts in the world that make them true or false.6
2 Hume 1739, p. 335.
3 Miller 1992, p. 56. In Black’s (1958, p. 163) words:
Now whether there be much or little to be said in favour of a theory of juries arrived at in this
way there seems to be nothing in favour of a theory of elections that adopts this approach.
When a judge . . . declares an accused person to be either guilty or innocent, it would be
possible to conceive of a test which, in principle at least, would be capable of telling us
whether his judgement had been right or wrong. But in the case of elections no such test is
conceivable; and the phrase “the probability of the correctness of a voter’s opinion” seems to
be without definite meaning.
See similarly Copp 1993; Estlund 1993.
4 For arguments in similar directions see Estlund (1993, pp. 72–80; 2008, ch. 2).
5 Or moral naturalist or cognitivist—whatever their subtle differences, they all converge on the
point being made in this paragraph.
6 As Talisse (2013, p. 512) says, ‘it is precisely our individual commitment to the truth of our
own beliefs on such matters that makes disagreement potentially volatile and uncivil’, which is
why ‘we skirt around [moral] issues for the sake of decorum and politeness’.
Limitations 39
One (now hardly credible) way of making that claim treats morality as a divine
commandment. Another way of making that claim is as a dictate of reason alone.
Some read Kant’s derivation of his Categorical Imperative in that way. Others
take utilitarianism as grounded on natural facts about happiness, as well as
principles of rationality.7 Whether any such form of moral realism leads to an
external standard of correctness for political questions is an additional issue,
of course, but it is not implausible that the answer to at least some political
questions depends on such moral facts.
Obviously, those are bold claims in their own right.8 And even if those claims
were themselves established as true, how well they connect to the CJT is an
open question. That would depend crucially on how reliable voters are in per-
ceiving and tracking such truths.9 But at least, on this account, there is a truth
for people’s views to track, even when it comes to matters of moral values.10
Those may be different kinds of truths than empirical facts about the world,
and we may come by them and hold our beliefs about them in very different
ways than we do our beliefs about empirical facts about the world.11 But that
would not matter, from the point of view of the CJT: all that matters for those
purposes is that there are moral facts and that people are more likely than not
to be right in their perception of them.
7 Smith 1994. Kant 1785. Harsanyi 1982.

8 Here, for example, is Robert Dahl (1989, p. 66) blustering against them:
As to moral propositions, few moral philosophers, and probably not many thoughtful and
educated people, now believe that we can arrive at absolute, intersubjectively valid, and
‘objectively true’ moral judgments. . . . Although some moral philosophers would make such
a claim, they have conspicuously failed to demonstrate the absolute and objective status of
any specific moral judgment. . . . Instead their ‘objective moral truths’ invariably prove to be
highly debatable; their pretense of intersubjective validity cannot be upheld. . . .
9 Dworkin (2009), for example, thinks:
it is extremely doubtful . . . that Condorcet’s theorem has any application at all to moral issues.
The crucial assumption on which his proof depends—that any individual in a specified
group is more likely than not to make a correct judgment—is indeed plausible when the
question is a matter of straightforward fact and perception. . . . But nothing in any plausible
account of how people form moral convictions—which are not a matter of perception or
logic—provides the slightest ground for assuming that people generally are more likely than
not to form correct convictions about controversial moral issues; and history hardly supports
that hypothesis either.
10 In writing ‘nature has connected, by a chain which cannot be broken, truth, happiness,
and virtue’, Condorcet (1795/1796, p. 279) has been widely interpreted—and widely excoriated
(Rothschild 2001, ch. 7)—as taking that view.
11 Mackie 1977. For one example: it would make no sense to speak of ‘forgetting the difference
between right and wrong’, as it would make perfect sense to speak of forgetting some other empir-
ical fact (Ryle 1958). For another example: in the realm of morality, perhaps we should not neces-
sarily be led to question our own conclusions when confronted with countervailing conclusions
of someone whom we regard our epistemic peer, as we should do in other realms. But perhaps all
that is simply to say that moral facts are ascertained differently than other facts about the natural
world. There can be an epistemological difference in that respect without any ontological one. The
moral facts of the world may remain facts, even if they are epistemically accessed in some different
way than other facts (Driver 2013).
4.1.2 Moral Conventionalism
A second approach would draw on a conventionalist account of morality. On that

account, values are social constructs, which vary from one place to another.
While there is no universal and timelessly true fact of the matter about what is
morally Right and Good, there is nonetheless some fact of the matter about what
is to be taken as Right and Good ‘around here’.12 Even if we just socially ‘invent
right and wrong’, once invented the invention is very much present among us,
and there is some fact of the matter about what form that invention has taken.13
That is how philosopher John Searle recommended we ‘derive “ought” from
“is” ’.14 It is not a truth-about-the-world, merely a truth-about-us. Still, such
truths matter, not just socially and politically but perhaps even morally.
Certainly the norms of ‘positive morality’, at least, have the felt force that they
do because and insofar as people in the relevant community think that they are
rightly held accountable to one another under them.15
Obviously, the Condorcet Jury Theorem has a strong bearing on moral norms
of that sort. There is, on this account, some fact of the matter about what are the
moral norms that are in force ‘around here’. Those facts are, on this account,
constituted by what people generally take a critical reflective attitude towards,
of a distinctively moral sort, within the community in question. And while it is
not analytically necessary, it is at least highly likely that people in that place will
be more likely to be right than wrong about that question. If so, when you ask
people—especially a lot of people—‘what are the moral norms in force around
here?’, their collective answer will highly likely be correct.
4.1.3 Moral Separability
Another modest approach would be to accept fully that facts and values
occupy separate domains—but then to make argumentative use of that fact.
Virtually any real-world political decision involves both factual and evaluative
components.16 If facts and values are wholly separate and wholly independent
12 Note well that it is the judgement rather than any social practice (which may or may not
follow from that) that grounds a norm of ‘positive morality’ and differentiates what we perceive
as a ‘moral norm’ from something that is merely a ‘social norm’ (Brennan et al. 2013, ch. 4).
13 Mackie 1977. 14 Searle 1964. 15 Brennan et al. 2013, esp. chs 2, 4.
16 As Landemore (2013b, p. 216) writes:
A lot of apparent ‘value pluralism’—for example the disagreement between Democrats and
Republicans over the legitimate size of the government—can arguably be explained by a
disagreement about facts, including complicated facts such as [the] causal relationship
between big government and efficient spending in a given social and economic context. This
disagreement about the facts of the world leads to a disagreement about political principles
that are dependent on those facts. At the bottom, however, both Democrats and Republicans
may share a common core of basic values.
Limitations 41
of one another, then we can use CJT-style majority rule procedures to get the
factual component of the decision right, without value judgements impinging
one way or another.17
That would work, if we could use one set of procedures to determine upon
what values we should act, but another (the majority or plurality rule) to deter-
mine the facts of the matter concerning what course of action is best designed to
satisfy those values.18 Maybe politics is not all about getting it right factually—
but much of it is.19
Of course, even if facts and values are logically distinct, in practice it may
prove difficult or impossible to prise apart issues of facts and values in such a
way that those two issues can be decided separately. In practice, we may well be
forced to decide on both issues at once.
4.1.4 Moral Majoritarianism
A final and even more modest approach would be to suppose that political
evaluative statements can neither be true nor false according to any external
standard, either in a timelessly true sense that is the same for everyone every-
where or even in the morally conventionalist sense that is the same for every-
one ‘around here’. Instead, on this account, different people have different
values within the community, and different answers are ‘correct’ from each of
those different perspectives.20
Even in that case, insofar as one of those alternatives has to be socially
adopted for the community as a whole, it is democratically desirable to choose
the one that is correct from the point of view of the larger segment of the com-
munity (subject of course to substantive constraints designed to prevent the
majority from becoming tyrannical). We can be assured that that will be the
17 If they are not, and for example values determine our perceptions of facts, there is more of
a problem. The literature on ‘motivated reasoning’ is rife with such examples. See e.g. Kunda
1990; Mele 1993; Kruglanski and Webster 1996; Kim et al. 2010; Bicchieri and Mercier 2013. But
for evidence that motivated reasoning may be more apparent than real in politics, see Bullock
et al. (2015).
18 After the fashion of Hanson (2013)—although his own particular suggestion is to ‘vote on
values’ while letting facts be settled by the operation of betting markets.
19 Lane (1999, p. 214) puts a version of this thought well when she writes:
Insofar as the resolution of [political] disagreement bears on the resolve to act, it requires
practical as well as evaluative reasoning. It is only a practical dimension which can make
sense of the broad and general powers claimed by political authorities in general. . . . Such
practical authority will necessarily have an epistemic dimension. The authority must have
the relevant knowledge of applicable reasons, facts and causal connections; and it must be
able to exercise good judgment, using this knowledge, about what is to be done.
20 The values in question need not be moral values: they might for example be aesthetic values.
case among a large electorate, by reasoning closely analogous to that underlying

the Condorcet Jury Theorem.21
The larger the number of voters, the more certain it is that the majority
(or plurality) among them will track what is true from the point of view of
the largest segment. That is a different guarantee than the classic CJT offers
where it is one and the same truth that everyone’s vote is supposed to be
tracking. Still, it is a valuable guarantee, from what we will call a ‘democratic-
epistemological’ perspective.
Notice that even if facts and values are intermixed in practice, a version of the
separation strategy might work in combination with moral majoritarianism.
Insofar as factual issues dominate the discussion, the democratic majority is
likely to be right about those facts for standard CJT reasons. And insofar as value
issues dominate the discussion, at least we can be confident that the outcome
that is chosen will be correct from the point of view of the values that are most
commonly held across the community.
By and large over the course of this book, we will be presuming one of the
other responses (moral realism, moral conventionalism, or moral separability)
rather than moral majoritarianism. But we will explore the implications of the
latter explicitly with respect to differences among voters deriving from their
differing interests (in Chapter 14), values, or priorities (in Chapter 13).
4.2 THERE MIGHT BE NO (SINGLE) TRUTH

AMONG THE ALTERNATIVES
The CJT tells us that, so long as its assumptions are met, the majority among a
large number of independent votes will settle upon the truth—if there is a truth
to settle on. But therein lies another rub. There may be no truth of the matter,
or the true answer might not be on the agenda for choice, or there might be
multiple truths among which competent voters might divide, or there might
not be any voter-independent fact of the matter with the truth being different
for different voters. If the assumption that there is some unique truth to be
found is confounded in any of these ways, then the aggregation procedures at
the heart of the CJT might fail to track the truth.22
21 Assuming each voter votes on the basis of his perception of what is the correct outcome in
terms of his own values, priorities, or interests (and not those of the majority). If some do one and
others the other, nothing coherent can be inferred (Wolff 1994). See further Section 4.6.2 below.
22 Even if there is a single, truly correct alternative to be found, if there are too many alterna-
tives under consideration for boundedly rational individuals to consider all at once, they still
might not settle on the truly correct alternative even if all the other conditions of the CJT were
satisfied (Condorcet 1789/1994, p. 172). We present models taking this factor into consideration in
Chapters 8 and 12.
Limitations 43
4.2.1 The Question Is Not Truth Apt
The first worry is that the question under discussion might simply not be
truth apt. It might simply not be something about which there can be any
truth of the matter.
We already discussed one version of this worry in Section 4.1, in relation to
the alleged fact/value dichotomy. There, the claim was that value judgements
might not be truth apt.
Similar questions arise when the issue is truth-tracking with regard to pref-
erences. There is an important difference between saying ‘I prefer X’ and the
proposition ‘X is good (even just for me)’. When stating a preference, you are
not expressing a judgement that can possibly be true, one way or the other. You
are merely stating your preference, and that is the end of the matter. So too
when making a demand, or issuing a command, or expressing a wish. None of
those have propositional content that can take on truth values.
Notice, however, that even if people’s preferences and wishes are not them-
selves truth apt, true statements can nonetheless be made about what people
wish and what they prefer. There may well be some good social purpose to be
served by finding out the truth of the matter about what most people wish or
prefer; and aggregating independent judgements about what that is can be a
good way of doing so.23
4.2.2 The Right Answer Is Not on the Agenda
In the classic CJT framework, and virtually all variants on it, the alternatives
among which voters choose are exogenously given.24 Even if there is one
‘correct’ alternative, voters cannot choose it if it is not on the menu of options
among which they are supposed to choose.25 That is another way that the CJT
might fail.
Of course, the CJT does not fail altogether in light of this consideration. It
remains the case that, if the correct alternative is among the options available to
be chosen, the majority of a large number of independent voters is highly likely
to vote for it. In that case, the probability that what a majority of a large numbers
of voters vote for is the correct option is that probability calculated in CJT fashion,
times the probability that the correct option is among the options for choice.
Depending on the latter, that might still be quite a high probability. In any
case, the probability that the majority of many voters will be correct will in the
23 We have already discussed that briefly in Section 4.1.4, and we will go on to say more about
it in Chapter 13 and Section 14.1.
24 We suggest ways of altering that in some of the models discussed in Sections 8.1 and 9.2.5.
25 Estlund 2008, p. 15; Fuerstein 2008; Goodin 2008, pp. 122–4.
classic CJT framework be invariably higher than the probability that any one of
them individually will be correct, when confronted with the same menu of options.
But even that is a little too concessive. True, the classic CJT framework is
described in terms of two alternatives, one of which is correct and the other of
which is incorrect. And even the extended CJT framework we have suggested
for dealing with more than two alternatives is most naturally taken to assume
that one of those alternatives is correct and the others are all incorrect. But note
that this is less demanding than it might first appear because at least sometimes
incorrectness can come in graded form.
Suppose that the truth to be tracked within the CJT framework is the truth
about which of the alternatives is ‘better’ than the others on the agenda.
(Condorcet himself was thinking in those terms, as the second half of his Essai
goes on to show.26) In that case, the object of the CJT exercise is not really one
of finding the needle in the haystack of the ‘one truly correct option out there
in the world’. Rather, the object of the exercise is then to select the best alter-
native among the alternatives offered for choice. Hence, even if the very best
option is not among the set of alternatives offered to voters, that in no way
undermines the CJT claim, reformulated to say that the majority is highly likely
to choose the best alternative among those under consideration.
4.2.3 There Are Multiple Truths
In both the classic CJT framework and our more-than-two-alternative refor-

mulation of it, there is an assumption that there is only one correct (or best)
alternative among the options available for choice. But what if there are multiple
equally good alternatives, none more nearly correct than any of the others? That
is to say, in terms of the ‘betterness’ relation just discussed, they are tied.
Here is how that possibility might upset the CJT result. Suppose there are
four alternatives. Suppose that three of them are coequal best among those on
offer, with the fourth alternative being worse than any of those top three.
Suppose each voter is pc = 0.6 likely to vote correctly, which in this case means
for one of the three coequal best alternatives. But since each is as good as both
others, voters divide randomly among those three coequal best alternatives.
Among a large number of such voters, each of the coequal best alternatives
would garner roughly 20 per cent of the votes, while the remaining ‘worst’
alternative would garner 40 per cent.27
How worrying is this scenario likely to be? Well, that just depends on how
many coequal best answers there are on the agenda, compared to how many
26 Condorcet 1785/1976, p. 52.

27 This is arguably what happened in the 2016 Republican primary elections in the US, where
there was a large majority for ‘anyone but Trump’, but that anti-Trump vote was split among several
alternative candidates (Maskin and Sen 2017).
Limitations 45
‘clearly incorrect’ ones. In the numerical example just given, there were three
ways to be right and only one way to be wrong, and that is what drove the result
there. But it seems likely that there are usually a great many more ways to be
wrong than to be right (even in those cases, themselves perhaps rare, where
there are multiple coequal best alternatives).
Assuming that is the case—and assuming that the items on the agenda for
decision are a roughly random selection of all possible answers to the question,
right and wrong in rough proportion to the proportion of each in the pool of all
logically possible answers—then wrong options will typically far outnumber
right ones (even if there is more than one that is coequally right). In that case,
the problem posed by our toy example is no problem.
But randomness, of course, assumes that there has been no manipulation of the
agenda. Perhaps that assumption is a bit heroic, in real-world politics. Notice,
however, that a very particular pattern of agenda manipulation would be required
for the problem posed in our toy example to arise. The manipulator of the agenda
would have to restrict radically the number of wrong items on the agenda, and to
maximize the number of coequal best ones on the agenda, for the trick in our toy
example to work. Only in that way could a manipulator who is trying to install his
own preferred wrong decision as the social decision succeed in that task.
Success in that strategy presupposes that advocates of error are univocal in
supporting the same wrong answer. It further presupposes that advocates of the
truth will not be able to detect any difference at all among the options that the
manipulator hopes that they will regard as all coequally best. And it finally
presupposes that, even if advocates of the truth regard multiple options as
coequally best, they will not be able to coordinate in support of any one of them
but will instead split their votes roughly equally among all of them.
While it is logically possible that all of those conditions might occasionally
prevail, it seems highly unlikely for all of them often to do so. Very much more
often, even if there are multiple coequally best options on the agenda, and even
if advocates of the truth distribute themselves randomly among them, there
will be vastly more clearly incorrect items on the agenda for decision that will
be clearly defeated by one or another of them. And notice that if they really are
coequally best, it simply does not matter which of them prevails—just so long
as one of them does.
4.3 FAILURES OF SINCERIT Y: STRATEGIC VOTING
The CJT works if, and only if, each voter with better-than-random competence
votes for the alternative that he sincerely believes to be the best among the
options available. That is the Sincerity Assumption of the classic CJT, introduced
in Section 2.2 and unmodified by anything we have said so far.
4.3.1 Beliefs versus Preferences
Economists and political scientists are familiar with many reasons for thinking
that that may not be the case. There are many settings in which voters can, by
strategically misrepresenting their own sincere preferences, secure an outcome
closer to the one that they genuinely prefer.28 In an early round of a multi-stage
election, they might strategically vote for someone whom they regard as a ‘bad’
candidate in order to set up their preferred candidate with an easy opponent for
the next round. Or in a constituency where their preferred candidate has no real-
istic chance of winning, they might strategically vote for their next-most-preferred
candidate in the hope of blocking a victory by their least-preferred one.29
All of that occurs in the realm of voters who are seeking to satisfy their
preferences. They have a preference ranking over the possible outcomes of the
voting, and they want to secure an outcome as high up their own preference
ranking as possible. In a CJT-like setting, in contrast, voters are pooling their
judgements to determine what the truth is of some matter before them.30
In that setting, each does indeed have his own beliefs on that matter. It is
these beliefs or judgements that are being pooled. But it is not clear why people
should have ‘stakes’ in their own beliefs prevailing, in any way analogous to the
stakes that they have in their preferences prevailing.31 Indeed, the whole reason
they are engaged in the pooling exercise is that they know that their own beliefs
might be fallible, and that the judgements of their epistemic peers are broadly
speaking as reliable as their own.32
28 Furthermore, no decision rule can prevent that possibility in general (Gibbard 1973;
Satterthwaite 1975). Approval voting (Brams and Fishburn 1978) is an exception only in the
dichotomous-choice case.
29 Farquharson 1969; Downs 1957, p. 47. Condorcet (1789/1994, pp. 174–5) contemplates both.
As a solution to the former problem, he recommends procedures that ‘enable the election to be
decided by a single ballot’. Apropos the latter he asks:
Should we have any hesitation in totally condemning a method which almost always offers
an honest man only the choice between either not voting in accordance with his conscience
or wasting his vote; which makes it our duty to vote for and pronounce as the most worthy,
not the candidate we consider deserving, but the least unworthy of those between whom the
factions force us to choose?
30 Arrow (1963, pp. 94–5) rightly observes that in the first half of Condorcet’s Essai in which the
CJT appears ‘voters are judges of some truth rather than expressing their own preferences’, which
is importantly different to the social-choice style discussion that occupies the second half of the
Essai and with which classic discussions of strategic voting are concerned.
31 Contrary to the central assumptions in classic ‘strategic models of talk in political decision
making’ (Austen-Smith 1992). People may sometimes have preferences over beliefs, as when
engaging in ‘wishful thinking’ (Elster 1983): but that is a cognitive error, which rational actors
(which strategic voters are paradigmatically supposed to be) will abandon once they notice it;
however much a person wishes something to be true, above all he wants his beliefs about its truth
to be true.
32 On average, if not on a case-by-case basis. Christiano (1995, p. 408) suggests that a voter who
knows that what she thinks is the correct answer is unlikely to garner sufficient votes to win might
Limitations 47
Of course, that is not true in our extended applications of the CJT to cases
where people have different preferences, interests, values or priorities.33 There,
what majority voting successfully tracks is not a single truth that is the same for
everyone but rather the answer that is correct from the point of view of the
preference, interest, values, or priorities of the larger number of people in
the community. In that sort of setting, there may indeed be incentives for people
to strategically misrepresent their beliefs when voting. But in the classic CJT
setting, the only potential reason to misrepresent one’s own belief is not in
order to help it prevail, but in order to increase the chance of the group getting
it right collectively, as we show in the next section.
4.3.2 Game Theoretic Complexities
The standard discussion of strategic voting in the CJT literature is game-theoretic

in form.34 The central insight there is that voting in line with your sincere belief
is not necessarily a Nash equilibrium. In particular, even if all voters are interested
in making a collectively correct decision, there can be incentives to unilaterally
deviate from the situation in which everyone votes sincerely according to their
own beliefs.
To see why, take the case of a group (of odd size) making decisions by
majority rule.35 A voter will realize that his vote will make a difference—will be
decisive in determining the outcome—only if the rest of the group’s votes
are exactly tied. He should therefore vote as he would if a tie were to occur
(because in all other cases his vote doesn’t matter). And the hypothesized fact
that everyone else’s votes are exactly tied serves as information in itself that
might affect the way that the voter votes.
Suppose, for proof by contradiction, that all voters vote in line with their
sincere beliefs. Consider one of these voters. For some reason that voter thinks
as follows: ‘If φ is true then that will be pretty much evident to all, and the vast
majority will vote for it; whereas if not-φ is true, then that will be less obvious
and only a narrow majority will vote for it.’36 Suppose now the voter’s own
have a strategic incentive to vote strategically so as to increase the chances that some alternative
she thinks best will win. But that would make good epistemic sense only if the voter thought that
she personally was more competent than the group as a whole (and of course if she was fairly
confident about the second-best alternative having a realistic chance of winning).
33 As in Section 4.1.4 and Chapters 13 and 14.
34 Owing to Austen-Smith and Banks (1996), Feddersen and Pesendorfer (1998), and a raft of
others in their wake.
35 This is the sort of case envisaged by Austen-Smith and Banks (1996). Feddersen and
Pesendorfer (1998) focus on the case of the unanimity rule, and show that strategic voting might
distort outcomes even more in that case.
36 Effectively assuming different competence levels for φ and not-φ. Another possibility is that
the voter starts with a higher prior probability for not-φ (Austen-Smith and Banks 1996). Or perhaps
private judgement is that φ is true, so voting sincerely he would vote for φ. But
thinking of the situation in game-theoretic terms, and focusing purely on the
case in which everyone else’s votes are tied, the fact of that hypothesized tie will
lead the voter to vote not-φ instead. And sincerely so, in a way: not-φ is indeed
what the voter truly believes most likely to be true, in the very special case (the
case in which everyone else’s votes are tied) that strategic thinking instructs the
voter to focus upon (assuming of course that all other voters vote sincerely, to
which we will come in the next paragraph).37 This shows that voting sincerely
is typically not a Nash equilibrium: it can be attractive to disregard one’s own
signal once one conditionalizes on the fact that all other voters are tied. Yet
that is true not because voters want to promote their own judgement, but—
surprisingly—because they want to get the collective decision right.
Of course, it would be minimally problematic if a single voter reasoned in this
fashion—that is unlikely to change the majority decision. The problems arise
only if a great many, or indeed all, voters think strategically in this way. But now
contemplate what would happen if everyone did likewise: in that case, the
hypothesized ‘fact of a tie’ would provide none of the same straightforward
informational content as in the previous case.38
Universal sincere voting is ordinarily not a Nash equilibrium, as we have seen.
Universal strategic voting in the fashion described is typically a weak Nash
equilibrium.39 But it is one that voters are not likely to remain in, because it is
collectively uninformative and the reasoning towards it is self-undermining.
Hence the would-be strategic voter, in trying to figure out what the hypothesized
‘fact of a tie’ would mean for how he should vote, has to make some complicated
surmises about the strategic choices of his fellow voters, as well.
Were we to continue down the game-theoretic path, we might attempt to
construct special ‘knife-edge’ cases in which sincere voting is a strong Nash
equilibrium. But those would rely on very specific assumptions (the knife-edge
cases are rare), so that approach is not very robust.
he thinks that the conditional probability of φ being true, given a tie, is substantially lower than
the conditional probability of not-φ being true, given a tie.
37 Different thresholds of reasonable doubt could lead to strategic misrepresentation that is not
aligned with beliefs about what is correct but about what is best: perhaps deciding for φ requires
crossing a higher threshold of reasonable doubt, so that a tie makes not-φ the safer bet. In that
latter case, however, the voters do not vote for the most likely alternative but for the one with the
highest expected utility.
38 It provides some information, but not information directly about what others think is the
correct alternative, as when each voter could count on all the others voting sincerely rather than
strategically. When others vote strategically, the information a voter could surmise from the
hypothesized ‘fact of a tie’ is at most information about what other voters would themselves do
were it a tie.
39 It is a Nash equilibrium in the sense that, if everyone behaves that way, no one can unilaterally
do better by defecting. But it is only a weak Nash equilibrium in that it is completely pointless to
behave that way—one might as well just randomize or follow one’s own opinion.
Limitations 49
It is probably wiser, at this point, to step back from complicated game-theoretic

reasoning altogether, if we are to construct a credible model of voter behaviour
in the real world. After all, voters are limited in terms of information and cog-
nitive abilities—much too limited to engage in complex strategic calculations
of this sort. Since such Nash equilibria as do exist are not easy to understand or
anticipate, they are not very likely to emerge under realistic conditions, even if
people were trying to reason strategically.
4.3.3 Sincerity as a Default
Given the complexities of strategic reasoning, and their limited capacities to

engage in it, voters much more often than not will fall back on the default rule,
‘Vote sincerely’.40
They might adopt that as a default for normative reasons (‘truth-telling is
good’). They might do so for the sake of speaking their mind (‘what I really
think is true’). They might do so out of sentiments of solidarity with the
group of people who, together, are pooling their judgements in search of the
truth.41 Or they might do so to avoid the reputational damage, among their
smaller circle of friends and colleagues, of getting caught in lies. All of those
considerations converge on ‘voting sincerely’ as the uniquely natural default
rule. Since strategic thinking provides no clear, unique alternative guide to
behaviour, even for voters who try their best to engage in it, the default rule
rules by default.
Notice one final fact that bolsters that conclusion. In any large electorate
the chance that any single voter will be decisive is vanishingly small. The game
theorist is right to say that, however small its likelihood, that is nonetheless the
only scenario upon which a strategic voter should focus in deciding how to
vote. That is the only time his vote will make a difference. But if the chances of
that scenario occurring are vanishingly small, the chances of the voter’s making
a difference with his vote are also vanishingly small. If any of those other reasons
for action just mentioned weigh at all heavily with the voter, then he should
rationally act on those other bases instead—because none of those other ends in
view depend on the outcome of the election, in the way that voting strategically
correctly to change the outcome of the election does.42
40 Taking advantage of special features of German parliamentary elections, Spenkuch (2017,
p. 22) demonstrates that ‘large, democratic elections are characterized by a “sincerity bias”, but
that voters can and will behave strategically if the situation demands it’. More specifically, he finds
that (in those elections, anyway) ‘at least two thirds of individuals’ do not vote strategically.
41 Sunstein and Ullmann-Margalit 2001. On this, see further Chapter 14.
42 Goodin and Roberts 1975; Brennan and Lomasky 1993.
4.3.4 Deliberation Induces Sincerity
Finally, there is experimental evidence to suggest that, while ‘in the absence of
communication individuals behave strategically in much in the spirit of [game]
theoretical jury models’, that is not the case where there is an opportunity for
members of the group to communicate with one another before voting. In that
case, ‘messages are public and truthful, [and] they are a powerful determinant of
the collective choice’.43 This effect of deliberation in dampening strategic behav-
iour is acknowledged even by some of the principal authors of the ‘strategic
voting’ challenge to the Condorcet Jury Theorem.44
Of course, not all group decisions are preceded by free and open communi-
cation among all members of the group. And in some sorts of groups (very large
ones, for example), perhaps it cannot be. So these experimental findings, even
if true, do not constitute a guarantee that strategic behaviour will never occur.
They merely provide some reassurance that there is a way to design decision-
making procedures—by making them more genuinely deliberative—that will
serve to discourage it. Furthermore, given all the other epistemic advantages
we adduce for deliberation in Chapter 9, that is an institutional design that
epistemic democrats will have many reasons for embracing.
4.4 FAILURES OF COMPETENCE
The assumption that voters are of better than random competence is rightly
seen as absolutely crucial for delivering the classic CJT’s optimistic result.
To see why, just go back to the classic CJT framework. Suppose there is a
large number of independent voters charged with deciding between two
alternatives, one of which is correct and the other incorrect. In the classical
CJT, it is assumed that each voter is of better than random competence. But
suppose instead that each voter is only 45 per cent likely to vote for the cor-
rect alternative, and 55 per cent likely to vote for the incorrect one. Further
suppose that the other assumptions of the CJT were met. Then the same
law-of-large-numbers reasoning that drives the classic CJT in the optimistic
direction would drive us in a highly pessimistic direction. On this scenario,
the correct option would lose (garnering only a roughly 45-to-55 share of the
votes in a large electorate).
43 Goeree and Yariv 2011, p. 919. This finding is of a cloth with earlier experimental findings
that prior discussion even induces cooperative behaviour in Prisoner’s Dilemma situations, where
non-cooperative behaviour is the strictly dominant strategy (Orbell et al. 1988).
44 Austen-Smith (2015, pp. 75–6), in effect recanting his earlier analysis (Austen-Smith 1992).
Limitations 51
0.5
0.4
0.3
0.2
0.1 pc = 0.45
0.0
0 100 200 300 400 500
Population size n
Figure 4.1 Probability that the majority of voters is correct, if voters are individually
pc = 0.45 likely to be correct, for electorates of varying sizes.
Furthermore, by that same law-of-large-numbers reasoning, that result would

be increasingly probable the larger the number of independent incorrectly
inclined voters there were—tending towards certainty that the result will be
wrong as the number of incorrectly inclined voters approaches infinity.45
Indeed, the probability that the majority of incorrectly inclined independent
voters will favour the incorrect alternative increases quite quickly with increases
in the number of such voters, as seen in Figure 4.1.
The same is true, by parity of reasoning, in a k > 2 alternative context. If each
independent voter is most likely to vote for some wrong alternative (the same
wrong alternative for all voters), then the more such voters there are the more
nearly certain it will be that that wrong alternative will win and the correct
alternative will lose. Hence, in both the classic CJT framework and the k > 2
alternative generalization offered in Section 3.2, it is absolutely essential that
each voter’s probability of voting for the correct outcome exceeds his probability
of voting for any wrong one.
45 Many have made this observation, starting with Condorcet (1785/1976, p. 49) himself:
If, on the contrary, the probability of the judgment of each voter is below ½ (i.e., if it is more
probable than not that he will be mistaken) then the more the number of voters increases, the
more the probability of the truth of the decision diminishes. The limit of this probability will be
zero. . . . If the probable truth of each voter’s vote is ½, then, whatever the number of voters, that
of the truth of each decision will also be ½.
We might well wonder what reason we have to worry on this score.46 After
all, if people vote purely at random, their probability of voting for each of
k alternatives would be 1/k. Even if voters are not much better than random,
surely it requires some pretty special story to explain how they—not just some
of them, but all or even just a majority of them—can manage to be worse than
random. ‘There must’, as Condorcet says, ‘be a reason why he decides less well
than one would at random.’47
There are clearly good grounds, of a procedural sort, for thinking that people
are at least a little better than random at tracking the truth. Those rest on the
fact that people are trying to form correct views about the state of the world,
both in order to advance their own private projects and in interacting with
others (in the cooperative game that is interpersonal conversation, believing
the proposition you assert to be true is precondition of asserting it).48 We can
observe people, accordingly, employing various procedures of investigation
and reasoning that, while they may of course fail on any given occasion, are
generally likely to lead them to reach correct conclusions.49 No special case,
beyond that observation, is required for thinking that people are probably on
average at least a little better than random in their reports about any matter.50
46 Estlund (2008, p. 228; see similarly Sunstein 2006c, p. 37; Vermeule 2009b, p. 7) offers a
‘disjunction’ worry that, while not giving us any specific reason to doubt that people are worse
than random, is designed to engender some concerns over the robustness of the CJT’s compe-
tence assumption across different ways of specifying the options. We discuss that concern in
Section 4.6.2 below.
47 Condorcet 1785/1976, p. 62.
48 It does, at least on Grice’s (1975) ‘cooperative principle’ of language use. Similar views have
been put forward by: Habermas 1990; Misak 2000; and Talisse 2005, ch. 6; 2009, chs 3–4.
49 Estlund (1993, p. 99, n. 44) remarks:
It may seem hard to imagine the average competence on any question being lower than [.5],
but it is actually quite possible. People use methods, principles and previous experience to
answer such questions. If they did not, they would be no better or worse than random. But
these very factors that make it possible for humans to be better than random also allow them
to be worse. The possibility of systematic correctness brings with it the possibility of systematic
error. If our methods or principles or experiences happen to be incorrect rather than correct,
we will be less competent than a coin flip.
Yes, it is possible that the methods we have chosen for discovering the truth might fail on any given
occasion. But possibility is not the issue, probability is: those procedures have been chosen precisely
on grounds that they are good epistemic guides, and would presumably be abandoned for those
purposes if they systematically proved to be otherwise.
50 Estlund (1993, p. 93; see also 2008, p. 16) insists ‘we don’t know whether [average individual
competence on some relevant class of social question is above .5] or not’. More precisely, the
problem he sees is this:
The Condorcet Jury Theorem gives us no epistemic mileage unless we can first publicly
establish in a way that is unreasonable to deny, that the average individual competence is
above .5. . . . [B]ut it is hard to see how such a thing could be established without inde-
pendent public knowledge of the answer key – the very facts we hoped to use democratic
voting to reveal. This . . . is a grave difficulty for the Condorcetian conception of voting as
strongly epistemic.
Limitations 53
What requires some special explanation is how people could be worse than
random, even averaging across a large number of people.51
The argument for thinking that the competence of large numbers of people
is worse than random is typically couched in terms of some systematic biases
that affect many people at once, misleading them all to vote for some alterna-
tive other than the correct one. Condorcet himself supposed that the masses
were afflicted by ‘great ignorance with many prejudices’, and he recommended
that they exclude themselves from social decision-making on those grounds.52
Sheer ignorance, however, provides no reason to think that people would be
worse than random on average (at least given a large number of people). It takes
something like widely shared prejudices and the systematic biases that they
introduce into many people’s judgements at the same time and in the same
direction to drive individual competence below random.53
Shared prejudices—racism, sexism, and such like—are all too common, of
course. So too are shared biases. Some derive from psychological heuristics that
are widely shared. Others derive from values that people share, which (through
‘motivated reasoning’) leads them to similar beliefs that they all wishfully believe
to be true given those shared values.54
But we do not have to know the right answers to the question at hand to know whether someone
is employing procedures that are generally likely to find the right answer. Ironically, Estlund
(2008, esp. chs 6, 9) himself depends on this very fact, in his own argument for democracy in
terms of ‘epistemic proceduralism’.
51 Thus, Estlund (1997, p. 186) is simply too glib in asking:
Why ever substitute the outcome of majority rule for one’s own moral judgment, if all that is
required in order to stick with one’s own judgment is to believe that the voters must probably
have been, on average, worse than random? A voter has no more solid basis for the probabilities
that the theorem requires than she has for her . . . judgment that the outcome of the voting
procedure is . . . mistaken.
If a lot of people who have been following reliable procedures for discovering the truth say
something different, that in itself counts as a solid basis for thinking that their view is right
and yours is wrong. (The ellipses in the quotation replace the words ‘moral’ and ‘morally’, but
in context it is clear that Estlund does not mean to deny that moral judgements can take on
truth values.)
52 Condorcet 1785/1976, pp. 49, 62. Note well, ‘exclude themselves’ rather than ‘be excluded’: as
Condorcet (1789/1994, pp. 107–8) says elsewhere, ‘In general every citizen has the right to elect
men to any public function. Any other election or nomination is only legitimate with their con-
sent. But this is one of the rights which they can ask their delegates to exercise for them.’
53 As many have said, starting with Condorcet (1785/1976, pp. 49, 62) himself. Mill (1872/1974,
bk. 3, ch. 18, sec. 3, p. 539) warns similarly of ‘some common prejudice or mental infirmity, [which]
if it acted on one judge, would be extremely likely to affect all the others in the same manner, or at
least a majority, and thus render a wrong instead of a right decision more probable, the more the
number was increased’. See also Waldron (1989, p. 1323), Sunstein (2006c, pp. 34–6), and Estlund
(1993, p. 99, n. 44; 2008, p. 16).
54 Kunda 1990. Mele 1993. Kruglanski and Webster 1996. Kim et al. 2010. Bicchieri and
Mercier 2013. Flynn, Nyhan, and Reifler 2017. Some experiments show, however, that only modest
incentives to ‘get it right’ break the chains of such motivated reasoning (Bullock et al. 2015).
Notice that shared prejudices or biases are, in essence, failures to satisfy the
CJT’s Independence Assumption. We will discuss violations of the Independ
ence Assumption in the next section, continuing through the next chapter.
But there is one crucial thing that bears emphasizing at this point in our
discussion.
If voters’ incompetence derives from a violation of CJT’s Independence
Assumption, then the CJT provides no reason to think that the CJT will ‘go into
reverse’ and group competence will asymptotically approach zero as the number
of voters who are incompetent in that way increases. On the contrary, the
probability that the voters are collectively right then simply flatlines at the prob-
ability that the common cause of all their votes is not misleading, as we will show
below (informally in Section 4.5.2, and more formally in Sections 5.3 and 5.4).55
4.5 FAILURES OF INDEPENDENCE
Much of the concern over the applicability of the CJT results to the real
world centres on its Independence Assumption. It is not hard to see the
cause for concern.
4.5.1 Worries over Independence
Mathematically, the Independence Assumption is absolutely crucial to the

working of the CJT. Suppose it is not satisfied. Suppose, for example, that there
is one super-charismatic opinion leader, who announces his voting intention
before every election; and suppose that each of the other million voters faithfully
follows his lead and votes exactly the same way as the opinion leader does. When
aggregating those votes, we are not pooling the independent information of
1,000,001 voters. Instead we are simply counting the opinion leader’s view
1,000,001 times. It is not a case of ‘many minds’—it is rather a case of ‘one mind
and many mimics’. And the probability that the majority among those 1,000,001
voters will be correct is identical to (and not more than) the probability that the
single opinion leader is himself correct.
55 For an example of this error, see Brennan (2014, p. 36): ‘If citizens are systematically mistaken,
then by definition their errors are not randomly distributed, and so the so-called miracle of aggre-
gation does not occur. . . . [T]hen their mean competence is less than 0.5, so . . . the probability that
democracy will get the wrong answer approaches 1.’ But that occurs only if voters are independently
incompetent. If they are systematically incompetent, i.e. if their incompetence derives from the
systematic influence that some common cause exerts, then the probability of the majority among
a large number of such voters being wrong converges to the probability that that common cause
is wrong—no more, no less. See further Sections 4.5.2, 5.3, and 5.4.
Limitations 55
The case of voters slavishly following some super-charismatic opinion leader,

of course, is an extreme example of the failure of the Independence Assumption
to hold. But there can be other less dramatic common influences that shape
all (or a great many) voters’ opinions in the same way. Here is a short list, which
considers the dependence of people’s opinions:
• on the same shared opinion leader (as just described);
• on the same shared ideology (Condorcet’s ‘prejudices’; Gramsci’s ‘dominant
ideology’);56
• on the same shared psychological mechanisms (‘heuristics’ and consequent
‘biases’);57
• on the same shared cues (precedent in courts; social dynamics leading to
cascades, more generally);58
• on the same more fundamental shared properties (such as ‘a common
social background’, or ‘common training’59);
• on the same shared evidence, background information, or theories; and
• caused by following other voters, so that they influence each other directly
(e.g. by voting exactly as one’s peer has voted).
Under such conditions the CJT’s assumption of statistical independence among
the votes will be violated. Just how damaging that violation is to the CJT’s results
is a separate question, to be assessed subsequently in this chapter. For the moment
let us simply say this. It would be wrong to jump to the conclusion that, just
because the Independence Assumption is violated, ‘all bets are off ’. While it is
true that some severe violations of Independence lead to the complete collapse
of any plausible jury theorem, if the violations of Independence are not so
severe then jury theorems with weaker constraints on independence still hold,
as will be shown in Chapter 5.
4.5.2 The Relation between Competence and Independence
It is important to be clear about the difference between the CJT’s Independence

and Competence Assumptions. Both are statistical assumptions about the
votes, but their content differs. Independence is a claim about the relation of
the votes to one another,60 while competence is a claim about the probability of
individuals each voting correctly in relation to the state of the world. One reason
56 Condorcet 1785/1976, pp. 49, 62. Gramsci 1971; Abercrombie et al. 1980.

57 On which see Kahneman, Slovic, and Tversky (1982) and many in their wake.
58 We discuss these in Chapter 10. Cf. our discussion of cue-taking more generally in Chapter 12.
59 Vermeule 2009b, p. 6.
60 More precisely, about the random variables that represent the votes.
why the relation between Independence and Competence can be confusing

is that the very same factors that undermine Independence can also lower
Competence. Estlund for example challenges the Competence Assumption in
these terms:
[I]f you were to ask, ‘How could a person be dumber than a coin flip?’ the answer
would be ‘easily’. People have more or less systematic views about many issues. If
their system is bad, so to speak, then they could easily be wrong all the time. If, for
example, people in some time and place were systematically racist or sexist or
both, it would not be surprising if their political decisions were worse than the
performance of a coin flip would be on political matters involving race or sex.
Who knows what other important biases or errors people might have in their
systematic thinking on issues?61
Estlund here talks about a single voter being influenced by some factor that
systematically leads that voter astray across several distinct choices. However,
merely a single voter among a great many voters getting things wrong is no
big deal—as long as the average competence across the electorate as a whole
remains greater than 0.5. What matters for the purposes of the CJT is whether
the same systematic bias—racism of sexism or whatever—afflicts many voters’
judgements at the same time and in the same way. If so, then our faith in the
classic CJT result would indeed be undermined.
But what undermines them is not merely that individuals subject to those
biases are individually less competent in consequence. The other important
way in which systematic biases undermine the CJT is by being shared
among many people at the same time, thus compromising the Independence
Assumption.
We can distinguish between four different settings here, as shown in Table 4.1.
Table 4.1 Interaction of the Competence and Independence Assumptions.

Competence Assumption
satisfied violated
Independence satisfied I II
Assumption violated IV III
61 Estlund 2008, p. 16, reiterating a point he made earlier (Estlund 1993, p. 99, n. 44; 2005,
p. 610). Here is how Condorcet (1787/1994, Letter 1, p. 293) himself describes failures of the
Competence Assumption:
False decisions are caused by interest, corruption, passions and error. Interest can be either
personal or professional, or related to a public post or function, or to the legislative body
itself. Passions can be either personal or public. Error can be the result of ignorance,
prejudice or the difficulty of making decisions. Most of these causes can act in two ways:
either directly on each individual, or initially on a few leaders of opinion who, for what-
ever reason, have acquired some influence and command the votes of a certain number
of members.
Limitations 57
In our initial discussion in this section, we will focus on the limiting case in
which the ‘violation’ of the Independence Assumption is complete, that is to
say, the voter in question follows without fail the racist bias, opinion leader, or
other common cause. Then, at the end of this section, we will comment on what
happens in situations short of that limiting case.
Throughout this discussion (and later in the book) it will be useful to distinguish
sharply between two different notions of individual competence. Competence
as the probability of voting for the correct alternative in the given decision envir-
onment is the standard notion of competence appealed to in the Competence
Assumption. As introduced above, we call that probability pc . Another notion
of competence we sometimes appeal to is the probability of a voter voting for
the correct alternative if he or she were to decide without influence of the specific
common causal factor under discussion. We denominate this partially idealized
notion of individual competence as pc*.62 For instance, in the following discus-
sion we distinguish between the probability of voting for the correct alternative
after having been influenced by an opinion leader as opposed to the probability
of voting correctly without being influenced by the opinion leader.
Case I in Table 4.1 is the felicitous case in which the classic CJT’s Competence
and Independence Assumptions are both satisfied. In that case (assuming all
other conditions are satisfied) the classic CJT result will follow.
Case II is a scenario in which the individuals are, on average, worse than
random in their judgements—but their votes are independent of one another,
in the relevant respect. For example, if I am delusional and you are delusional,
but we are each delusional in different ways, then our respective delusions
might lead each of us to vote wrongly often, although not necessarily in the
same way; our votes remain independent of one another.
Should that happen, it would be terrible news from an epistemic point of
view. If everyone is delusional each in his own idiosyncratic way, and each vote
is therefore independent of every other but also worse than random, then the
probability that a majority vote will be correct would converge to zero as the
number of voters increases, just as shown in Figure 4.1 above.
But just how credible is this scenario? It requires that voters all be systematically
incompetent, but at the same time fully independent. That would require voters
all to be confused each in his or her own utterly idiosyncratic way.
Much more likely are cases where the confusion stems from some common
source. Suppose one and the same common factor (or factors) lowers the com-
petence of every individual at one and the same time. Suppose, for example,
that all voters were influenced by the same racist bias (to vary Estlund’s example
above). The racist bias that they all share in common would then make them
individually incompetent—it would make each of them more likely to be wrong
62 We speak of it as a ‘partially’ idealized notion because sometimes voters are influenced by
other common causes that remain in place.
than right, when casting a vote on issues that they see through their racist
frame. But that same racist bias that they all share in common would also tend
to make all of them vote in the same way—thus undermining the Independence
as well as the Competence Assumptions of the classic CJT. That would be an
instance of case III in Table 4.1.
In case III sorts of situations, the probability of the majority being correct
does not normally converge to zero, as it would in case II where individually
incompetent voters act independently of one another. Suppose that all voters
follow without fail one and the same racist opinion leader, whose own probability
of being correct is only 0.40. Then the probability of each individual voter
voting correctly would be 0.40. By assumption, their individual competence
pc is therefore also worse than random, but since everyone follows the same
opinion leader, that does not do any additional harm. Contrast a case II sort of
scenario: with any large number of voters who are that incompetent, the prob-
ability that a majority of them would be correct would approach zero. But that
is true only if their votes are independent of one another—which in case III
they are not. Instead, they are all following (following without fail) one and the
same racist opinion leader. Then the probability that the group is wrong is simply
the probability of the racist opinion leader that everyone follows is right. So
if the racist opinion leader gives a true signal 40 per cent of the time, a majority
of the people following him slavishly will be right 40 per cent of the time—
and that figure is impervious to how many people there are who follow that
opinion leader.63
Finally, in case IV the individuals are competent but their votes are not
independent in the relevant respect. Such settings are likely to arise, for example,
if individuals follow a competent opinion leader (or some other causal factor
or factors). In case IV, neither of the classic CJT results necessarily obtain. The
probability that the majority will be correct does not necessarily converge to 1
as the number of voters grows very large: instead, if each voter follows the same
opinion leader with absolute certainty, the probability that the majority of even
a very large number of voters is correct is once again simply equal to the probability
that the opinion leader is correct. Furthermore, the majority is not necessarily
more likely to be correct than each voter would be acting independently (since
voters might have been individually more likely to be right, had they acted on
their own, than the opinion leader is likely to be right).
For a graphic representation of these four cases, consider Figure 4.2. In the
scenario there depicted, there are 200 voters each of whom has the same indi-
vidual competence level when not being influenced by an opinion leader pc*
(with various such levels being depicted on the x-axis). The probability that the
majority of such individuals will be correct, for each of the scenarios represented,
is given on the y-axis.
63 That is just a special case of the ‘Best Responder Corollary’ set out in Section 5.3.
Limitations 59
If voters vote independently of one another, the probability that the majority
vote of the group will be correct is represented by the S-curve in Figure 4.2,
with the top-right part showing what happens when each voter is of better-
than-random competence (case I) and the bottom-left part showing what
happens when each voter is of worse-than-random competence (case II). Case I
is of course just the classic CJT, and case II is the inverse of that already presented
in Figure 4.1.
For the other two cases, suppose that all those 200 voters follow without fail
one and the same opinion leader. That is to say, they are certain to vote the same
way as the opinion leader (so the individual voter’s own individual competence
level, as represented on the x-axis, ceases to be relevant). The higher line shows
the effect of all voters being guided by a relatively competent opinion leader
(case IV), the lower line the effect of a less competent one (case III). Figure 4.2
shows what being herded by the opinion leader brings about in terms of epistemic
costs or benefits when compared with the S-curve of independent voting: For
incompetent voters it is almost always better to follow the opinion leader, for
competent voters it is almost always worse.
Of course, the results described above for cases III and IV arise with necessity
only if the Independence Assumption is violated completely—that is to say, only
if each voter follows the opinion leader (or other common cause) ‘slavishly’ or
‘with absolute certainty’. If the violation of the Independence Assumption is
1.0
Case I
0.8
competent OL (Case IV)

0.6
0.4
incompetent OL (Case III)
0.2
Case II
0.0
0.2 0.3 0.4 0.5 0.6 0.7 0.8
Individual competence without OL pc*
Figure 4.2 Probability that the majority of 200 voters is correct, depending on whether
voters vote independently (cases I and II) or are certain to follow an opinion leader
(cases III and IV) who is 0.40 likely to be correct or who is 0.60 likely to be correct, for
electorates of varying levels of homogeneous individual competence.
less complete than that—if voters follow the opinion leader some but not all of
the time—then something like the classic CJT results might nonetheless emerge.
In a scenario where voters are partially independent of any common causes and
therefore of one another, and where each voter is incompetent when she does
vote on the basis of her own judgement (the left-hand side of the figure), the prob-
ability that the majority will be correct might still converge to zero—just a little
more slowly than in the classic CJT setting. In a scenario where voters are partially
independent of one another and where each voter is of better-than-random
individual competence (the right-hand side of the figure), the probability that
the majority will be correct might still converge to one—just a little more slowly
than in the classic CJT setting. The crucial cutting point for just how much of a
violation of the Independence Assumption is consistent with those outcomes is
given in the next section.
4.5.3 Living with Dependence
One good reason not to worry overly much about failures of independence is
that the most extreme cases of dependence might just not be that common.
Condorcet contemplated the problem of opinion leaders, imagining the case
of an elector who is so ‘totally devoted to his faction’ that ‘no form of election
can prevent him from following the master he has chosen, whether by corrup-
tion or enthusiasm’. But as Condorcet goes on to add, ‘such slavish followers are
rarely found’.64
Such slavish following of the opinions of others may be more of a problem in
our day than Condorcet’s, with the consolidation of media ownership and the
rise of dominant cable news channels. But it is an open—and empirical—question
just how widespread it is across the electorate. It is all well and good for philo-
sophers to point to the possibility that the problem is sufficiently widespread to
vitiate the CJT results. But logical possibility is one thing; empirical probability
is another. That is something that cannot be established purely from the armchair,
although our discussion of the victories of Trump and Brexit in Chapter 21 give
us pause on this score. It happens. Just how frequently is the question.
People who follow an opinion leader or succumb to some other common
causal factor rarely do so slavishly, 100 per cent of the time. More typically, they
do so (as Condorcet says65), in a more-or-less, upon-some-occasions fashion.
And if people’s probability of following some common causal factor is less than
64 Condorcet 1789/1994, p. 178.

65 Thinking in terms of multi-option rather than dichotomous choices, Condorcet (1789/1994,
p. 178) goes on, from the passage quoted for n. 64, to say: ‘Other electors will only be influenced
into ranking certain candidates, or candidates of a certain party, among the first in their lists. This
kind of intrigue would scarcely have a dangerous effect on the resulting will of the electors’ (i.e. the
electoral outcome).
Limitations 61
100 per cent, then—depending on just how likely they are to follow that common
causal factor, compared to how likely they are to vote for the correct alternative
when not doing so—the conclusions of the CJT may still obtain. So the real
question is just how much influence by one opinion leader can be tolerated
without breaking the Asymptotic Result of the CJT.
Call the probability of each individual’s following one and the same opinion
leader π and call the individual’s probability of voting for the correct alternative
(if they do not follow the opinion leader) pc* as before. Assume for convenience
that all voters have the same π and the same pc* and this is true for both states,
in the spirit of the classic CJT framework. Then it can be proven66 that group
competence still converges to 1 as the number of voters approaches infinity, just
so long as
π < ( pc * − ½ ) / pc *.
That is not hideously demanding.67 If, for example, pc * = 0.55 then the CJT’s
Asymptotic Result would still obtain for any π up to 9 per cent; and if pc* = 0.6
it would still obtain for any π up to 16 per cent. Of course, if the voters hang on
the lips of the opinion leader and follow her most of the time ( π ≥ 0.5 ) then no
value of pc* can compensate for that; and the probability that the majority
will be correct would then converge to the probability that their opinion leader
(or ideology or heuristic or whatever they are so faithfully following) will be
correct. And even if π is only half that large, the values of pc* required to com-
pensate for that would be unrealistically high. But for relatively modest values
of π, it is altogether reasonable to expect that the value of pc* might well be
sufficiently high for the CJT’s Asymptotic Result still to hold true.
A second reason for not worrying too much about failures of independence
is that not all patterns of dependence are equally bad. As we have seen, one strong
opinion leader moving all voters in the same direction is disastrous. However,
things would look far less bleak if for example there were several opinion leaders
pulling in different directions: equal and opposite influences might cancel one
another out. We say more about both of these points, providing numerical
examples, in Section 11.2.
In a pluralist society, failures of Independence are less likely to be worrying
for this sort of reason. It is all right that some people’s votes are subject to com-
mon causes, as long as different people are subject to different common causes
and enough people remain independent in their judgements. That is one way
in which social diversity contributes to the epistemic quality of collective
66 Boland 1989, pp. 185–6. Spiekermann and Goodin 2012, p. 568, with the proof given in
Appendix 2 of the supplementary online material.
67 Anyway, it is not for even moderately high levels of pc*. If pc were very low—pc* = 0.501 for
example—then π as low as two-tenths of a percent would suffice to foil the CJT.
decisions,68 and it shows how freedom of speech and opinion can contribute to
collective correctness.69
The next chapter will explore much more fully, in more theoretical fashion,
various other ways in which the Independence Assumption might be violated
and yet a version of the CJT results can still be obtained.
4.6 LIMITATIONS ARISING FROM THE CHOICE

SITUATION
Finally, there is a suite of limitations on voters’ epistemic performance arising

from the nature of the choice situation itself. Those have three principal sources.
4.6.1 The Choice Situation Is Systematically Misleading
Voters might systematically err if the choice situation is itself systematically

misleading. That occurs when everyone’s votes are all subject to some com-
mon causes that systematically lead everyone simultaneously to vote in the
wrong direction.
Such failures of Independence have already been discussed briefly in
Section 4.5.2 and will be discussed much more fully in Chapter 5. For now, let
us confine ourselves to one telling example. Suppose the evidence upon which
everyone must base his or her choice is itself systematically misleading. Suppose
it is biased or partial or just downright wrong.
If the evidence is systematically misleading, then voters will be systematically
misled. If even the best responder to that evidence base would form an incorrect
view on the basis of it, so too inevitably will voters, even if they are themselves
highly adept at tracking what the best responder to the evidence would decide.
That is the best that they can do.
But that is also the worst that a large group of faithful followers of the best
responder to the evidence are likely to do. The likelihood that a majority of
them is correct will be no better—but no worse—than the probability that the
evidence is truth conducive.
68 Ladha (1992) shows the epistemic advantages of including voters even if they are more likely
to be wrong than right, so long as that reduces the correlation in votes across the electorate. See
further Chapter 7 on ‘diversity’.
69 This is the classic argument for freedom of speech, within the CJT tradition. Ladha (1992)
provides a formal demonstration, but the basic point is made by Condorcet (1787/1994, Letter 2, p. 319)
himself: when discussing ‘common prejudices which exist in a country’ that might lead to systematic
failure to track the truth, he remarks, ‘prejudices of this kind are quite easily dispelled in countries
where a free press is in frequent use’.
Limitations 63
4.6.2 The Options Are Ill Formulated
The CJT presupposes that all the voters are voting on the same proposition.
Then, and only then, does it make any sense to add their votes together.70
‘[W]here individuals are not addressing the same issue, Condorcetian consid-
erations do not apply.’71 Condorcet himself expressed a concern that voters
have not ‘all addressed the same question’.72
Beyond that, voters are further limited by the way the question on which
they are to vote is cast. Obviously, voters can only vote for the correct option
if that is one of the options on the ballot that is presented to them (as we have
said in Section 4.2.2). Enriching the agenda for decision is one of the particular
contributions that might be made by pre-vote discussion and deliberation, as
will be discussed in Section 9.2.5.
It is epistemically important not only that the agenda be ‘complete’ (par-
ticularly, that it contains what is truly the best option). It is also epistemically
important that the agenda itself be ‘well formulated’. The epistemic success of
decision-makers depends, in particular, upon each distinct option being pre-
sented separately, in its own right, on the agenda for choice. If multiple distinct
options are bundled together and presented as a single option disjunctive in
form, then that may lead to incorrect votes even by voters who would have
voted correctly had each of the distinct options appeared separately, in its own
right, on the ballot.
Here is an example of that.73 Suppose there are three distinct options. Let the
correct option be xc and the two incorrect options be xy and xz. Suppose that
each voter has probability 0.40 of voting for xc and probability of 0.30 of voting
for each of the other options, and this is true for all possible states. Each voter
is thus competent, in the sense required for the plurality rule results discussed
in Section 3.2—that is, each voter is more likely to vote for the correct option
than she is to vote for any of the incorrect ones. But only if she is allowed to—
which here is to say, only if the three options appear, and appear as distinct
options, on the ballot paper.
70 Goodin and Saward 2005.

71 Estlund 1989, p. 1319. See similarly: Estlund 1993, pp. 403, 411–16; Christiano 1995, p. 404;
Cohen 1986; Grofman and Feld 1988; Waldron 1989; Estlund 1997; Gaus 1997; Swift 2014, p. 223;
cf. Vermeule 2009b, pp. 8–9. This is most commonly expressed as a Rousseau (1762/1997, bk. 4, ch. 1)
style worry that people might be voting on the basis of particular interests rather than any com-
mon good that they all share (Condorcet 1785/1976, p. 61; 1789/1994, p 170; 1793/1994, pp. 192–4).
As we show in Chapter 13, a weaker jury theorem obtains even in those circumstances, however.
72 Condorcet 1789/1994, p. 172.
73 That will not always happen, of course. Here is an example where the phenomenon we
describe in what follows could not occur: there are three options; each voter thinks the probability
of the correct option being correct is 0.52 and the probability of each of the incorrect options
being correct is 0.24. Then the correct option would be expected (with increasing confidence the
larger the number of voters) to beat a disjunct of either of the two wrong ones by 52 to 48 votes out
of a hundred.
If all three options appeared separately on the ballot, and if the decision were
made by plurality rule (so whichever option gets the most votes wins), then xc
would be likely to win, and increasingly certainly so the more voters there are.
That is the Plurality Vote Asymptotic Result presented in Section 3.2.
But suppose instead some of the options were presented disjunctively. Suppose
that voters could vote only in two ways:74 one would be for xc; the other would
be for the disjunction ‘xy or xz’. Here is one seemingly natural way to extend the
voting probabilities described above to that new choice situation: let each
voter have a probability of 0.40 of voting for xc just as before; but given that
each voter had a probability of 0.30 of voting for each of xy and xz taken separ-
ately, let each have a probability of 0.60 of voting for the disjunct of ‘xy or xz’. If
so, each voter would thus be inclined to vote for the wrong option ‘xy or xz’
instead of voting for the correct option xc.75
We hasten to add that there is no necessity whatsoever in that ‘seemingly
natural’ way of extending those voting probabilities from the case where each
distinct option is considered separately to the case where some of the distinct
options are combined in a disjunct.76 When the choice is recast in that new way,
it is a different choice situation, and it would be perfectly reasonable for a per-
son to respond to it differently. (Indeed, we presume that voters do precisely
that in our analysis of how a voter may be less likely to choose the correct
option from a larger compared to a smaller set of options.77) There is thus no
reason that a voter necessarily will, or rationally should, be 0.60 likely to vote
for the disjunct ‘xy or xz’, simply because she was 0.30 likely to vote for each
element of that disjunct when presented separately. Still, the voter may become
more likely to vote for the disjunct, and if she does then bunching those two incor-
rect options would lead such a voter to vote for the disjunct of the two incorrect
options rather than for the one correct option.78
74 Nothing turns on the fact that a multi-option choice situation has therefore been recast as a
binary choice. The same problem described below would occur if there were five options, a correct
one which each voter judges to be 0.24 likely to be correct and four incorrect ones which each
voter judges to be 0.19 incorrect. Suppose the voters were then asked to make a four-way choice
between one disjunct combining two of the incorrect options and each of the remaining options
listed separately. The disjunct of the two incorrect options would win by plurality rule, garnering
roughly 38 votes out of a hundred to the correct option’s 24 votes.
75 In this example as constructed, voters presented with a series of choices involving different
disjuncts would always favour the disjunct over any single, determinate option. The fact that no
unique option would ever emerge as the winner counts as a further argument against allowing
choices over disjuncts—but it is the damage that such choices can do to voters’ epistemic perform-
ance that is the focus of our present discussion.
76 The disjunction problem invites people to confuse the probability of voting for an alternative
with the subjective beliefs that individuals may hold over different alternatives. If these probabil-
ities were subjective beliefs, then of course individual voters would be constrained by principles
of rationality regulating these beliefs. But in the CJT framework votes are instead determined by
a private signal (one for each voter) about what is the correct alternative. The relevant probabilities
are probabilities with regard to those signals only.
77 See Sections 3.2 and 8.2.1.
78 Note that this example also relies on the unrealistic assumption that the agenda setter knows
in advance which of the three options are the incorrect ones.
Limitations 65
Some commentators seem to suppose that this ‘disjunction problem’ poses a

major challenge to the CJT framework as a whole.79 One answer to the chal-
lenge is to say that competence is agenda-relative. If we take competence as
agenda-relative, then the disjunction challenge merely draws attention to
another limitation of the CJT: it will work only if the options are not miscast.
But that is an argument for not miscasting the options that we present to the
electorate, and nothing more. And we should want to avoid miscasting the
options that we present to the electorate, not merely to ‘save the CJT’, but
instead to enable democratic majorities to make sensible decisions. Putting
each distinct option on the ballot separately, in its own right, rather than dis-
junctively ‘bunched’ with some other, is the right rule for doing that. In any
case, the fact that rearranging the agenda can turn a competent voter into an
incompetent one should not be too surprising. After all, it is well known that
agenda manipulation can be highly effective—it would be surprising if that was
not the case in the epistemic setting.
Avoiding a miscasting of alternatives might be easier said than done, of
course. Certainly there is no ready formula—like a rule that says, ‘no option
on the ballot may contain the word “or” ’—that we can use to be sure of achieving
that goal. Suppose in the three-option case described above that the question that
appeared on the ballot read, ‘Should x1 be enacted?’80 There is no telltale ‘or’
appearing on the face of that question. But implicitly, in the decision situation
in view—where there are actually two distinct options x2 and x3 implicit in the
‘no’ choice fork—the structure of the choice being presented to voters is none-
theless disjunctive. But to repeat our point from above—it is by no means clear
that putting alternatives into agendas of different disjunctive forms leads to a
violation of the Competence Assumption. It might or might not do.
Another more forceful response to the ‘disjunction’ challenge would be to
reconceptualize a person’s competence as a matter of how much better than
random that person is at making a decision. A person with no competence will
be no better than random, over a long run of choices. The probability that she
will choose correctly among k options is just 1/k. But a person with competence
of ε above random will be ε +1/ k likely to choose the correct option.
Estlund writes:
Consider a choice among three alternatives: A, B and C. If we suppose, a priori,
that voters are a little better than random, we might let them have, say, a .34 chance
of getting the right answer and a .33 chance of each of the wrong answers. But
suppose we presented the choice differently: alternative A versus the disjunction
of B or C. By leaving the choice between B and C for later, the choice is now binary.
Since the choice is now a binary one, are we suddenly entitled to suppose voters must
79 Sunstein 2006c, p. 37. Estlund 2008, pp. 228–30, 232–6. Vermeule 2009b, p. 7.

80 Estlund 1997, p. 189. Notice that that is the form of the question put to people in many states
of the US, where judges are initially appointed for a certain number of years by the governor, and
then the question is put to the voters whether that judge should be confirmed to continue in that
role for life or not.
be at least a little better than .5? Is it a minimal, modest assumption that they are
more likely than .5 to choose A, which is the right answer? Quite a promotion.81
But conceiving competence in the more disaggregated way we have suggested,
the ‘promotion’ is not for anything peculiar to the voter himself but rather merely
for the ‘random’ element in the pc = ε + 1/ k equation. It is the 1/k component
of pc , not the ε component, that has risen so dramatically when shrinking the
number of options.
81 Estlund 2008, p. 228.

Independence Revisited
The Independence Assumption is the most contested and most misunderstood

aspect of Condorcet’s jury theorem (as well as many other more recent
jury theorems). Here we hope to dispel some of the misunderstandings
surrounding it.
In this chapter we will explain ways to change the Independence Assumption
and make it more realistic. Even though there are good theoretical reasons for
using a more sophisticated notion of independence, we will, however, for the
rest of the book, often revert to the simpler assumption on which Condorcet’s
original jury theorem was based. Working with this simpler jury theorem will
still yield many interesting insights without the additional technical baggage
of the more refined theorems. Nonetheless, readers with a more technical
inclination could replace the simple models with the more sophisticated assump-
tions we discuss over the course of this chapter.
The central argument of this chapter (especially Sections 5.2.1, 5.3, and 5.4)
builds on technical results owing to Christian List and especially Franz Dietrich,
who with Kai Spiekermann did further work elaborating and extending those
results. The underlying papers, which are cited at the appropriate points in what
follows, remain authoritative where the current gloss deviates from them.
5.1 UNDERSTANDING THE INDEPENDENCE

ASSUMPTION
Having explored the issues surrounding the CJT’s Independence Assumption

discursively at the end of Chapter 4, let us now approach them more formally.1
Let us begin by setting out exactly what is involved in the Independence
Assumption as it figures in the classic version of the CJT.
1 The discussion here and in what follows draws on earlier and more technical accounts in
Dietrich and Spiekermann (2013a, b), which builds in turn on the earlier work of Dietrich and List
(2004) and Dietrich (2008).
5.1.1 Independence Is Not Absence of Interaction
In many naive accounts of independence, the assumption is described in terms

of the absence of any interaction between the voters. Independence is supposed
to be met if the voters do not influence each other. Influence could range from
seemingly benign (voters talking to each other2) to severe (one voter forcing
the other to vote in a specific way).
Interpreting independence as literally the absence of voter interaction is
mistaken in two ways:
1. It is too permissive, because it does not rule out dependence between
voters caused by factors beyond direct interaction.
2. It is too strict, because it rules out interaction even if the interaction does
not cause increased dependence between the votes.
The first mistake occurs because a major source of dependence is overlooked.
Voters can be influenced by the same common causes even if the voters have no
direct contact with each other whatsoever. Opinion leaders, shared evidence,
or similar theories for interpreting the world can cause people’s votes to be
positively correlated and make voters vote the same way, without any direct
communication between them.
The second mistake occurs because of a conflation between the interaction of
voters and statistical independence of the votes. Suppose, for example, voter 1 says
to voter 2, ‘Don’t just vote the way I do, make up your own mind’; and suppose that
voter 2 follows this advice. Then this is undoubtedly a direct interaction between
voters. But it is one that does not increase dependence between the votes.
What we should ultimately be concerned about is statistical independence of
the votes. Statistical independence of the variables to be aggregated is the cru-
cial premise for the law of large numbers that underlies the Condorcet Jury
Theorem. Variables are independent when learning the values of some variables
does not teach us anything about the values of others. For example, the results
of tossing two unbiased coins are independent: learning the result of one does
not make our guesses about the outcome of the other one any better.
By contrast, the coin tosses of a skilled con artist may not be independent.
Suppose the con artist can probabilistically influence the results in one unknown
but fixed direction. Then in a matching game, if the first coin comes up heads it
is a good guess that the second one will come up heads (because, in a matching
game, the first result is evidence that the con artist is interested in bringing
about matching heads). The coin tosses of the con artist are not independent.
They are, however, independent conditional on the influence of the con artist.
2 Echoing Rawls (1971, p. 358), Cohen (2010, p. 79) for example writes, ‘the Condorcetian
argument is very limited, in several ways. First, the theorem requires that individual judgments be
independent. But if people are talking to each other the judgments do not meet that condition.’
Independence Revisited 69
Once we know that a con artist is aiming at producing heads, we do not learn
anything additional about the second result when we observe the first result.
5.1.2 Independence Is Not Unconditional
To avoid the mistakes noted above, we need an account to show how the causal
interaction between voters and their environment leads to (or undermines)
independence of the votes—the statistical property we are ultimately interested in.
This can be accomplished with the aid of diagrams depicting causal networks.
A causal network shows the causal relation between different phenomena.
Nodes in the network represent the phenomena, arrows the causal effect of one
phenomenon on another.3
Figure 5.1 shows a very simple network with the votes of three voters (it is
easy to extend this and the following diagrams to include more voters). For
the voters to be competent, their Votes need to be causally influenced by the
‘State’ of the world. The Votes need to ‘track’ the State of the world, otherwise
the voters cannot have a better-than-random probability of voting for the cor-
rect alternative.
This, in turn, means that the Votes must be positively correlated and therefore
dependent. If Vote 1 is for alternative 1, it is more likely that Votes 2 and 3 are also
for alternative 1, provided that all voters are competent. In other words: learning
that Vote 1 has been cast for x tells us something about how Vote 2 will be cast,
namely that it is more likely than not that Vote 2 will be cast for x, too.
There is an important lesson in this. If the voters are competent, their votes
cannot be unconditionally independent. That is simply an incoherent way of
specifying independence, in the context of the CJT.
5.1.3 Independence Conditional on the State of the World
Conditional on the true State, however, the Votes in Figure 5.1 are independent.
Once we know the State, we do not learn anything additional by observing
State
Vote 1 Vote 2 Vote 3
Figure 5.1 Independence conditional on the State of the world.
3 More technical details are given in Dietrich and Spiekermann 2013a, b.

State
Figure 5.2 Direct violation of independence conditional on the State of the world.
some Votes about any other unobserved Votes, in the same way as in the
previous example we do not learn anything about the con artist’s next coin
toss by observing his previous toss once we know the con artist’s intentions.
This, in fact, is the sort of independence presupposed by the classical form
of the Condorcet Jury Theorem: independence conditional on the state of the
world.4
However, independence conditional on the state of the world rules out direct
causal links between the votes that would arise if voters simply copied each
other’s votes. Consider as an example of that Figure 5.2. There, Vote 1 influences
Vote 2 and Vote 2 influences Vote 3. Hence, the Votes are no longer independ-
ent conditional on the State of the world. This is easy to grasp intuitively: if
Vote 1 influences Vote 2, and we hear the outcome of Vote 1, we learn something
about Vote 2, and more than we already knew given the State of the world.
5.2 DEALING WITH DEPENDENCE BY

FURTHER CONDITIONALIZING
A direct causal influence between votes is the most blatant violation of the
Independence Assumption. In the extreme case, if the voters blindly follow each
other in a sequential vote, the dependence of votes is severe and its source easy
to observe. However, dependence due to common causes, while more difficult to
spot, is at least as important.
Suppose that the voters are not only influenced by the state of the world but
also by one opinion leader. To make matters worse, in this example the opinion
leader (‘OL’ in the figures) does not have any evidence himself—there is no
causal link between the State and the opinion leader (OL) in Figure 5.3. There
too independence conditional on the state of the world is violated, in Figure 5.3
indirectly (via the shared influence of the opinion leader) rather than directly
as in Figure 5.2.
4 More formally put, the classic CJT assumes that ‘the probability of each voter’s vote, conditional
on the state of the world and conditional on any given other voter’s vote’ is simply equal to ‘the
probability of each voter’s vote, conditional on the state of the world’.
State OL
Figure 5.3 Indirect violation of independence conditional on the State of the world.
When common causes other than the state of the world influence the votes,
the votes are not independent conditional on the state of the world because
they are all correlated with the common cause and thus correlated with each
other. Opinion leaders are not the only possible common cause that can induce
dependence. The same problem would arise if we replaced ‘OL’ with ‘ideology’
or with ‘Fox News’,5 or with ‘heuristics’ or with ‘cues’—or perhaps even with
freak factors such as ‘the time of day’ (there is some disconcerting evidence that
judges are less likely to grant parole when they are hungry, just before lunch6).
5.2.1 Independence Conditional on the Available Evidence
Epistemically, we want people’s votes to be causally determined by the state of

the world. That is what it is for them to be ‘competent’. And if everyone’s vote
is causally influenced by the state of the world, there will inevitably be some
interdependence among them by virtue of that common cause. Yet that sort of
interdependence is epistemically desirable.
Suppose now, however, that voters do not have direct access to the state of
the world. Suppose that instead they only have access to ‘Evidence’ about the
state of the world.7 This case is as depicted in Figure 5.4.8 Here the State of the
world causes some ‘Evidence’ to take on certain values. This Evidence, in turn,
causally influences the Votes.
To see how plausible this setup is, consider jurors in a criminal trial. The state
of interest is whether the defendant is innocent or guilty. However, this state of
the world does not directly cause the votes to be in favour of conviction or
acquittal. Rather, the jurors will try to infer the state of the world (guilt or inno-
cence) by observing the evidence available. Does the defendant have a motive?
Was she in possession of the murder weapon? Does she have an alibi? Is there
forensic evidence linking the defendant to the scene of the crime? Are there
witnesses reporting sightings of the defendant? And so on. The important upshot
of the example is that we hardly ever observe the state of the world directly.
5 Beatty (2007) analogizes Fox News to the Soviet propaganda organ, Pravda (which translates,
of course, as ‘truth’).
6 Danziger, Levav, and Avnaim-Pesso 2011. 7 Ladha 1992; 1993; 1995.
8 Adapted from Figure 3 in Dietrich and List (2004).
State
Evidence
Figure 5.4 Independence conditional on the available Evidence about the State of
the world.
Instead we try to get at the state of the world by more indirect means, especially
evidence and testimony.
This has significant implications for how we should conceptualize inde-
pendence. Note first that in Figure 5.4 the Votes are no longer independent
conditional on the State of the world. To see this, suppose that the defendant
is innocent. If we now observe many votes for convictions we can conclude
two things: first, the evidence the jurors face must be of a misleading nature
(perhaps the defendant’s twin brother left DNA traces on the scene9); second,
the presence of misleading evidence makes it more likely that other votes are
also incorrect, a violation of the intuitive independence check. Thus, condi-
tionalizing on the state of the world alone does not ensure independence.
Furthermore, the dependence problem in Figure 5.4, unlike that in Figure 5.3,
cannot be solved by simply removing the common cause—the evidence is
essential! Observing the evidence is, ex hypothesi, the only way to make infer-
ences about the state of the world because the effect of the state of the world on
voters is completely mediated through the evidence.
Supposing that votes are entirely relying on the available evidence, rather
than on the state of the world itself, leads to an importantly different inter-
pretation to the CJT. On that understanding, the CJT is not about tracking the
truth (the state of the world) directly. Instead, it concerns the ability of individ-
uals and groups to reach the same judgement as the best responder would on the
basis of the shared evidence that is available to them.10 The CJT should then be
interpreted as showing that the majority of a large group of competent and
independent voters (now independent conditional on the evidence) converges
toward the judgement of the best responder, who processes that evidence in
9 As was the case of Elvin Gomis, who was also held in preventive custody for ten months after
DNA evidence found at the scene of multiple rapes in Marseille could not distinguish him from
his twin brother Yoan, who eventually confessed to the rapes (Lichfield 2015).
10 Dietrich and List (2004). We call their ‘ideal responder’ the ‘best responder’ because we
generalize that notion of best response in what follows.
the best possible way and who chooses the option most likely to be true given
that evidence.11
The CJT modified in this way changes the Independence Assumption. Votes
no longer need to be independent conditional on the state of the world, but
instead independent conditional on the available evidence. Correlation induced
by the fact that voters all share the same pool of evidence thus ceases to be a
problem for the CJT modified in that way. This is a special case of the ‘Best
Responder Corollary’, to be discussed in Section 5.3.
This conditionalize-on-the-evidence extension of the CJT has another
desirable feature. It explains why even large groups can be wrong, despite the
fact that the classic CJT’s Asymptotic Result suggests that they should be nearly
infallible. In the CJT modified to conditionalize on the evidence, the probabil-
ity of the majority’s being correct is upper-bounded by the probability that the
evidence available to it, interpreted in the best possible way, is actually pointing
to the correct result. Insofar as the available evidence is less than perfect in that
respect, Pn will to that extent asymptotically approach some value well short of 1,
even as n approaches infinity.12
Conditionalizing on the evidence also helps to explain away another puzzle
arising with the classic version of the CJT, which is why it can ever be epistemically
advantageous to democratically revisit issues once they have been initially resolved
and to take a second vote on the same issue. ‘Democratic decision-making
needs to recognize its own fallibility, and hence needs to institute feedback
mechanisms by which it can learn how to devise better solutions and correct
its course in light of new information about the consequences of policies’, as
Elizabeth Anderson rightly remarks.13 And while the classic CJT might pro-
vide no explanation as to why that should be necessary, a revised version of the
CJT in which the probability that the majority is correct is upper-bounded by
the probability that the available evidence is truth-conducive can provide
a ready explanation—namely, it can be epistemically advantageous to take a
second vote on the matter when we get new evidence.
5.2.2 Independence Conditional on an Opinion

Leader’s Interpretation of the Evidence
Often voters’ access to the true state of the world is mediated not only by the
evidence but also by opinion leaders who report on the evidence. In Figure 5.5
we depict the case of a single opinion leader (OL) being the only way the voters can
learn about the Evidence and, through it, the State of the world. The voters are
11 Competence must now be understood as the ability to track the judgement of the best
responder. We make this more precise below.
12 Dietrich and Spiekermann 2013a. 13 Anderson 2006, p. 12.
State
Evidence
OL
Figure 5.5 Independence conditional on the Opinion Leader.
now entirely reliant not only on the quality of the Evidence but also on the
ability of the opinion leader to interpret the Evidence and relay the relevant
information to the voters. The Votes themselves are, however, independent of
one another, conditional on ‘OL’.
Assume now that we increase the number of voters in Figure 5.5, with all Votes
being influenced by OL, while keeping everything else equal. In the best case,
the collective competence of the voters converges on what the best responder
could make out of the information provided by the opinion leader (OL). The
success of the best responder in correctly ascertaining the true State of the
world is determined not only by the quality of the Evidence but also by the
epistemic quality of the opinion leader (OL).
5.2.3 Independence Conditional on All Common Causes
Let us now extend that same logic to the case in which several common causes
combine. In Figure 5.6, for example, we have evidence that informs an opinion
leader (OL). It also influences a joint ‘Heuristic’ that the voters use and a shared
‘Cue’ they rely on. To obtain independence of Votes, we need to conditionalize
on all common causes (OL, Heuristic, Cue, Evidence, State) together.14 Once
we conditionalize on the common causes, however, the Votes are once again
independent.
Just as in the case of a single direct common cause (the Opinion Leader OL
in Figure 5.5), so too in the case of multiple common causes: if we increase the
14 This is the central insight behind the jury theorem in Dietrich and Spiekermann (2013a).
State
Evidence
OL Heuristic Cue
Figure 5.6 Violation of independence due to multiple common causes.
number of voters such that all votes are influenced by ‘OL’, ‘Heuristic’, and ‘Cue’,
the collective competence of voters converges on what the best responder could
make out of the information provided by the set of these common causes. The
success of the best responder in correctly ascertaining the true state of the world
is determined not only by the quality of the evidence but also by the epistemic
quality of the common causes taken as a whole.
Let us now consider the possibility that, while voters are influenced by the
common causes, they also have some independent access to the evidence
themselves. Figure 5.7 shows causal constellations of that kind.
The votes in Figure 5.7(a) and 5.7(b) are independent of one another, condi-
tional on all the common causes (Evidence, OL, and State). The interesting
question is how the Evidence and the common cause OL would influence the
upper bound of group competence, if we increased the group size but keep a
similar causal structure as in Figure 5.7 (i.e. all Votes are influenced by the
Evidence, and there is one opinion leader OL who is (5.7(a)) or is not (5.7(b))
influenced by the Evidence, and who influences all Votes).
In one polar case the private access the voters have to the evidence is
dominant—the opinion leader does not play any significant role. In that case,
competent voters, by definition, have a probability of greater than 0.5 to vote
for the judgement the best responder to the evidence would arrive at, as dis-
cussed above. With the number of voters going to infinity the competence of
the group approaches the competence of the best responder to the evidence.
How good a large group of competent voters does depends on how often the
best responder would be misled by the evidence.
In the other polar case, the influence of the opinion leader (OL) is dominant—
the private evidence the voters have does not play any significant role. In that case,
competent voters have a probability of greater than 0.5 to vote for the judgement
of the best responder to the information provided by the opinion leader (OL).
(a) State (b) State
Evidence OL Evidence OL
Vote 1 Vote 2 Vote 3 Vote 1 Vote 2 Vote 3
Figure 5.7 Voters are influenced by both Evidence and a common cause.
Whether that is good or bad, epistemically, depends on two things. First, does
opinion leader OL have access to the Evidence (as in Figure 5.7(a)) or not (as in
Figure 5.7(b))? Second, if the opinion leader has access to evidence, how good is
the opinion leader in responding to the evidence and relaying the results to the
voters? If the influence of the opinion leader on the votes is not based on any
evidence, then he (and voters following him) cannot collectively do better than a
random guess. If, however, the opinion leader is informed by evidence, then it all
hinges on what the opinion leader does with the evidence and how good the
evidence is.
Most of the cases do not lie at either of these two extremes, of course. In
many situations the votes are partly influenced by the opinion leader, partly by
the private evidence. This combination of causes can be epistemically beneficial
or detrimental, depending on how they causally interact with each other.
It bears emphasizing that it is perfectly possible that the common causes as
a whole might point in the correct direction with a very high probability. It
is perfectly possible that opinion leaders, heuristics, cues, and all the other
common causes of people’s decisions might be highly reliable epistemic guides.
(People have often chosen them, insofar as they are matters for choice, for
precisely that reason.) If so, the probability that a majority of voters who follow
such common causes will be right might be high. Still, that probability is upper-
bounded by the probability that the common causes taken together point in the
correct direction.
5.3 THE BEST RESPONDER COROLLARY
Let us sum up the important points so far. Condorcet’s classical Independence

Assumption is typically violated because the votes are influenced by common
causes such as shared evidence, opinion leaders, cues, and many more. We have
suggested, informally, that there is a fix: conditionalize on all common causes.
Revising the Independence Assumption in this more realistic way requires,
however, a different way of thinking about competence. We must now think of
the voters as being competent in the sense that they are better than random in
following what the best responder would be able to establish about the state,
given all the common causes.15
This revision has a further important implication. That is that the wisdom of
the crowds is upper-bounded by the probability that the best responder would
himself reach the correct answer. Take the common cause of ‘the evidence
about the state of the world’, for example. The best that even the largest group of
competent voters can do, epistemically, is only as good as the best responder
could do on the basis of the quality of the evidence that is available. Put that
way, the point is obvious and unsurprising. And revising the CJT in such a way
as to reflect that obvious truth avoids the counter-intuitive implication associated
with the classic CJT that the majority among a large number of voters meeting
its condition is always utterly certain to be correct.
Let us now make this revised version of the CJT—which for ease of back-
reference later in the book we will call the ‘Best Responder Corollary’—a little
more precise.16 As formulated it applies to decisions between two alternatives,
although it could presumably be generalized.17
Let us call the set of all actual common causes influencing the votes the
decision situation (or the situation for short).18 From an epistemic point of view,
the situation can be either truth-conducive (good) or misleading (bad). Truth-
conducive situations are those where the evidence points to the right result.
Opinion leaders might back the correct alternative, for example, or cues that
the voters use point to the right choice. Misleading situations are those where this
is not the case.19 For example, the evidence may be misleading because the core
witness is a convincing fantasist or because the voters follow a cue (e.g. ‘the
candidate is tall and therefore a natural leader’) that misfires.
Let us define the best responder as someone who does the best epistemically
that can be done, given the decision situation.20 If the situation is truth-conducive,
15 We are borrowing and extending an idea first stated by Dietrich and List (2004).
16 We do not make it fully precise as this requires a technical apparatus beyond the scope of this
book. See Dietrich and Spiekermann (2013a, b) for technical background and formal proofs.
17 As in Section 3.2.
18 Formally, the decision situation must contain all common causes and the state of the world.
19 When we say that a decision situation is systematically misleading, that assumes that that
decision situation does not itself contain meta-information that it is systematically misleading.
If it did, it would not really be systematically misleading at all. In that case, the best responder
should simply toss a fair coin, ignoring all else in the decision situation; and the probability that
he would reach a correct decision, doing that, would be 0.5 rather than the smaller probability of
being correct following the other components of the decision situation.
20 That is to say, given the existing structure and instantiation of common causes. It might be
better yet epistemically to alter that structure, and we suggest how discussion might do that in
Chapter 9; but someone doing that would cease being a mere ‘responder’. There are certain common
causes whose causal power is beyond our control. The best responder simply has to take those as
given, and exercise the ‘best’ choice she can within those constraints. Consider for example the
case of the common cause of ‘evidence’: the best responder can choose among alternative inter-
pretations of the evidence; but she has to take the evidence itself as given (at least for the purposes
the best responder’s vote will be correct. But if the situation is misleading,
even the best responder’s vote will be incorrect. (Given the misleading com-
mon causes that constitute the situation, she has no chance.) Let ω be the prob-
ability that the best responder will vote correctly, which is the probability of the
decision situation she is confronting being truth-conducive.
Using the device of the best responder, we can now express the competence
of the voters very intuitively as their ‘probability to track the vote of the best
responder’. This value is analogous to the standard CJT’s competence param-
eter, except that the voter now tracks the vote of the best responder rather than
the state of the world. Let pBR be the probability that voters vote in the same
way as the best responder (for all possible decision situations), and let us call
the voters best-responder trackers if voters are more likely than not to follow the
vote of the best responder, so that pBR > 0.5.21 In other words: while the aver-
age voter is not a best responder, she is more likely than not to follow the best
responder.
This leads us to the following proposition:
The Best Responder Corollary:22 If the votes are independent conditional on
the situation23 and the voters are best-responder trackers, then the probability
of a majority vote being correct converges to ω, the probability that the decision
situation is truth-conducive, as the group size increases.24
Figure 5.8 provides a diagrammatic illustration of the implications of the Best
Responder Corollary. The classic CJT implicitly assumes that the decision situ-
ation is truth conductive, in the sense that the probability that the best responder
will make the correct choice in that situation is ω = 1.0. And in that case, the
probability that the majority of voters who track the best responder will be
correct approaches 1.0 as the number of such voters increases towards infinity.
That is the classic CJT result. But the Best Responder Corollary invites us to
consider cases where the truth situation is something less than perfectly truth-
conducive. If the probability that it is truth-conducive is ω = 0.9 , for example,
that is the best that the best responder can do—and that constitutes the upper
bound of the probability of the majority of any number of voters tracking the
of the decision immediately at hand: in the longer term she can always invest in acquiring more
evidence). Something similar applies to other common causes.
21 And that is true for all possible decision situations. Also, p BR must be well defined and
bounded away from 0.5.
22 We call this a corollary because it follows directly from the jury theorem based on the easy/
hard dichotomy in Dietrich and Spiekermann (2013b) and the ‘Remark’ on the New Jury Theorem
in Dietrich and Spiekermann (2013a). Unlike Dietrich and Spiekermann, we focus only on the
asymptotic part here and therefore do not need the more restrictive homogeneity assumptions of
Dietrich and Spiekermann’s ‘New Competence’.
23 As they are, under some mild additional technical assumptions; see Theorem 1 in Dietrich
and Spiekermann (2013b).
24 We omit the knife-edge case of ω = 0.5 for simplicity.
1.0
ω = 1.0 (classic CJT)
0.9
ω = 0.9
0.8
Pn
0.7
ω = 0.7
0.6
0.5
0 200 400 600 800 1,000
n
Figure 5.8 Convergence of probability of the majority among n voters being correct,
for various values of the probability that the decision situation is truth-conducive (ω),
shown with homogeneous pBR = 0.55.
best responder being correct, as well. Similarly, if the probability that the decision
situation is truth-conducive is only ω = 0.7, then that once again represents the
best that can be done by the best responder or by the majority of any group of
voters (however large) tracking him.
5.4 EPISTEMIC IMPLICATIONS OF THE BEST

RESPONDER COROLLARY
The Best Responder Corollary is surprisingly powerful for analysing the

epistemic performance of groups in quite different situations. Let us here briefly
sketch some settings that will become relevant in later parts of the book to
demonstrate the wide reach of the corollary.
5.4.1 A Single Common Cause
This is the simplest setup in which a common cause undermines voter

independence.
Assume that the voters have no independent access to the state of the world.
Instead, all information they receive is provided by one single common cause
influencing them all. Suppose the common cause tends to provide useful
information, but does not do so without fail. By assumption, the best responder
identifies the correct solution if and only if the information provided by the
common cause is truth-conducive. Let the average voter track the best responder
effectively and assume that, conditional on the common cause, the votes are
independent.
This means:
1. The best responder has a probability ω, the probability that the common
cause is truth-conducive, to indicate the true state of the world.
2. The voters are best responder trackers with probability pBR > 0.5 .
3. Therefore, according to the Best Responder Corollary, with increasing n,
the group competence (to track the state of the world) converges to ω.
5.4.2 The Evidence-Limited Case
A straightforward and plausible application of the single common cause case

arises if the common cause is evidence, as in Figure 5.4. The group competence yet
again converges on the probability ω, which can now be interpreted as the prob-
ability that the evidence is truth-conducive, that is, such that the best responder
will arrive at the correct solution when provided with the evidence. This is an
intuitive insight: the group can only be as good as the evidence available.
5.4.3 Multiple Common Causes Cases
Suppose that each voter is influenced by the same set of multiple common
causes. The best responder bases her judgement on the information provided
by all these common causes. The average voter tracks the best responder of the
decision situation effectively. Also, conditional on the common causes, the
votes are independent.
This leads us to the following analysis:
1. The best responder has probability ω to vote for the correct alternative
where ω is the probability that the decision situation with all common
causes is truth-conducive.
2. The voters are best responder trackers with pBR > 0.5 .
3. According to the Best Responder Corollary, with increasing n, the group
competence (to track the state of the world) converges to ω.
As a special case case of multiple common causes, assume that all common
causes provide a binary signal that indicates which of the two alternatives is
correct, just like a vote.
Let the vote-like signals of the common causes be competent with pCC > 0.5
and independent conditional on the state of the world.25 Further assume that the
voters make use of the information provided by the common causes by always
voting with the majority of the common causes (and tossing a coin to break ties).
These assumptions provide us with more information about the probability ω:
it can be calculated precisely with the standard CJT competence formula, provided
we are given the number of causes c and their competence pCC. In addition, we
know that, as the number of common causes c grows, the group competence
among the causes grows and converges to 1.
Following a great many competent binary common causes therefore leads to
ω approaching 1, which in turn lets group competence approach 1, provided
that the voters effectively track the best responder. If the voters must be influenced
by common causes (such as opinion leaders or cues) at all, then it is typically
better to be influenced by many independent ones rather than fewer.
Now go back and reconsider Figure 5.6 in this light. We initially introduced
that figure to illustrate its violation of the assumption of independence among
voters themselves; and that is certainly true. Yet notice that the common causes
‘OL’ , ‘Heuristic’, and ‘Cue’ all are independent of one another conditional on
Evidence. And the more such independent common causes there are, the
more likely it is to be the case that a majority of them (and a majority of voters
following them, by extension) is correct.
5.5 SOME PEOPLE ARE IN BET TER DECISION

SITUATIONS THAN OTHERS
As we said in introducing the Best Responder Corollary, how well the best
responder can do epistemically is a function of the decision situation at hand. And
it bears emphasizing that some people can be in better—more truth-conducive—
decision situations than others. The evidence before them is simply a lot better
than the evidence before others.
This suggests two potential sources of differences between experts and
laypersons. One is that the experts may be more competent, in the sense of more
closely replicating the judgement of the best responder to any given decision
situation, in ways discussed earlier. The other (and arguably more important)
difference between experts and laypersons might be that experts face a better—
more truth-conducive—decision situation than laypersons. The difference is
between experts being more competent at interpreting the same body of
evidence and them having a larger and more reliable body of evidence available
to them for interpretation.
25 Let the common causes have direct access to the state of the world.
(a) State (b) State
Expert evidence Expert bias Lay evidence Lay bias
Figure 5.9 Experts versus laymen facing different decision situations.
The populace has numbers on their side, of course. But sometimes smaller
numbers of experts have a distinct advantage: they are better informed, or they
are less likely to be confused or to be overwhelmed by the plethora of information
and misinformation available. More technically speaking, the experts track a
best responder with better evidence; the ω variable for the experts is therefore
greater than the ω variable for the population.
Figure 5.9 visualizes some potential reasons for why experts can face a more
truth-conducive decision situation. Because they are experts, they face ‘Expert
Evidence’, with access to better information and better cognitive tools to pro-
cess that information, as in Figure 5.9(a). This influence can be quite strong, as
indicated by the bold arrows. They are also influenced by ‘Expert Bias’. If the
experts are too similar in background, education, theories, ideologies, or data
used they will be dependent in their judgement to some extent. But if the group
of experts is chosen well, this dependence may not be too strong.
The populace by contrast—although confronting exactly the same ‘State’ as the
experts—only has access to inferior ‘Lay Evidence’, as indicated in Figure 5.9(b).
In addition, they are influenced by potentially quite strong ‘Lay Bias’ (as indi-
cated by the bold arrows). This may be due to their inexperience, the fact that
they are less able to invest time into the investigation of the relevant facts and
have to rely on cues or second-hand information, or even deliberate attempts
by interested parties to manipulate the populace.
The decision situation could be much improved for the population if they
could weaken the influence of the lay biases, while at the same time benefiting
from expert evidence. This may well be possible when experts first provide advice,
rule out implausible options, weaken biasing factors, and help to facilitate
an evidence-based public debate. In the ideal scenario, the population would
benefit from expert evidence, shielded from preventable lay biases, and also
make use of their superior numbers. Whether this strategy succeeds depends
on many empirical facts. But if it succeeds it would be better than just leaving
the decisions to experts, who are not only less numerous but also often influenced
by expert-specific biases of their own.
Part II
Epistemic Enhancement
Traditional discussions of the classic Condorcet Jury Theorem tend to suppose

that there are just two ways in which we can improve the epistemic perform-
ance of groups in making decisions. One is simply to increase the number of
voters.1 The other is to make the existing individual decision-makers more
competent than they previously were.
The latter strategy comports well with long-standing programmes of ‘civic
education’ in response to worries about voter competence.2 Most of those
worries are fixated on factual, propositional knowledge of voters. It is easy to
make monkeys of voters by asking them ‘easy’ factual questions that they get
wrong, in embarrassingly large numbers. But as we point out in Chapter 6,
voter competence—in the sense of voting for the correct alternative—is what
counts for the purposes of the Condorcet Jury Theorem. And as we show in
Chapter 6, that can be relatively high even in the face of substantial ignorance
about particular propositional facts about politics.
The two ways sketched above do not exhaust the options available to increase
group competence, however. A third generic way in which group competence can
be improved is by systematically putting voters in epistemically more favourable
circumstances. A fourth is by reducing the dependence of people’s votes on
the votes of one another or on common causes that affect the votes of many
people at the same time.
There are various ways of enhancing the epistemic competence of the elect-
orate that operate at this more systemic level. One, discussed in Chapter 7, is by
increasing diversity within the electorate. Another, discussed in Chapter 8, is
by dividing epistemic labour. Yet another, discussed in Chapter 9, is by promoting
discussion and deliberation across the community. Through the first two devices
we can simplify the epistemic task, reduce the influence of any particular biases,
and increase the quality of evidence or educate voters to interpret the evidence
better. Through the last device we can enrich the decision situation in ways that
can be epistemically advantageous.
1 So long as their mean competence does not decrease too quickly; see Section 15.2.
2 Galston 2001. Civic education is a long-standing (Farr 2004), and recently renewed (Macedo
et al. 2005), concern of the American Political Science Association.
Improve Individual Competence
The Condorcet Jury Theorem shows us how groups can be far more competent
than are the individuals who comprise them, just so long as the group is com-
posed of a large number of individuals whose votes are minimally (better than
random) competent and independent of one another. Still, group competence
is built on individual competence; and the more competent its members are
individually, the fewer members will be required for the group to achieve any
given level of competence. So it is a natural first thought, when asked how to
enhance the competence of the group, to explore measures for improving the
competence of individual members within it.
In the ways that matter most for CJT purposes, that might not be asking
terribly much. After all, remember the lesson of Section 3.1.1: we do not have to
improve the competence of each individual, only the average. Furthermore,
we do not have to improve mean individual competence to any huge degree
over random. Remember Figure 2.1, showing that (even in a group with only
100 members) bumping mean individual competence up from pc = 0.51 to pc = 0.55
will help group competence enormously. Modest gains of that sort may not be all
that difficult to achieve. More heroically bumping mean individual competence
up from pc = 0.6 to pc = 0.7 will make much less difference, in comparison.
Still, the evidence suggests that voters are pretty wedded to their errors
and hard to shift away from them once they are set in their ways. Over the
very long term, civic education in schools may make a difference, catching
kids before they are set in their ways and attuning them to politics more
effectively.1 But that is a pretty long-haul strategy. All in all, if we are looking
for promising shorter-term strategies for epistemic enhancement of the
electorate as a whole, we would probably do better to look elsewhere.
That will be the more pessimistic conclusion of this chapter. The more optimistic
conclusion is this: none of the evidence standardly offered for thinking that individ-
ual voter competence is urgently in need of improvement is itself remotely con-
clusive. Superficial appearances to the contrary notwithstanding, it may well be
that voters in general are already competent enough for the purposes of the CJT.
1 On the effectiveness of this, opinion is divided: cf. Langton and Jennings (1968) and Niemi
and Junn (1998).
6.1 D ON’ T WORRY
There is a long tradition of making fun of the ignorance of voters, and that has
only intensified with the advent of sample surveys. ‘Voter information scales’
have been constructed, showing time and again that voters are woefully lacking
in information we think that any informed citizen really needs to know. But
why think that voters need to know as much as ‘we political junkies’ know? It
may well be that voters need to know far less than the standard information
scales measure, in order to vote competently in ways that matter for the pur-
poses of the CJT.
6.1.1 Voter Ignorance
Humorist H. L. Mencken once famously quipped:

Democracy is a pathetic belief in the collective wisdom of individual ignorance.
No one in this world, so far as I know—and I have researched the records for years,
and employed agents to help me—has ever lost money by underestimating the
intelligence of the great masses of the plain people. Nor has anyone ever lost public
office thereby.2
We beg to differ with the Sage of Baltimore, and with a raft of political scientists
who have made careers out of ‘proving’ him right.
It is not that we disagree with them that it is easy to show that the average
voter is shockingly ignorant of basic factual propositions that, for the cognos-
centi, seem fundamental to making an informed choice in politics.3 At the
height of Cold War tensions, only 38 per cent of Americans knew that the USSR
was not a member of NATO.4 A 1980 survey found that roughly the same
proportion of Americans believed that Israel was an Arab nation5—and so
on, and so forth.6 We fully agree: the depth of ignorance is enough to make
grown men weep.7
Realistically, there may not be much that can be done to increase the propos-
itional knowledge that voters have about matters of consequence when voting.
The reason is as Anthony Downs gave, many years ago: the chances of any given
individual being decisive to the electoral outcome, in any large electorate, is
vanishingly small. So even if acquiring more information would enable the
2 Mencken 1926; see further Mencken 1922.

3 Charges made by many casual observers over the years and reiterated with evidence since the
beginning of modern survey research (Berelson et al. 1954, p. 308; Campbell et al. 1960, p. 170;
Converse 1964).
4 Page and Shapiro 1992, p. 9. 5 Sniderman et al 1991, p. 15.
6 See e.g.: Delli Caprini and Keeter 1996, esp ch. 2; Achen and Bartels 2016.
7 Conspicuously among them, of late, Caplan (2007) and Somin (2006; 2013). Jason Brennan
(2011b, ch. 7) relies heavily on those works.
Improve Individual Competence 87
voter to cast a more correct vote (even just from his own self-interested point of
view), the costs of acquiring that additional information would ordinarily
vastly outweigh the benefits of doing so, after discounting those benefits by the
improbability that his making a more informed vote would make a difference
to the outcome of the election.8 There might be more hope, however, if one
embraces an expressive theory of voting: it may well be the case that not many
people like being wrong, whether they influence the outcome or not. Such an
expressive motivation could lead to some efforts to inform oneself, in the same
way as an expressive theory of voting may explain why people vote despite the
extremely low chances of any given voter’s changing the outcome.9
Still, voters may well know enough to vote accurately, notwithstanding
their embarrassing ignorance of those particulars so beloved of the political
cognoscenti. As Lupia rightly points out, it is not strictly necessary for you to be
able to name the Chief Justice of the US Supreme Court in order for you to
know whether you approve of the Court’s major decisions—and still less to
know what party you want to be in control of the choice the next members of
that court.10 Even if they seem woefully ignorant about specific policy proposals
and confused about what the politically sophisticated see as the overall con-
tours of ‘ideological space’, voters can be (and arguably are) nonetheless guided
very effectively by abstract ‘policy principles’.11
Furthermore, there are good reasons to think that voters vote accurately in
aggregate, notwithstanding the depths of their individual ignorance.12 The
explanation again is due to Downs. While any sensible cost–benefit calculus
would lead people not to go out of their way to collect additional evidence
purely to improve the quality of their vote, the same calculus should lead them
to acquire all sorts of information relevant to progressing their own private
concerns in their private lives where such information becomes more immedi-
ately relevant to them. This information, in turn, has implications for their
political choices.13 When aggregating, uninformed votes constitute ‘noise’ and
8 Downs 1957, chs 11–13. See further Wittman 1995, ch. 2.

9 Brennan and Lomasky 1993. 10 Lupia 2006; 2015, ch. 1.
11 Goren 2013, p. 6 and passim.
12 Page and Shapiro (1992) and Erikson et al. (2002) offer a wide range of evidence of that. Feld
and Grofman (1988) show that the placement of candidates on the ideological spectrum is almost
exactly correct when averaging across all voters’ placements, although the individual placements
were often wildly wrong.
13 This is an old thought. Galton was confronted with it, in reply to his Nature letter reporting the
‘guess the weight of the ox’ contest we mentioned in Chapter 1. One F. H. Perry-Coste (1907), from
Polperro in Cornwall, wrote in reply: ‘I do not think that Mr. Galton at all realizes how large a per-
centage of the voters – the great majority, I should suspect – are butchers, farmers, or men otherwise
occupied with cattle. To these men the ability to estimate the meat-equivalent weight of a living
animal is an essential part of their business . . .’ (cf. Galton 1907a). Hochschild and Einstein (2015a,
ch. 2) quote Weingarten (1996) pointing out that, while ‘40 percent of adult Americans may be
unable to name the vice president . . . “72 percent of the residents of greater Helena, Montana, were
able to identify, on one of those creepy diagrams, every known slice of cow” ’. We are unsure of the
relevance of that to their voting choices (although we note, in puzzlement, that one of VP candidate
cancel out; votes informed by this private information constitute the ‘signal’
that is left after that.14
Thus, we should not leap to the conclusion, based on survey results showing
that large numbers of people are factually ill informed about matters of great
consequence politically, that voters are ignorant in ways that would compromise
the Condorcet Jury Theorem’s results. The electorate as a whole can still be com-
petent enough for the CJT to apply in two ways. The first of these is if most or all
of the voters are just a little bit better than random in voting for the correct
alternative. The second is if many voters are completely incompetent and vote in
random ways, but a sufficient minority of voters is well informed and pushes the
collective result in the right direction. This latter setting works because, as long
as the many who are wrong are wrong in ways that are random,15 they cancel—
and as long as there are enough who are well informed and whose votes decide
the contest after uninformed votes cancel, the CJT result still goes through.
Whether or not there is enough competence present in the electorate is of
course an empirical question.16 The classic paper testing this proposition comes
from Larry Bartels.17 His methodology was to split respondents in successive
American National Election surveys into multiple demographic groups and to
examine the voting intentions of those in each demographic who subjectively
seemed to interviewers to be relatively well or poorly informed. Bartels takes
that as his measure of how a well-informed voter in that demographic would
vote. Bartels then calculates what difference it would have made to the electoral
outcome if all members of each demographic had voted the same way as the
informed segment of that demographic voted.
The assessment as to whether a voter has voted correctly is tricky, of course.
To make such a judgement, we not only need to make the theoretical assumption
that an independent standard of correctness exists, but we also need to know
what the correct decision would be for each specific election analysed for each
demographic. Unsurprisingly, Bartels sidesteps these issues, even though he
briefly mentions the CJT in his introductory discussion.18 Instead, Bartels tests
a more restricted hypothesis—that informed voters cast systematically different
votes than uninformed voters.
Sarah Palin’s prime claims to office seemed to be that ‘she knew how to field dress a moose’—‘How
do you do that? In white tie?’ one East Coast wag quipped self-deprecatingly in reply).
14 Silver 2012. ‘Noise’ is assumed to be randomly distributed. Uninformed votes might not
be, owing to some common cause, such as the Kennedy campaign’s false claim that there was
a ‘missile gap’ between the US and USSR (Page and Shapiro 1992, pp. 226–7, 367; Kinder 2006,
p. 212).
15 Insofar as voters are simply ignorant, they will vote randomly. Ignorant voters may be more
easily manipulated, or more prone to prejudice, and if so they might not vote randomly after all.
But that is the result of a violation of the Independence Assumption, and is not (purely) the result
of a violation of the Competence Assumption.
16 As Somin (2006; 2013) is right to emphasize. 17 Bartels 1996.
18 A more direct experimental investigation of the epistemological questions is offered by Lau
and Redlawsk 1997.
Nevertheless, Bartels’s research raises interesting epistemological issues.

Suppose the candidate whose winning would be in the best interest of most
voters is the objectively correct candidate to choose.19 Against that backdrop,
Bartels’s results are a cause for concern from a truth-tracking perspective. If
(1) most voters are typically uninformed, and (2) we assume that more
informed voters are more likely to vote for the alternative best for them, while
(3) informed voters choose systematically differently from uninformed voters,
then our typical election with largely uninformed voters might not track the
truth as to what is best for most. Whether a majority of largely uninformed
voters can still track the truth depends on whether the uninformed voters are
systematically biased or whether they just introduce random noise. The for-
mer is, of course, problematic in a way that the latter is not, from a CJT point
of view.
Here is Bartels’s own report of his findings:
In four of the six [US presidential] elections examined [1972 through 1992], the
aggregate deviations from hypothetical ‘fully informed’ election outcomes are
both large (with absolute deviations ranging from 2.7 to 5.6 percentage points)
and statistically significant (with p-values of .16, .09, .02, and .02 for separate two-
tailed t-tests of the null hypothesis of no aggregate deviation).20
In short, Bartels’s conclusion is that the errors made by uninformed voters do
not always completely cancel out (although it should be noted more firmly than
he does himself that in two out of those six elections they pretty much did so).
But from a substantive point of view, do they come close enough to doing so
as not to matter, politically? That would be the case if the difference ignorant
voters’ errors made to the outcome of the election was less than the winner’s
margin of victory. We examine this matter in Table 6.1. That table displays: in
the first column, the difference Bartels estimates it would have made to the
electoral outcome if all voters had been as fully informed as the most informed
members of their own demographic;21 and in the second column, the winning
candidate’s margin of victory.22 The final column sums those first two, thus
showing what would have happened in that election if all voters had been as
informed as the most informed members of their demographic group.
In no case does the adjustment in Table 6.1 change who wins the election.23
Furthermore, in four out of those six elections, the adjustment would only serve
to increase the margin of victory for the party that won the election anyway.
Only in two elections—the ones in which the winner’s actual margin of victory
19 As discussed briefly in Section 4.1.4; we return to this in Chapters 13 and especially 14.
20 Bartels 1996, p. 218.
21 From the second column in his Table 3 (Bartels 1996, p. 216).
22 Calculated as the percentage of popular vote for the winning candidate minus the percent-
age for the next-nearest candidate.
23 A fact also noticed by Althaus (2003, p. 126, n. 18) and Bendor and Bullock (2008, p. 13).
Table 6.1 US presidential election outcomes if all voters voted the same way as
informed voters of the same demographic.
Presidential % change in outcome winning candidate’s adjusted margin (%) of
election if all voters voted like actual margin (%) victory if all voters voted
informed voters of victory like informed voters
1972 + 1.71 to Dem. Rep. by 23.15 Rep. by 21.44

1976 + 0.35 to Dem. Dem. by 2.07 Dem. by 2.42
1980 + 5.62 to Rep. Rep. by 9.74 Rep. by 15.36
1984 + 4.87 to Dem. Rep. by 18.21 Rep. by 13.34
1988 + 3.01 to Rep. Rep. by 7.72 Rep. by 10.73
1992 + 2.73 to Dem. Dem. by 5.56 Dem. by 8.29
was huge (1972 and 1984)—would the adjustment serve to reduce somewhat
the winner’s margin of victory.
In short, even if voting errors introduced by uninformed voters are not
completely random and do not completely cancel out, at least in the six elec-
tions studied by Bartels they come close enough to doing so as not to make any
substantive difference to the electoral outcome. Bartels may be right that the
difference between how informed and uninformed voters vote is, at least in
some elections, statistically significant. But statistical significance is one thing,
substantive difference another.24
Using a similar methodology to Bartels, Scott Althaus claims to have found
that the collective preference would have been different on fully a fifth of all
policy questions in the 1988 and 1992 American National Election Studies,
had every voter the same preferences as his or her demographic counterparts
with the highest levels of (politics quiz-style) political knowledge.25 That
report is highly misleading, however. Here is why. Althaus considers ‘three
possible collective preferences . . . : a majority in favor, a majority opposed, or
a tie between opinions’, where ‘a tie was defined as any marginals falling
within plus or minus 3 percentage points of the 50% mark’.26 His count of
‘different collective preferences’ includes transitions from any of those states
to any other.
On that basis, Althaus reports differences between ‘actual and fully informed
preferences’ on fully nine out of the forty-five issue questions in the ANES
studies.27 But upon closer inspection, it turns out that the vast majority of those
differences amount to movement into or out of that artificial ‘tied’ category. If
you drop the bogus ‘tie’ category and count as ‘preference differences’ only
cases in which the majority opinion changes from one side to the other of the
24 For other critiques of Bartels’ approach on different grounds, see Landemore (2012, pp. 272–82)
and Mackie (2012).
25 Althaus 1998, p. 552. The figure rises to a quarter when, in subsequent work, Althaus (2003,
p. 23 and ch. 4) adds the 1996 ANES study to his sample.
26 Althaus 1998, p. 552 and n. 9. 27 Althaus 1998, p. 552, Table 2.
question, there are only two genuine collective ‘preference differences’ on the
forty-five issue questions under discussion.28
Philip Converse, the reluctant poster child for the ‘voter ignorance’ hypoth-
esis, hails these two studies as ‘path-breaking studies that should herald a wider
range of research’.29 We are always happy to sign up to the proposition that ‘more
research would be good’. But for now, we are singularly unimpressed that this
strand of ‘cutting-edge research’ really has established what it is thought to have
established, which is that collective judgements would often have been very
different if all voters had had more information.
6.1.2 Knowledge ≠ Competence
A second reason not to worry that voters lack seemingly crucial propositional
knowledge about politics is this: what matters for the CJT result to go through
is not knowledge, but competence.
Competence is the tendency to make the correct choice—nothing more,
and nothing less.30 Once you know enough to vote the right way, further
knowledge is simply superfluous for the purposes of the task at hand. Judging
the adequacy of voters’ knowledge on a need-to-know basis, it is simply not
the case that voters need to know the ‘minute details that excite [the] sophisti-
cated’ in order to perform effectively the task before them, which is to make
the correct choice.31
Voters can be competent (at voting for the correct option) without being
knowledgeable (well informed about myriad detailed facts about politics) for any
of many reasons. Perhaps the most important—to be discussed more fully in
Chapter 12—is that uninformed voters take cues. Those may be cues from others
who are more informed. Voters may not know the right answer themselves, but
maybe they know who knows. Alternatively, voters may cue on other elements in
the external environment: on party labels, or some particularly salient factoid, or
‘the look of the guy’.32 By relying on these sorts of cues, relatively uninformed
voters might be able to exercise political judgement more reliably than they
would by relying on their own private information alone.33
28 Althaus 1998, pp. 554–7, Appendix B. Althaus (2003, pp. 325–6) employs the same procedure,
so presumably those findings are similarly flawed—although full item-by-item results are not
reported there, making it impossible to check.
29 Converse 2000, p. 351.
30 I.e., ‘the same vote she would have cast if she had all relevant information available at the
time of her choice’ (Lupia 2006, p. 221; see further 2015, ch. 3).
31 Popkin 2006, p. 253. Similarly, what Goldman (1999, pp. 320–39) dubs ‘core voter know-
ledge’ is simply the knowledge of which option to vote for; as he goes on to say, ‘Nothing follows
concerning other types of political knowledge on the part of voters’ (p. 329).
32 Popkin and Dimock 1999, pp. 125, 127. See further Popkin 1993.
33 It just depends on whether the cue is more likely to be right than wrong in the direction it
points the voter.
6.2 PRIORITIES IN IMPROVING

VOTER COMPETENCE
All that is simply to say that the extant levels of voter competence across the
electorate are probably ‘high enough’ for the CJT to work. Evidence about
propositional ignorance at the individual level is simply not enough to prove
decisional incompetence at the aggregate level.
Our next discussion is meant to be similarly reassuring. Its theme is that huge
improvements in competence are not necessary. Given the way that the math-
ematics of the CJT calculations work, improving mean voter competence from
‘barely better than random’ to ‘just a bit better than that’ is what most matters.
Further improvements in average voter competence beyond that are nice but are
decreasingly consequential for the group’s reaching the correct conclusion.
Note well that it is mean individual competence of voters that matters for
CJT purposes (see the discussion in Section 3.1.1). Thus, we do not necessarily
need to improve the individual competence of all or even the majority of voters
to achieve substantial effects. Instead, we can improve mean competence in
whatever manner is the most efficient, by focusing on increasing the compe-
tence of voters who are most responsive to such measures. Sometimes that may
involve focusing on those who are already quite competent (perhaps because
they are more interested). Sometimes that may involve focusing on those who
are least competent and whose competence can therefore be increased more
easily (perhaps because ‘silly’ errors are easier to eradicate). Sometimes there
may be many more people in the middling groups whose competence could
be improved by one and the same civic education campaign—in which case
more of an improvement in mean competence across all individuals might be
achieved by concentrating on doing that.
No matter what the most efficient target group is, increasing the competence
of some can lead to remarkably high increases in group competence—as
Figure 6.1 demonstrates. The dark grey area shows the competence distribution
of 300 voters. On average, these voters have competence pc = 0.51, but many of
them (on the left) are significantly below and some (on the right) significantly
above that competence level. With this competence distribution, the group has
competence Pn = 0.64 , which is about the same competence that would arise if
all voters had homogeneous competence pc = 0.51, in line with the results on
heterogeneous competence presented in Section 3.1.1.
Our interest here is in comparing different competence-raising interven-
tions. In each case, we ensure that the average competence increases by one
percentage point, to pc = 0.52. As a baseline (not shown in Figure 6.1), consider
what happens if we increase the competence of each individual by one percent-
age point—in which case the group competence increases to 0.76. We compare
this with three other scenarios in Figure 6.1 (marked ‘1’, 2’, and ‘3’). In scenario 1,
the individual competence of the hitherto least competent 10 per cent of voters

0.9
0.8
Pn = 0.76
0.7
Pn = 0.76
3
0.6
Pn = 0.75 2
pc
0.5
1
0.4
0.3 Pn = 0.64
0.2
0 50 100 150 200 250 300
Voter
Figure 6.1 Different distributions of individual competence and the resulting group
competences.
is increased by 10 percentage points each, as indicated by light grey area 1—thus

increasing the average competence of the group as a whole by 1 per cent once
again. The resulting group competence of that intervention is Pn = 0.75 , which
is virtually the same as that resulting from a uniform 1 per cent increase of
competence across the group as a whole. In scenario 2 we increase the compe-
tence of the middle 30 among the 300 voters by 10 percentage points; and in
scenario 3 we do the same for the most competent 30 voters. In both scenarios
the group competence is Pn = 0.76 , which is exactly the same as in the base.
The bottom line of this example is just this. Apart from some freak cases
(which are only likely to occur in small groups), it does not matter much how
the average competence pc gets increased, as long as it increases. That is very
good news from a public policy perspective, because it suggests that we do not
need to succeed in making everyone more competent. Increasing the compe-
tence of a minority ordinarily suffices.
A final way of improving mean individual competence among voters would
be by the less competent simply refraining from voting. Voting is compulsory
only in a very few countries in the world. And studies show, time and again,
that ‘knowledge about politics stands out as a consistently strong factor shaping
the decision to vote’.34 In the 1988 US presidential election for example ‘nearly
nine out of ten of the most knowledgeable ten percent of respondents voted; by
34 Popkin and Dimock (1999, p. 142) continue: ‘The dominant feature of nonvoting in America
is lack of knowledge about government; not distrust of government, lack of interest in politics,
lack of media exposure to politics, or feelings of inefficacy.’
comparison among the least informed decile only two [out of ten] did so’.35 For
CJT purposes, of course, what matters is mean individual competence of those
who do vote (not of those who are entitled to vote). So the more of those whose
own individual competence is low who do not take part in the election, the
higher the mean individual competence is of those who do vote.
6.3 WHAT MORE CAN BE D ONE TO IMPROVE

INDIVIDUAL COMPETENCE?
As we have said, the most useful thing to do from a CJT perspective is to improve,
if only marginally, mean individual competence in electorates where that is
barely better than random. How exactly can we do that?
One natural way to approach that question is to look at what we know about
dynamics surrounding the acquisition of propositional knowledge more gen-
erally. Evidence from there is not very encouraging.36
Studies show that it is hard for people to unlearn erroneous facts, once they
have embraced them.37 A non-negligible proportion of people insistently ignore
correct knowledge even after it is presented to them and are nonetheless polit-
ically active in pressing their misinformed agenda.38 Among the underlying
reasons for such an active resistance to available evidence are biased media
reports, segregation into echo chambers, and a remarkable individual ability to
avoid factual information that does not fit into one’s world view.39 The fact that
people are persistent in their errors—seemingly impervious to updating their
factual beliefs in light of new evidence—and insistent in acting politically on
those errors is bad news from an epistemic-democratic point of view.
If unlearning errors is hard, the better strategy might seem to be to put
voters on better epistemic footing from the start, through civic education as
part of the formal school curriculum and through encouraging informal dis-
cussions of current events there. Evidence suggests that the former alone
increases overall political knowledge by 4 per cent and, when combined with
the latter, by 11 per cent.40 But of course it will take a long time to make much
difference to mean individual competence across the community, improving
the mean competence of successive waves of seventeen-year-olds one at a
time—even on the assumption that those improvements in propositional
knowledge translate into similar improvements in competence.
35 Delli Carpini and Keeter 1996, p. 224.

36 Sunstein and Hastie 2014. 37 Nyhan and Reifler 2010; 2015. Flynn et al. 2017.
38 Kuklinski et al. 2000. Hochschild and Einstein 2015a, ch. 3; 2015b. Sunstein and Hastie 2015.
39 See Anderson (2012) for a detailed analysis of these three factors.
40 Niemi and Junn 1998, ch. 7. See also: Galston 2001; Nie and Hillygus 2001.
Beyond formal schooling, there is ‘learning from experience’. The more

people are called upon to engage in politics, and the more they answer that
call, the more they will come to know about the political issues before them.
John Stuart Mill famously made this claim in defending the extension of the
franchise in the nineteenth century, and this was a familiar theme among latter-
day participatory democrats.41 Evidence from places with direct democracy
seems to support the conjecture.42 We will later offer some modelling to sug-
gest how genuinely helpful this effect might be, in Section 15.6.
41 Mill 1861/1977, ch. 8, pp. 467–9. Participatory democrats echoing the claim include: Pateman
(1970, p. 27); Thompson (1970, pp. 19–22); and Macpherson (1973, esp. chs 1, 3).
42 Benz and Stutzer (2004); Smith and Tolbert 2004; and Donovan et al. 2009.
Diversity
Multiculturalists have taught us many reasons to welcome diversity in our

social environment. In part it is a matter of fairness to the Other: insofar as
there are diverse people present, it is only proper that they be permitted to
display their authentic selves in social space rather than living a lie.1 In part it is
a matter of enhanced enjoyment for Self to have a more interestingly diverse
social environment (did anyone ever get a good meal in Australia before
multiculturalism?).2
While those other considerations are also in play, our focus in this book
is purely on the epistemic implications of diversity within the group whose
opinions are being aggregated to form a collective judgement. Diverse groups
do not always outperform less diverse but more competent groups. Still, diver-
sity clearly has an epistemic contribution to make, at least in certain (perhaps
tightly confined) situations, as we shall show in this chapter.
The epistemic value of diversity has long been appreciated. In Federalist no. 73,
Hamilton discussed the advantage of the US president’s qualified power to veto
legislation in precisely these terms:
The oftener the measure is brought under examination, the greater the diversity in
the situations of those who are to examine it, the less . . . probable, that culpable
views of any kind should infect all the parts of the government at the same moment
and in relation to the same object, than that they should by turns govern and
mislead every one of them.3
Drawing on his experience as sometime head of the White House Office of
Information and Regulatory Affairs, Cass Sunstein reports that just such
diverse perspectives are fruitfully brought to bear in deliberations across the
US Executive Branch over proposed new regulations, for example.4
Diversity of knowledge bases among people within the group can be an
epistemic advantage for two reasons. One is that, when different people know
different things or search in different places, and each recognizes the truth
1 Parekh 2000. 2 Goodin 2007b.

3 Hamilton, Federalist no. 73, 1788. 4 Sunstein 2017b.
Diversity 97
when it is presented, such a group is in a better position to deal with a larger class
of problems. The larger number of knowledge bases means that a larger num-
ber of candidate ‘truths’ will be presented to the group, and this makes it more
likely that the genuinely correct one will be among those offered to (and hence
chosen by) the group.5
That is the reason why John Stuart Mill attaches so much importance to
‘individuality’ in his essay On Liberty, for example. There he argues
that mankind are not infallible; that their truths, for the most part, are only
half-truths; that unity of opinion, unless resulting from the fullest and freest
comparison of opposite opinions, is not desirable, and diversity not an evil, but a
good, until mankind are much more capable than at present of recognising all
sides of the truth. . . .6
A second way in which diversity can improve the epistemic performance of a
group is by providing a corrective to any positive correlation among voters7
(or even a supplement, in the form of providing negative correlation among
them) in judging which alternative is correct. Less positively correlated—and
in that sense, more diverse—voters tend to have the epistemic edge. Extending
the decision-making group beyond a narrow set of people sharing much the
same training and experiences is epistemically valuable, in that way.8
Those are clearly two very different mechanisms, and we shall discuss them
over the course of two separate chapters (negative correlation in this chapter,
diverse search strategies in the next). Unfortunately, those two very different
mechanisms are typically conflated in popular discussions of the ‘diversity
trumps ability’ theorem owing to Hong and Page.9
5 That is why we need racially diverse juries to try cases in racially diverse communities, for
example (Lever 2016, pp. 6–9).
6 Mill 1859/1977, p. 260. For an extravagant extension of that sort of claim to ‘a dialog about
transrational (spiritual) knowing within socio-ecological decision making’ see Barrett (2013).
7 Putting the point somewhat colloquially. Strictly speaking, the correlation in question is
between votes, not voters (Sections 5.1.1); and the correlation in question is that that remains
‘even after conditionalizing on the state of the world’ and perhaps ‘on the evidence’ as well
(Sections 5.1.3 and 5.2.1).
8 Vermeule 2011a, p. 1454. Lever 2016. Early theorists of the English common law such as Hale
(1716/1971, p. 252) thought it was a great advantage in keeping ‘both the rule and administration
of the law of the kingdom uniform’ that
those men are employed as justices, who as they have had a common education in the
study of law, so they daily in term-time converse and consult with one another; acquaint
one another with their judgements, sit near one another in Westminster Hall, whereby
their judgements are necessarily communicated to one another, and by this means their
judgements and their administrations of common justice carry a constancy, congruity and
uniformity one to another, whereby both the laws and the administrations thereof are pre-
served from the confusion and disparity that would unavoidably ensue, if the administration
was by several uncommunicating hands, or by provincial establishments.
Be that as it may, there was an epistemic cost to achieving uniformity of the law in these ways.
9 Hong and Page 2004; Page 2007.
There are several reasons to be wary of following that theorem too closely.
One distinguished mathematician has argued that the theorem is trivial; that
the interpretation of the theorem in terms of diversity, as ordinarily under-
stood, is questionable; and that Hong and Page’s computational exercises are
based on very particular modelling assumptions.10 Of course, every model
idealizes and simplifies to some extent, and different specifications can lead to
different results.11 What worries us, however, is how frequently the theorem
has been misunderstood in the less technical literature—and especially
that it is so rarely acknowledged just how limiting the assumptions of the
theorem really are.12
Nevertheless, something in the neighbourhood of the Hong and Page results
seems likely to be both interesting and at least in some (perhaps not-too-
common) cases important.13 In this and the following chapter, we offer some
more simplified and streamlined models to show what that might be.
7.1 CLONES AND COMMON CAUSES
As the Federalist suggests, less diverse groups tend to arrive at worse decisions
than more diverse groups. That is most obviously true in extreme cases: a group
of perfect clones in which all always vote exactly the same way cannot benefit
from voting and aggregation. In that case, the collective result is only as good
as that of any single group member. Completely positively correlated clones do
not get wiser in numbers since their votes are completely interdependent.
In less extreme cases, a lack of diversity can be interpreted as an unhealthy
dependence on common causes.14 Consider a few examples. In a situation
where all the voters rely on the same TV channel about what the evidence is
(say, Fox News as in Figure 7.1(a)15), even the most competent voters are limited
10 Thompson 2014 and cf. Page 2015; Thompson 2015; Kuehn 2017.

11 For example, in Keuschnigg and Ganser’s (2016) setup, ability trumps diversity in small
(under sixteen-person) groups under plurality rule voting.
12 For instance, the theorem has been linked to binary votes, even though the theorem’s setup
is such that it simply cannot have any bearing on two-option choices (Weymark 2015). Even for
small numbers larger than 2, the theorem is applicable but almost pointless. That Hong and Page
have separate technical results about those kinds of setups has largely been lost in the noise. For
those results see Hong and Page (2009; 2012). The ‘diversity trumps ability’ theorem has been cited
widely and often without acknowledging its limitations; see e.g. Landemore (2013b, ch. 4; 2014),
and Anderson (2006). For a critique of Landemore’s use, see Quirk (2014, pp. 134–44; cf.
Landemore 2014, pp. 214–21 in reply).
13 Differentiating carefully between the search-based arguments for diversity in Hong and
Page (2004) and the signal-based arguments for diversity in Hong and Page (2009; 2012). The
latter are the subject of this chapter, the former of the next.
14 As discussed in Section 4.5 and Chapter 5. 15 Which is analogous to Figure 5.5.
Diversity 99
(a) State (b) State
Evidence Evidence
Fox News Fox News MSNBC CNN
Figure 7.1 One common cause and many common causes.
in their epistemic capabilities by the quality of the Fox News report. If the
report is biased or just plain false, the majority is bound to be incorrect.
The situation in Figure 7.1(b)16 is better in that regard. Each voter watches
two different TV channels, with three channels in total. There are thus more
paths leading from the state of the world to the voters. Even if one TV channel
conveys incorrect information, this can be balanced by the fact that two other
channels report correctly. If there must be common causes at all, it is better to
have many of them that are independent from one another (as discussed in
Section 5.4.3).
Another way to improve the epistemic situation of the voters is for them to
access directly the evidence (or even the state of the world, if possible), thus
partially bypassing the common causes that would lead their votes to be posi-
tively correlated. For instance, suppose voters 1, 2, and 3 watch the same TV
channels as in Figure 7.1(b), but suppose they also investigate the evidence dir-
ectly, so that direct arrows point from the Evidence to the Votes, as in Figure 7.2.
That is a better epistemic scenario, ceteris paribus, because the voters are not
entirely reliant on some TV channel to relay information but, instead, can
access that information on their own. If the voters are good at assessing the
evidence for themselves, it might even be epistemically useful not to have any
of the TV channels involved. But whether that is true depends on many differ-
ent factors: the comparative competence of the voters to interpret the evidence
directly or through a TV programme, the competence of the various TV pro-
gramme makers at tracking the true state of the world, and so on.
Analysing the flow of information in terms of causal networks thus allows
us to distinguish between decision environments that lack diversity and
where voters are highly dependent on one or a few common causes and
decision environments in which different voters have access to different sources
16 Which is analogous to Figure 5.6.

State
Evidence
Fox News MSNBC CNN
Figure 7.2 Several common causes and direct access to Evidence.
of information. The former tend to bring about dependent, positively correl-

ated votes. The latter are more likely to ensure low correlation, approaching the
ideal of votes that are independent conditional on the evidence (if not state of
the world).
7.2 NEGATIVELY CORRELATED VOTES
Our discussion in Chapter 5 has shown that correlated votes are usually
bad news epistemically. But more precisely, it is positive correlation that is bad
news because, if the votes are positively correlated, there is less to learn
from additional votes. Negative correlation is (or anyway can be) a different
matter, however.
7.2.1 Epistemic Benefits of Negatively Correlated Votes
If the votes of overall competent voters are negatively correlated, then that
implies that if someone makes an error someone else is more likely not to
make that error. In other words, when one votes for an incorrect alternative
the other votes the opposite way—thus ‘cancelling’ the first voter’s error.
Negative correlation is therefore epistemically very good news, because it
means that errors tend to be cancelled more systematically than without the
negative correlation.17
17 But note that negative correlation and competence constrain each other. Intuitively, if the
voters are quite competent they tend to agree more often, ruling out strong overall negative
correlation. A more technical account of the bounds can be found in Kaniovski (2010).
Diversity 101
Suppose voters have individual competence pc = 0.75. If all the voters were
correct about all the same alternatives, and incorrect about all the same alter-
natives, there would be a perfect positive correlation among their votes; and the
majority would be right about approximately 75 out of 100 propositions put
to them. If the voters are uncorrelated, they each, independently, vote for the
correct alternative with probability 0.75, which means that the voters agree
often but not always. The group decision is correct, in line with the CJT, with a
probability increasing in group size. For there to be a negative correlation
among votes, voters must vote in different directions from one another more
often than they would without negative correlation. Then whenever voters are
split on some issue like that, their errors tend to cancel each other more often
than they would in the uncorrelated scenario. The correct answer would thus
be chosen by the group more often, with votes being split in that fashion.
In a sense, the pattern emerging here is similar to that which results from
gerrymandering legislative districts. To maximize the number of seats they win
in the legislature, a party wants to draw electoral boundaries in such a way as to
concentrate the opposing party’s voters in a few districts and spread their own
voters widely. Spreading their own supporters in that way, they will win nar-
rowly in many more seats. In terms of controlling the legislature, that is much
more advantageous than winning big in a few seats.
We can think of negative correlation among competent voters in similar
fashion. Instead of the Truth winning most contests by an expected 75 per cent
margin, having correct backers of the Truth split up and pitted against less
numerous mistaken voters on each issue allows the Truth to prevail (albeit by a
narrower margin) more often. Of course, however, for this to work requires
reliably choreographed coordination, just as in the gerrymandering example.
Perhaps even more interestingly, negative correlation can lead to success for
voters with individual competence below 0.5. With sufficient negative correl
ation, the voters can compensate for each other’s mistakes in certain fortunate
(but very special, and not necessarily at all typical) voting patterns.
Some simple numerical examples are offered in Table 7.1.18 In that setup,
there are three voters. The first three columns show all possible vote profiles
(that is, all combinations of the ways the voters can vote). ‘1’ represents a correct
and ‘0’ an incorrect vote. So the profile ‘1,1,1’ represents the case in which each
of the voters votes correctly; the profile ‘1,1,0’ the case in which the first two
voters vote correctly and the third incorrectly; and so on.
Let us compare, first, the independent (uncorrelated) and the negatively cor-
related setting I. For these setups, we assume quite a high level of competence
for all voters, pc = 0.75. Column four (‘independent’) in Table 7.1 shows the
probability of each vote profile if the voters vote independently, conditional on
the state of the world, with competence 0.75. For example, the probability that
18 The example is from Zaigraev and Kaniovski (2012).

Table 7.1 Independent and negatively correlated votes.

Vote profiles Probabilities
independent Negatively correlated
I II
1 1 1 0.4219 0.25 0
1 1 0 0.1406 0.25 0.17
1 0 1 0.1406 0.25 0.17
0 1 1 0.1406 0.25 0.17
Pn 0.844 1.00 0.51
1 0 0 0.0469 0 0
0 1 0 0.0469 0 0
0 0 1 0.0469 0 0
0 0 0 0.0156 0 0.49
all three voters will vote correctly (vote profile ‘1,1,1’) is 0.753 ≈ 0.4219; the prob-
ability that only two of them will vote correctly in each of the three ways that
might happen is 0.75 × 0.75 × 0.25 ≈ 0.1406; and so on. One can see by simple
addition that the majority of voters will be correct with probability 0.844.
Compare the results of such ‘independent’ votes with the ‘negatively correl-
ated votes I’ scenario (column 5 in Table 7.1). For purposes of that example, we
keep individual competence at the same level of 0.75, but we make votes
dependent upon one another in the following very specific way: whenever one
voter votes for the wrong option, the other two voters ‘compensate’ for this
error and vote for the correct option. In this way negative correlation works to
epistemic advantage, enabling the majority to offset individual errors of the
minority in a wider range of cases. Indeed, in this example the majority of
the three voters always votes for the correct alternative, so that Pn = 1.
Now consider the case of ‘negative correlation II’ (column 6 in Table 7.1). In
constructing this case, we intentionally do so in such a way that the group
competence is just above 0.5 (0.51 to be precise), but each individual compe-
tence is much below. The goal of this exercise is a ‘proof of possibility’. The aim
is to demonstrate how even incompetent voters can be collectively competent
if the voters are negatively correlated in just the right way.
For the purposes of this ‘negative correlation II’ construction, we begin by
setting the probability for vote profile ‘1,1,1’ to 0. We do that in order to ensure
that the three voters never agree on a ‘yes’ vote (that would entail positive
correlation). We then complete the construction of this scenario in such a way
so as to ensure that the correct answer wins as often as possible, given that
constraint. To achieve that, we give each of the remaining profiles in which
the correct alternative wins—that is, vote profiles ‘1,1,0’, ‘1,0,1’ and ‘0,1,1’—an
equal probability of 0.17 each. There are three such ways in which the correct
Diversity 103
a lternative might win. So the total probability of a correct majority is 3 × 0.17 = 0.51.
That makes the group a bit more competent than a random coin toss. Interestingly,
however, this group achieves better-than-random group competence even
though every single voter only has a probability of 2 × 0.17 = 0.34 of voting
correctly (since the vote of any given individual is correct in only two of the
vote profiles in question).
The important upshot is that negative correlation can make a group compe-
tent even when all its members are individually incompetent. Column 6 of
Table 7.1 shows how: in the 51 per cent of cases in which the majority is correct,
there is always one incorrect vote; but the mistakes of the one are compensated
for by the correct votes of the other two.
We hasten to add, however, that this is a quite specific constellation, not
likely to be observed often in real-world settings. If the setup in column 6 of
Table 7.1 seems contrived, that is precisely because it is. It is fiendishly difficult
to concoct a scenario in which it is simultaneously the case that individuals are
incompetent but the group is competent, owing to negative correlation among
the votes of the group’s incompetent members. It is possible that that might
occur, but it hardly seems likely.
7.2.2 Are Negatively Correlated Votes Plausible?
Let us start with the simple observation that there are strict logical limits to the
amount of pairwise negative correlation there can be among votes. Furthermore,
the maximum level of pairwise correlation across the group as a whole must
go down with increasing group size. Two voters can be perfectly negatively
correlated with one another in the way that they vote, for example; three voters
cannot. Hence, the more voters there are, the less diversity—understood as
pairwise negative correlation among their votes—is possible.
Even within those limits, questions arise as to why (and how often) the
votes would be negatively correlated. Usually, there are no systematic reasons
for negative correlation, and we will assume that there is none in most other
parts of this book. That assumption is particularly apt insofar as the truth to
be discovered is a plain perceptual fact. In such cases, the solution to the
problem relies, not on any more complex cognitive work (such as reasoning,
applying heuristics, etc.), but merely on the simple operation of one’s own
ordinary senses.
Suppose, for example, the task is to hold two different objects in the hand
and judge which one is heavier. Then the subjects approaching this task will
rely on their immediate perception of relative weight. It seems very unlikely
that the votes in this or similar examples are negatively correlated—their
guesses, informed by their senses, are plausibly independent since there exists
no underlying mechanism that could choreograph voting patterns in the way
needed for negative correlation. In particular, there is no good reason why the
error of voter X should make the correct judgement of voter Y more likely.19
In addition, in many situations the odds are stacked against negative correl
ation because there are so many reasons for assuming positive correlation.
Whenever there is reason to believe that the voters are influenced by common
causes, it is plausible, ceteris paribus, that they are positively correlated. We
already listed a small selection of possible common causes in Section 4.5 and
Chapter 5—opinion leaders, similar heuristics and cues, the same evidence,
reliance on the same theoretical framework, even random influences (such as
being jointly influenced by the latest chart hit) can correlate votes positively.
There are therefore quite a few good reasons to believe that positive, not nega-
tive, correlation is more likely.
Nevertheless, in more cognitively complex settings it may be thought that
there are ways in which systematic negative correlations among people’s votes
might arise. People might harbour opposing ideologies or follow opposing
opinion leaders, for example.
In Chapter 11 we investigate the effects of opinion leaders more closely. To
foreshadow one core result: if two equally sized groups (which between them
account for substantially less than the total electorate) tend to follow two
diametrically opposed leaders, the votes of ‘followers’ will tend to neutralize
each other, leaving the decision to the remaining independent minds. Note
however that in this case—and in most other cases of opinion leadership (or
ideology, etc.)—the effect on the overall correlation among votes is actually
ambivalent. The voters following the same opinion leader are all pairwise posi-
tively correlated, while voters following different opinion leaders are negatively
correlated. To achieve negative pairwise correlation between all voters, a more
subtle coordination is required, akin to our example in Table 7.1.20 The next
section presents one proposal of how such coordination might work.
7.2.3 Epistemic Benefits of Diverse Cognitive

Models of the World
One reason for systematic negative correlation has been suggested by Scott
E. Page and Lu Hong.21 Stated very roughly, they claim that votes are typically
19 Also, Dietrich’s (2008) argument that it is usually impossible to justify the Competence and
Independence Assumption at the same time also applies to the justification of the assumption of
negative correlation.
20 Also, as Kaniovski (2010) shows, if we assume equal pairwise correlation, there are bounds
on the highest possible negative correlation, and these bounds tighten with increasing group size
and increasing individual competence.
21 Hong and Page 2009; 2012; see similarly Landemore and Page 2015. For a critique of attempts
to apply these findings to problems of ‘justice in a diverse society’ see Gaus (2016, ch. 3).
Diversity 105
negatively correlated because they are based on diverse cognitive models of

the world.22 If each voter conceptualizes the problem in a different way, then
negative correlation of their votes is a plausible (although far from certain)
implication.
For this argument to work, Hong and Page need to appeal to cognitively
richer environments. The voters do not simply ‘intuit’ which alternative to
support or have access to an immediate perceptual fact. Rather, they have
different cognitive representations of the problem and different underlying
reasons for voting one way or the other.
Let us return to the example from Table 7.1 and give the voters diversity in a
way that reflects Hong and Page’s approach. Assume that there are two alter-
natives, A and ¬A, such that exactly one of them is correct. In the classic
Condorcet Jury Theorem framework the voters have a signal that pertains to
the correct alternative overall. In Hong and Page’s framework, in contrast,
the voters have more partial information. Let there be three attributes to the
problem that determine whether A or ¬A is the correct choice and call these
attributes a1, a2, and a3. Each attribute ak can either be present (ak) or absent
(¬ak). Let A be the correct choice if at least two attributes are present and ¬A
if at least two are absent. One can think of these attributes as representing
reasons for one of the two alternatives, so that the correct choice is the one
supported by more r easons. The complete relation between attributes and cor-
rectness is set out in Table 7.2 in the first four columns. We assume that all
possible combinations of attributes are equally likely (and so, by implication,
is the truth of either A or ¬A).
The crucial element of the Hong and Page approach is the limited but differ-
ent ways in which different voters access the truth. In our simple example let
voter 1 observe only attribute 1, voter 2 attribute 2, and voter 3 attribute 3. (So in
this example, diversity among voters is well distributed, with equal numbers
observing different attributes: that, or something very much like that, matters
crucially to the Hong and Page result, too.) Each voter always votes for the
22 This is a not-uncommon thought among political philosophers. Arendt (1967/1977, p. 242)
for example writes:
In matters of opinion [about factual truth] . . ., our thinking is truly discursive, running, as it
were, from place to place, from one part of the world to another, through all kinds of conflicting
views, until it finally ascends from these particularities to some impartial generality. . . . [I]n
this process, . . . a particular issue is forced into the open that it may show itself from all sides,
in every possible perspective, until it is flooded and made transparent by the full light of
human comprehension. . . .
On the advantages of a multiperspectival approach, Bohman (2006, pp. 179–80) says similarly:
The point of inclusion . . . is not to find the right perspective but to have such perspectives
interact and inform each other, and in that way open up deliberation . . . to correction.
Various experiments have been constructed to show how subjects find solutions to problems
through novel information that is accessible only through the uptake of the perspectives
of others.
Table 7.2 Voters with partial information.

Attributes Correct Profile Probability
alternative
a1 a2 a3 A 1 1 1 0.25
¬a1 ¬a2 ¬a3 ¬A
a1 a2 ¬a3 A 1 1 0 0.25
¬a1 ¬a2 a3 ¬A
a1 ¬a2 a3 A 1 0 1 0.25
¬a1 a2 ¬a3 ¬A
¬a1 a2 a3 A 0 1 1 0.25
a1 ¬a2 ¬a3 ¬A
alternative supported by the attribute she can observe.23 This results in the vot-
ing profiles shown in Table 7.2. In terms of voting profiles and probabilities, the
result is identical to the negatively correlated votes (column I) from Table 7.1.
Hong and Page’s approach provides a rationale for votes being negatively
correlated. How plausible one deems this rationale to be depends on the nature
of the problem and the voters involved. If the question to be answered pertains
to a very immediate fact that does not lend itself to a disaggregation into under-
lying attributes, then the Hong and Page approach is inapplicable. This is most
obviously the case when, as we have said in Section 7.2.2, the fact to be deter-
mined is a basic perceptual fact.
Hong and Page are right to point out, however, that in many situations indi-
vidual voters do use different underlying models of reasoning to answer even
simple yes/no questions. The voters might employ different heuristics, they
might be focused on different attributes of the available evidence, or they might
approach the problem from different theoretical backgrounds. If these differ-
ent ‘interpretations’ present in the population are sufficiently diverse and no
particular interpretation is substantially more common than any of the others
across the voting population as a whole,24 then the resulting votes will be nega-
tively correlated and the epistemic performance of the group increases.
The demandingness of these assumptions should not be underestimated,
however. If the population is not diverse enough, or if the diversity is not evenly
spread across the voting population, the epistemic performance can easily get
23 Note that, unlike the standard jury theorem approach, there is no probabilistic element of
competence involved here. On their respective attributes the voters are 100% competent.
24 In Section 8.3 we will suggest a way of ‘dividing epistemic labour’ by giving groups of people
with particular expertise over some dimension of the problem power to veto options that they are
confident are wrong on the dimension that they are expert upon. In that setup there is no need for
the groups to be of similar size, as there is in the Hong–Page setup; but there is of course a need to
identify in advance what dimension (if any) each voter is expert upon, which is more demanding
than the Hong–Page setup in another way.
Diversity 107
worse—and each of those seems to us quite likely.25 In our toy example, if the
voters focus, for instance, only on either a1 or a2 but no one focuses on a3, the
majority vote will do much worse than in the ideal scenario described above, as
they ignore one important aspect of the evidence entirely. In addition, it may be
hard to know in advance whether a population is diverse in the specific way
required to make them collectively more competent. Thus, the general applic
ability of Hong and Page’s diversity framework with regard to voting is open to
serious questions.
7.3 ENGINEERING DIVERSIT Y
Given that diversity can have epistemic advantages, what can we do to promote
more of it?
In one sense, the answer is easy: include more different people among the
electorate, and encourage the honest expression of genuinely held dissenting
opinions.26 But remember, the goal here in view is not diversity for the sake of
diversity; instead, it is diversity for epistemic advantage.
Insofar as we are concerned purely with the epistemic advantages of diver-
sity in the mechanical process of aggregating votes, there is nothing to be
gained from having artificially dissenting voices. What we want is something
more complex: voters who tend to get it right more often, exactly when others
get it wrong. So the ideal negatively correlated voter is not a stubborn contrar-
ian. Rather, a valuable addition to the group is the negatively correlated voter
who tends to make different errors than his companions.
There may be other good epistemic reasons for institutionalizing (as does
the Roman Catholic Church in considering candidates for sainthood) the role
of a devil’s advocate, who deliberately argues against the prevailing view simply
because it is the prevailing one, even if he does not himself believe what he is
arguing. That might well be useful in getting neglected items onto the agenda
(Section 8.1); it may well be useful in slowing down a rush to judgement
(Sections 10.4.2 and 17.4.3). But in the mechanical processes of aggregating
votes that have been the concern of this chapter, having someone who votes ‘no’
simply because someone else voted ‘yes’—someone whose vote does not track
the truth, or even the evidence about the truth, but merely (negatively) tracks
someone else’s vote—adds no information to the decision process.
25 Thus, unequivocal enthusiasm for the Hong–Page result seems to us unwarranted
(cf. Landemore 2012; 2013b).
26 Just as management scientists recommend diversity along various dimensions in work
groups (Knippenberg and Schippers 2007). Banks perform better under ‘conditions of significant
uncertainty’, for example, when their boards of directors do not have too many members who
have the same domain-specific expertise (Almandoz and Tilcsik 2016).
Consider, in this context, the workings of parliamentary democracy.

Opposition parties often take the view that ‘it is the Opposition’s job to oppose’,
and hence react with knee-jerk negativity to any proposal coming from the
Government. Again, there may be good democratic warrant for that: maybe
that is democratically desirable, because it gives voters a clearer choice of
competing agendas come the next election.27 But again, say what you may
about the epistemic advantages of dissenting voices and negatively correlated
votes, there is no epistemic benefit to be had from someone saying the opposite
of someone else, not because he genuinely thinks the other wrong, but merely
because he thinks it is his job to say the opposite.
There is, however, definitely a case to be made for devil’s advocates and
dissenters, insofar as having someone enunciating those opinions (even if they
themselves do not believe them) encourages others who genuinely hold those
opinions to voice them—and to vote for them. In this way devil’s advocates help
avoid groupthink. In Asch’s famous conformity experiments (‘which line is
longest?’), if there was just one other person in the group who responded truth-
fully, experimental subjects would voice their true opinion rather than falling
in with the falsehood that all the other confederates in the experiment were
enunciating.28 Encouraging others who genuinely share that view to speak, and
to vote, the way they genuinely believe to be correct is one of the principal
epistemic reasons ‘why societies need dissent’29—it is needed to break the grip
of conformity on the voters, knocking out one important reason for positive
correlation among votes. That is one of the strongest epistemic arguments for
regimes of free speech.30
To engineer diverse groups, however, we need to know a lot about the poten-
tial voters among whom social engineers can choose. We need to know the
different cognitive models they use to make up their minds and ensure that no
specific model dominates the electorate.31 We need to know that these models
are different enough to cause negative rather than positive correlation; if they
are not different enough, these models might end up being common causes
introducing positive correlation. Finally, we also need to know that the voters
are at least somewhat competent in the application of their own model (even
though, as we have seen, they do not necessarily need to be competent on the
question to be decided for the benefits from diversity to arise). It does not
suffice to create an electorate that is diverse in some respect—it needs to be
diverse with respect to the problem at hand. Engineering such an electorate is
asking a lot from the social engineer. In fact, it might be asking for so much
27 Schumpeter 1950. Dahl 1972. 28 Asch 1955.

29 Sunstein 2003; Sunstein and Hastie 2014. 30 Ladha 1992.
31 Notice that simply increasing the number of voters, which is (other things being equal)
always epistemically advantageous within the classic CJT, might not be within these sorts of
models. It would not be, if the increase unbalances the electorate and causes one model to have a
disproportionate number of backers.
Diversity 109
that, if the engineer had all that knowledge, one may wonder why she could
not accurately predict the answer to the question they are voting on in the
first place.
A final approach to ‘engineering diversity’ would not strive to increase
the amount of diversity present in the decisional group but would instead
empower such diversity as is already present. The Hong–Page diversity setup
depends crucially on each different ‘cognitive model’ being roughly equally
common among the voting population. If that condition is met, then the
group’s decision will benefit from all those diverse models being brought to
bear on the complex task before it; but if not, it may well not. We could artifi-
cially equalize voting power among all the ‘cognitive models’ by identifying
(if we could) which voters harboured each model, and giving those voters a
veto over that dimension of the problem.32 Think of it as an epistemic equivalent
of Calhoun’s ‘concurrent majority’.33 We will discuss a version of that approach
in Section 8.3. But again, knowing even this much about voters is asking an
awful lot of social engineers.
32 But note that the strategy cannot give them a veto over the decision as a whole, since that
would almost necessarily lead to gridlock—certainly in the examples we discussed in this chapter,
and quite often in the real world as well.
33 Calhoun 1853/1992.
Division of Epistemic Labour
Human rationality is notoriously limited. People have limited time and attention
and other resources for making good decisions.1 They rationally allocate those
scarce decisional resources in ways that they expect will be most useful from their
own points of view. And, as discussed in Section 6.1.1, that ordinarily means that
they will not allocate those scarce decisional resources maximally to political
choices before them. Condorcet himself was acutely aware of this fact.2
A natural solution to problems of people being overwhelmed epistemically
is for them to engage in some form of ‘division of epistemic labour’.3 One way
to do that is to send out ‘search parties’ in search of good options, and let them
report back to the group as a whole as to what they found. We discuss this
approach in Section 8.1.
Another approach is to divide up the epistemic task, so that voters are
deciding among fewer options. One version of this strategy would be to break
up the group, assign different people different parts of the problem, and then
let the group as a whole choose among the best options each subgroup throws
up (Section 8.2.2). Another version of this strategy would be to break up the
problem, letting the group as a whole vote on options a few at a time over a long
sequence of votes (Section 8.2.3). Given certain crucial assumptions about
separability of different parts of the decision problem, both of these strategies
1 Simon 1982.
2 And he explicitly acknowledged the need to reduce the number of options available for
choice, accordingly. See Condorcet 1789/1994, p. 172; 1793/1994, pp. 190 ff.
3 Brennan 2010. Warren and Gastil (2015) recommend deliberative ‘minipublics’, such as
the British Columbia Citizen’s Assembly that deliberated on electoral reform proposals, on
these grounds. For experimental evidence that division of labour within an organization can
produce better results than either hierarchical or committee-style decision-making, see Becker
and Baloff (1969).
Notice, however, that simply splitting the group into smaller groups, and taking a majority
vote within each of those groups and then a majority vote among those groups, always leads to
epistemically worse overall decisions, so long as individual competence and the agenda con-
sidered by each group remains the same (Boland 1989, pp. 186–8). Thus, the epistemic advantages
of splitting the group that we will be discussing in this chapter all turn crucially on one or both of
those parameters changing as a result of splitting the group.
Division of Epistemic Labour 111
can produce epistemically better outcomes than letting the group as a whole
vote on all options at once. But they do so only where the number of options is
modest or the number of voters is large.
A final alternative is to devolve power to subgroups, giving each the authority
to make decisions for the group on the portion of the decision problem assigned
to it, either by choosing what they think is right or vetoing what they think is
clearly wrong within their sphere of expertise (Section 8.3). This approach can
work, given once again certain crucial assumptions about separability of different
parts of the decision problem and the identifiability of special area-specific
competence among voters. But if the general population is numerous and even
minimally better than random in competence, the epistemic advantages of
such an approach vanish.
8.1 LO CALIZED SEARCH
When the agenda is incomplete and the best options are missing, the voters
cannot vote for them. Getting the agenda right is the first epistemic problem,
even prior to the epistemic performance of the voting process.4
Much of the rich literature on learning within organizations concerns the
optimal mix of ‘exploitation’ and ‘exploration’—taking advantage of the best
option you currently have, as against searching for better options.5 The latter is
ordinarily best accomplished by having several teams working in parallel to
develop new alternatives. That latter strategy is ubiquitous, found everywhere
from beehives through scientific laboratories to high courts. Swarms of bees
identify the best site for a new hive through scouts going out to search in vari-
ous different directions, and reporting back to the hive through their famous
‘waggle dance’.6 In the socially efficient (and professionally prudent) division of
scientific labour, different labs are permitted and indeed encouraged to explore
to the fullest the potential of different plausible conjectures.7 The US Supreme
Court identifies cases worthy of its attention by, in the first instance, the justices’
clerks sifting through piles of applications for certiorari (judicial review) and
4 As we have argued in Section 4.2.2. See also: Goodin 2009, pp. 122–4; and Lippert-Rasmussen
(2012), arguing that democracy might be better than epistocracy if the set of alternatives before
the former is larger than before the latter.
5 March (1991). 6 See Seeley (1996) and List et al. (2009).
7 See: Latour and Woolgar (1979); Kitcher (1990); Weisberg and Muldoon (2009); Zollman
(2010). On the social efficiency of mechanisms to encourage more exploration see Kremer,
Mansour and Perry (2014). Wittman (1995, ch. 5) argues that what psychologists call the ‘overcon-
fidence bias’ is, in fact, only rational—you shouldn’t give up your previously held belief at the very
first indication to the contrary.
bringing interesting cases to the attention of their justices.8 Here we discuss

analogous ways in which a diversified set of political search parties might
improve the group’s epistemic performance by identifying better items for the
decision agenda.9
8.1.1 Incomplete Agendas
The agenda on offer can be context-dependent and therefore restricted. Political

processes are often like that—the system might be in a certain state, and this
state limits which other options are available. Context-dependence can create
epistemic problems because it can block good alternatives from making it onto
the agenda, as we will show now.
We discuss two ways how context-dependent agenda limitations can come
about. First, the political actors might be the victims of political myopia,
making them unable to see alternatives that are much better but also very
different from the current state. Call this the ‘myopia-induced context-
dependence’. It arises when the perception of what is possible is restricted
by what is currently the case. Second, the alternatives on the table may be
context-dependent because the political transition costs to better but very
different alternatives are too high. Call this the ‘inertia-induced context-
dependence’. We will focus on myopia-induced context-dependence for now
and return to inertia later.
To model the context-dependence of alternatives, it is useful to imagine the
policy options available as a ‘rugged policy landscape’, inspired by so-called
rugged fitness landscapes in evolutionary biology.10 We can plot this in a figure,
restricting ourselves to a simple one-dimensional landscape, but higher dimen-
sions would also be possible (and perhaps more realistic). The horizontal axis
displays the possible policy alternatives such that the distance between any two
alternatives represents how different the two alternatives are. The vertical axis
measures the goodness of the policy on offer.11 Each point on the curve is a
possible policy and its goodness, and the nearby points on the curve are feasible
alternatives and the goodness of each. The political actors will vote to leave the
current policy point if there is a better policy point that they can ‘see’ from their
8 List and Vermeule (2014).

9 Our discussion here is similar in spirit to (but differs in setup from) the models discussed
by Hong and Page (2004) and Page (2007).
10 Weisberg and Muldoon (2009) apply such models in similar ways to the division of scientific
labour. For an application of such models to problems of justice, see Gaus (2016, ch. 2).
11 We assume throughout this section that all political actors agree about the goodness of the
alternatives. Of interest here is the epistemic problem of finding the best alternative, not resolving
disagreement over which is best.
Basin of attraction
for A
Figure 8.1 Rugged policy landscape.
own vantage point.12 This transition process continues until a vantage point is
reached from which there is no better policy alternative visible.
The policy landscape of Figure 8.1 provides an example. Here we assume that
the political actors can see all policy options that are no further than 10 per cent
of the axis length away to the left and to the right. We call the points the policy-
maker can see from the status quo (the current policy point) the ‘feasible set’
relative to the status quo. The set of elements in the feasible set of the status quo
that beat the status quo is the ‘winset’ of the status quo. If the winset is empty (that
is to say: there is no alternative in the feasible set that beats the status quo) the
status quo is an equilibrium. In the example provided in Figure 8.1 there are three
equilibrium points: A, B, and C. The horizontal lines through the points indicate
the intervals of the feasible sets. In each case there are no better policy points in
the feasible set, as the policy curve remains below these lines. Beneath the fitness
landscape, at the bottom of the figure, is a transition diagram. The arrows give an
indication in which direction a transition through the policy landscape would
proceed until one of the equilibria A, B, or C is reached. To obtain the transition
diagram, for each policy point we determine whether the highest visible point
(given the political myopia under which the agents operate) lies to the left or the
12 In this model, there is assumed to be no communication among agents, and they only
move from their current position if they can themselves see a better one from their own present
vantage point.
right. The arrows in the transition diagram show these directions and therefore
indicate the direction of travel in a search process.
There is one obviously best equilibrium in Figure 8.1, namely, point A. However,
despite there being a best alternative, it is by no means guaranteed that this best
equilibrium will be reached. For example, let the political transition begin at a
point somewhere between B and C. Depending on where precisely the starting
point lies, the policymakers will either climb up to equilibrium B or equilib-
rium C, and then the political transition stops. It would obviously be better to
move towards the global maximum A, but since A is not visible from B or C the
policymakers do not even have A on their agenda. This implies, of course, that
the policymakers might easily get stuck in C, which is the worst equilibrium by
far. Such an outcome is clearly suboptimal, but that is precisely the problem
with myopia-induced context-dependence.
If all policymakers start at the same point on the political landscape, the risk
of ending up in a suboptimal equilibrium is quite high. One can consider the
‘basin of attraction’ for A—the starting points that eventually lead to a transi-
tion to A. In the example of Figure 8.1, the basin of attraction for A is the grey
area on the left. If the search for the best alternative begins in this area, A will
be reached eventually. The basin for A is roughly one-third of the whole policy
space. If we assume that each starting point is equally likely, the best equilibrium
will never make it onto the agenda in two-thirds of cases.
8.1.2 Diversified (or Many Random) Search

Parties as a Solution
The problem of myopia-induced context-dependence can be solved if the

search for the best equilibrium is diversified. Suppose that, instead of everyone
starting at one and the same point in the policy landscape, there are several
‘search parties’ starting at different points. These different parties then ‘report
back’ to the group as a whole, informing them of the equilibrium they found.
These reported options are added to everyone’s feasible set, so that a successful
search will include many (ideally, all) equilibria available.13
There are different ways of thinking about this in political terms. One could
interpret different political parties as ‘search parties’, starting at different points
of the ideological spectrum to develop their best policy proposals.14 Or one
13 And of course the more such search parties that are sent out, the more likely it is that one will
come across the correct option and report it back to the group (Thompson 2013).
14 As discussed in Section 18.1.1. A potential problem with this analogy is that while parties can
develop policy proposals, they cannot typically test them and show their superiority until they are
in government and implement them. If not in government, parties can only add theoretical
options. That may not be enough to convince voters that these theoretical options are indeed
feasible. One could interpret this as an epistemic reason why it is desirable that different parties
are in government at different times—it allows parties to show that their policy proposals are
implementable and lead to better outcomes than other policies.
could argue that a federalist political system provides different ‘search parties’
to find the best equilibria. This was famously the view of Justice Louis Brandeis
in his praise of federalism:
It is one of the happy incidents of the federal system that a single courageous state
may, if its citizens choose, serve as a laboratory; and try novel social and economic
experiments without risk to the rest of the country.15
If, for example, fifty states were to try fifty different versions of health insurance
and each state were to start at a different place in the political system, then after
a couple of years fifty different practice-tested ways of running a health insur-
ance scheme would be available to compare.16
How many search parties do we need? That depends, of course, on the search
radius of a single search party, which in turn influences how large the basin of
attraction for the global optimum is. It also depends on where the different
search parties start their search, specifically on whether their starting positions
are sufficiently diverse.
As we have seen, in the setting related to Figure 8.1, the search will reach the
global maximum as soon as one party starts to search within the basin of attrac-
tion indicated in grey. If one were to space out the starting points evenly, one
would need at least three search parties (one at 0.25, one at 0.5, one at 0.75) to
succeed.17 Notice, however, that if one were to assign starting points randomly,
the chance of missing the basin of attraction is roughly ( 2 / 3 ) , where x is the
x
number of search parties. The chance of failure rapidly converges to zero with
increasing numbers of search parties. With five parties it is about 13 per cent,
for example, but with ten it is 1.7 per cent.
Suppose there are 1,000 different patches (intervals), only one of which contains
the global maximum. Suppose search parties get assigned randomly each to search
one patch. How successful the search parties will be depends on the number of
search parties there are. If there is only one search party the probability of success
is 1/1,000. But x search parties will find the global maximum with probability
1 − (999/1,000)x.
Figure 8.2 displays the probability of success based on the number of search
parties being assigned a random search patch, on that scenario. The upshot is
that there is no need to enforce diversity in starting positions, insofar as we can
rely instead on randomness and large enough numbers.18 Of course it would be
even more efficient to have each of 1,000 search parties be assigned to look at
15 Brandeis 1932. Allard-Tremblay 2017, p. 6.

16 Of course, if local circumstances varied in such a way as to make the right choice of health
insurance system for one state the wrong choice for another state, then the results of the one state’s
search would have no bearing on the other state’s search.
17 Two search parties put at position 1/3 and 2/3 would narrowly miss the mark: the threshold
of the basin of attraction is just below 1/3 in the example.
18 Note that Hong and Page (2004, pp. 16, 387) do not impose diversity but instead achieve it
through randomization. Cf. Thompson (2014, p. 1028) and Kuehn (2017, pp. 80–2).

1.0
Probability of finding global maximum
0.8
0.6
0.4
0.2
0.0
0 1,000 2,000 3,000 4,000 5,000
Number of search parties
Figure 8.2 Search on 1,000 patches, probability of finding correct patch depending on
numbers of search parties.
one or another of the 1,000 different patches. But if enforcing maximum search
diversity is infeasible, randomness coupled with larger numbers of search parties
will yield nearly as good results. In our toy model in Figure 8.2, 5,000 search
parties with random starting points are almost as good.
8.1.3 Transition Costs
This diversified search can lead to much better outcomes. But that will happen
only if the best option identified by some other search party (in the federalism
example, some other state) is added into your own feasible set of policy
options.19 That may not necessarily happen, insofar as transition costs intro-
duce policy inertia.
In cases of myopia-induced context-dependence, the distance in the political
landscape determines which alternatives are ‘on the radar’ of the policymakers.
This limited visibility of options is an epistemic problem. In case of inertia-induced
context-dependence, by contrast, the accessible alternatives from the current
state are restricted not because they are invisible but because the costs of the
transition can be prohibitively expensive. This is (primarily) a political, not an
epistemic, problem. However, there may be an epistemic element to the solu-
tion: once the political actors know that a much better alternative is available,
19 Walker 1969.
status quo bias may be less strong and the transition to that alternative may
meet with less resistance.20
Political inertia may be due to the slowness of the legislative process
combined with the need for constant political support throughout the transition
process. Imagine, for example, a reformist politician considering replacing
a working but inefficient suboptimal health insurance system with a better
system. In order to do that, many legislative and executive changes need to be
implemented; there will be delays in setting up the new system; there will
be winners and losers; there will be rent-seeking and organized resistance, and
so on. All in all, the costs can be daunting for the reformist politician and the
whole political community. Unless it is possible to reach a clearly better out-
come within a reasonably short period of time and without too many sacrifices
in the transition period, the reformer is likely to refrain from embarking on
that journey if he wants to keep his post. In fact, a political community may be
‘trapped’ in a suboptimal equilibrium even if all relevant actors realize that the
policy change would be in everyone’s interest in the long term. In those cases,
good or best solutions do not make it onto the political agenda because politi-
cians are aware that attempting this transition is likely to lead to failure in the
short term.
In the case of inertia-induced context-dependence the path on the curve
between the current state and the potential future state indicates the costs of the
transition. The longer and deeper the path is below the goodness of the status
quo, the more difficult the transition is. In other words: the longer and more
severe the sacrifices one needs to make to get from one equilibrium to the other,
the higher the transition costs. In political terms, a reformist politician has to
decide whether she has enough ‘political capital’ to embark on the journey
through the trough of the transition and make it to the other side before the
political capital is spent.
One simple way to model this is to consider the area below the goodness
level of the status quo and the curve representing policy options. Figure 8.3
is an example. As in Figure 8.1, we see a rugged policy landscape. However,
this time the feasible set for a status quo is determined not by myopia but by
how far the political actor can move from the status quo with the political
capital available. Each step away from the status quo down the policy curve
reduces the political capital budget. In Figure 8.3 the political capital of an
agent in, for instance, equilibrium B is represented by the grey area between
the constant y-value of B and the policy curve. Suppose a political actor in B
moves to the right in order to reach point C. Each point below the goodness
value of the status quo costs political capital. This starts adding up in form of
the grey area between B and the new position to the right of B. Here the
political capital is insufficient to reach a better point: it is exhausted where
20 Walker 1969, p. 890. Harsanyi 1969, pp. 528–9.

C
A B
Figure 8.3 Rugged policy landscape and inertia-induced equilibria.
the grey-shaded area ends. Since this is true both to the left and to the right
of B, B is an equilibrium.
As before, there are several suboptimal equilibria (A, B, and C). D is the best
policy point to reach, but it is cut off by a deep political trough and therefore
difficult to get to. In the first instance, this is not an epistemic problem: it is
perfectly consistent with this model that the political actors know of the better
equilibria—they are out of reach for reasons of political feasibility, not for
epistemic reasons. However, there might still be an indirect epistemic aspect to
the problem of inertia-induced context-dependence: if it is demonstrated to the
political actors stuck in low equilibria that a high equilibrium is available and
has been tried and tested, then the situation might change in two ways.21 First,
the transition process might become faster, reducing the transition costs. In
terms of the model this would mean compressing the policy curve on the
dimension of the horizontal axis, so that the actors’ political capital can reach
further. Second, knowing about a preferable high equilibrium might make the
intermediate steps to reach it less costly. This could be modelled as a lift of the
policy curve, especially in the regions of deep troughs.
Whichever way one chooses to model this idea, it is clear that the knowledge
of a proven attractive policy option changes the political landscape and makes
transitions in that direction easier. Yet again it is plausible that the ‘laboratory
of federalism’ can help to create, investigate, and test policy alternatives that
21 See again Walker 1969; Harsanyi 1969.

will eventually render the transition to the best alternative easier. For example,
once it is known that a preferable healthcare system exists in one state (say,
Massachusetts), the costs for other states of moving towards that high equilib-
rium may be lowered because the political resistance is weakened if the relevant
alternative has been tried and tested elsewhere. The transition period may also
be faster because one can copy a lot of legislation from the successful example
rather than working towards a new system in a piecemeal fashion.
The important upshot of this analysis is simply this: context-dependence can
lead to suboptimal results because the agenda does not contain the best options
genuinely available. This problem can be overcome with a two-step procedure:
first, a diversified search needs to be conducted; and second, the available alter-
natives thus unearthed need to be added to the group’s agenda. This approach
increases the probability that the agenda will contain the best policy option
available.22 It does not, of course, guarantee that the best item on the agenda
will be chosen. How epistemically successful the political system is in choosing
the best alternative on the agenda still depends on the usual factors: voter com-
petence, group size, and independence. But without putting the best alternative
on the agenda in the first place, it can never be chosen.
8.1.4 Recognizing the Best When We See It
There is an assumption underlying all these search models that has so far
remained largely implicit. That is that when we come across a better option we
will recognize it as such—either with certainty or with a probability not too far
off certainty.23 That assumption is at work at two points in our search models.
We assume that the search parties will recognize better options when they
come across them in the course of their search; and we assume that the group
as a whole will recognize which search party has found the best option, when
they all report back.24 In the medical insurance example, we assume that each
state finds the best policy available to it within its constraints; and we assume
that when the federal government chooses which state model to adopt as
national policy, it will choose the best among those models that are available to
it (within the constraints under which it operates).
22 Or even if not the best, better ones than before, thus giving the group a chance to choose a
‘more nearly correct’ alternative than it would otherwise have been able to do.
23 There is a parallel assumption in the work of Page (2007, pp. xxix, 160), which—in
Landemore’s (2013b, p. 102) gloss—requires that ‘the best solution must be obvious to all of [the
participants] when they are made to think of it’. As Ancell (2017, pp. 166–9) argues, that may well
not be the case.
24 Thompson’s (2013) model retains the assumption that individuals are more likely to vote
for the correct option than any incorrect one, once the correct item appears on their agenda for
choice, even if each is unlikely to recognize it ‘in the wild’ as an option that merits being put on
the agenda.
Sometimes such assumptions are well warranted. There are things we are
searching for—a lost child whom we know well, for example—that are such that
‘we know them when we see them’.25 And solutions to some policy problems
are like that; obvious to everyone once they’re mentioned, but not obvious until
then. But sometimes we are searching for a lost child based on some relatively
vague description that any number of children might meet. Sometimes when
prospectors are searching for gold they strike upon iron pyrite, ‘fool’s gold’. And
a lot of policy problems are like that, as well.
If making choices were costless and the assumption that we ‘know the best
when we see it’ unequivocally true, then it would follow that it’s always better
adding extra options to the agenda for choice. It could never make things worse,
and it might improve things. Of course, adding all relevant alternatives might
not be easy: as we have seen in Section 8.1, the best alternatives are sometimes
invisible due to political myopia. And even if we could add all alternatives,
there may be good epistemic reasons for not overburdening the agenda.26 One
reason is that there are decision costs, in terms of time and attention, that are
higher when you have to sort through more options. Another reason is that
your competence—the probability of your making the right choice—might
decrease with more options on the table. We will discuss models built around
that assumption in Section 8.2.
The main thing to notice, for now, is just this. Insofar as voters are anything
less than perfect at ‘recognizing the correct option when it is presented to them’,
the probability that they will make the correct choice is a joint function of the
probability that the correct choice will be on the agenda and the probability
that they will choose that option if it is on the agenda. Yet getting the correct
option onto the agenda is only half of the game, as we shall next show.
8.2 NARROWING THE FO CUS
Crucial to the success of the ‘local search party’ strategy discussed in the previous
sections was that different search parties went to search different locales,
either by design or by chance. Each concentrated its attention on some subset
of options, rather than on all options at once. Given limits on each search party’s
visible or achievable alternatives, deliberately organizing the search in that way
maximized the chances for success by some party or another, and hence the
group overall.
25 So are correct solutions to mathematical puzzles, at least for trained mathematicians; and
social psychological experiments confirm that groups are better than isolated individuals at
finding correct solutions (Laughlin and Ellis 1986).
26 Dworkin 1988.
Next let us explore a different reason for dividing epistemic labour: the
assumption that each voter’s individual competence is higher the fewer options
he has to consider. If that is so, then dividing the decision task up and consid-
ering fewer options at a time would typically lead each individual to choose
more accurately.
8.2.1 Letting Individual Competence Vary

with the Number of Options
We have seen in Section 3.2 that the Condorcet Jury Theorem can be extended
to more than two alternatives and plurality voting. But as we also speculated
there, more items on the agenda may well dilute individual voter competence.27
In that passage in Chapter 3 we already discussed one simple way of modelling
the dilution of competence: with each additional alternative the probabilities
would be distributed over more options, so that a vote for the correct option
becomes less likely overall.28 However, our approach there preserved (for sim-
plicity) a fixed competence advantage—the correct alternative would always be
voted for with the same probability margin above the probability of a random
choice (1 percentage point above random, in Figure 3.2). Now let us ask what
happens if that margin itself drops as the number of alternatives increases.
One reason why a glut of alternatives can reduce competence is that if
there are more incorrect options voters may be more likely to confuse them
with the correct option. Everybody will recognize this effect from their own
consumer choices: deciding between two ice cream flavours is easy, but
deciding between thirty is hard. A similar effect might kick in for epistemic
judgements: the correct option may be easy to spot out of two, but hard to
spot out of twenty alternatives.
Several mechanisms might play a role here. First, if the voter has limited time
or attention, less time will be spent examining each alternative when there are
many of them. Second, if there are many options, it becomes more likely that
there are some incorrect options available that are quite similar to the correct
option, increasing the risk of further confusion. Third, even if the voter has
sufficient time, weighing all the pros and cons of many options may be compu-
tationally very demanding.
27 For evidence of that see Lau et al. (2014, p. 254), whose examination of people’s propensity
to ‘vote correctly’ (i.e. in line with their own expressed policy preferences) in sixty-nine elections
across thirty-three countries find that that is strongly affected by the number of alternatives on the
ballot. ‘When there are only two alternatives on the ballot, all else equal the model predicts almost
79 per cent correct voting, but when there are nine alternatives competing in an election, the
probability of a correct vote drops precipitously to under 57 per cent.’
28 At least for small to moderate group sizes; see Figure 3.2 in Chapter 3 for the ambivalent
effects when alternatives are many and the group is large.
How precisely the number of available alternatives impacts voter compe-

tence is ultimately an empirical question, of course. Here we employ one simple
functional relation to produce a numerical example. Suppose that a group of
voters presented with k alternatives votes for the correct alternative with a
probability that is 10%/k higher than the probability of a random choice being
correct, and with equal probability for all other alternatives. Thus, the voters
vote with probability 1/k + 10%/k for the correct, and with probability 1/k –
(10%/k)/(k – 1) for each of the incorrect alternatives. This leads to competence
distributions over the correct and incorrect alternatives as stated in the first
three columns of Table 8.1.
While the margin is fully 10 per cent in the two-option case, it shrinks as
more options are added and approaches 0 for very many options. The reduc-
tion in the margin has a toll on the epistemic quality of the group decisions for
smaller group sizes. It remains the case of course that, among any appreciable
number of voters, the correct alternative is likely to be the plurality winner,
getting more votes than any of the incorrect alternatives, as argued in Section 3.2.
Be those relativities as they may, however, the absolute probability of the plur-
ality winner being correct can decline dramatically with increasing numbers of
alternatives.
The other columns in Table 8.1 show results for these competence distribu-
tions and varying numbers of voters, n. There we observe the familiar pattern
of larger groups performing better than smaller groups, epistemically, for any
given number of options k. However, since on this model individual compe-
tence decreases with the number of alternatives on the agenda, so does the
group competence.
For quite large populations of, say, 1,002 voters, the loss of epistemic perform-
ance is initially moderate (as, e.g. when we increase the number of alternatives
from two to three). But group competence then begins to fall quite swiftly,
Table 8.1 Group competence as a function of declining individual competence with

many alternatives.
k= Probability of each voter voting for the: n=
Correct alternative Incorrect alternatives 12 101 303 501 1,002
2 0.55 0.45 0.64 0.84 0.96 0.99 1.00

3 0.3666 2 × 0.3166 0.42 0.59 0.77 0.85 0.95
4 0.275 3 × 0.2416 0.31 0.45 0.60 0.69 0.84
5 0.22 4 × 0.195 0.25 0.35 0.48 0.57 0.72
6 0.1833 5 × 0.1633 0.21 0.29 0.40 0.48 0.62
7 0.1571 6 × 0.1405 0.18 0.25 0.34 0.41 0.54
8 0.1375 7 × 0.1232 0.15 0.21 0.29 0.36 0.47
12 0.0917 11 × 0.0826 0.10 0.13 0.18 0.22 0.31
16 0.0688 15 × 0.0621 0.08 0.10 0.13 0.16 0.22
20 0.055 19 × 0.0497 0.06 0.08 0.10 0.12 0.17
dropping below half with k = 8 alternatives. For smaller (e.g. parliament- or

committee-sized) groups of voters, the rapid deterioration of group compe-
tence is even more pronounced. Take n = 101 for example: group competence is
0.84 for k = 2 but only 0.21 for k = 8.
The upshot is: if putting more alternatives on the agenda exacts an ‘epistemic
price’ in terms of lower individual competence, the group competence tends to
go down as well. Of course, this is just one numerical example among the many
possible. Still, it is suggestive in showing how, if more alternatives lead to
decreased competence, group competence can suffer.
8.2.2 Considering Options a Few at a Time
Given these results, it might be advantageous to divide epistemic labour in either

of two ways. One is by splitting the decision-making group, letting subgroups
first vote on a smaller set of options, and then the whole group reconvene to
decide among the options recommended by each subgroup out of its assigned
subset of options. That strategy will be discussed in Section 8.2.3.
Alternatively, we might split the agenda rather than splitting the group.29 In
this model, the group as a whole always remains in plenum. But instead of
voting on all options at the same time (as in the model in Section 8.2.1), the
group undertakes a series of what we will call ‘subvotes’ pitting just a few of
those options against one another each time.30
Imagine a legislature composed of n = 303 members facing a choice among
k = 6 options. From Table 8.1 we know that the probability of them choosing the
correct option if voting on all of them at once is 0.40. But suppose that, instead
of doing that, the whole legislature of n = 303 members considers the options
in pairwise fashion, the first subvote pitting x1 against x2, the next subvote
pitting the winner of that round against x3, and so on until all six options have
been exhausted.
We know from Table 8.1 that the probability of n = 303 voters being correct
in any given binary choice is 0.96. But the precise probability of the group as a
whole choosing the correct option through such a pairwise procedure depends
on when the correct option is entered into the sequence of votes. Suppose the
correct option is the very last option entered. Then the group only has to get the
29 What we describe here is, of course, just the procedure of Condorcet pairwise comparison,
introduced briefly in Section 3.3.2.
30 The question is, of course, whether one may assume that the individual competence remains
high throughout this series of votes. If the reason for reduced competence in many-alternative
settings is lack of time or attention, then a similar reduction in individual competence will bedevil
the route taken here. By contrast, if the reason for individual competence reduction lies in the
computational challenge of comparing many alternatives at once, then reshaping the decision into
a sequence of pairwise votes can be epistemically beneficial.
correct answer once, in the very last subvote; and the probability it will choose
the correct option through this pairwise procedure is therefore fully 0.96. In that
case the group would be vastly more likely to choose the correct option through
the pairwise procedure than by voting on all six options at once.
Suppose however that the correct option was, instead, one of the two options
considered in the very first subvote. Then the group as a whole has to make the
correct choice not only in that first round, but also in each of the four following
rounds, for the correct option to emerge as the overall winner out of this pair-
wise procedure. And the probability of that happening is 0.965 which is approxi-
mately 0.82. Still, even in that worst case, the group would be much more likely
to choose the correct option through the pairwise procedure than by voting on
all six options at once. And in other cases in which the best option is introduced
at some intermediate point in the voting sequence, the probability of the group
getting the correct answer is in between those two values—averaging across all
cases, group competence is around 0.90. Therefore, to get around the deteriorating
group competence for decisions with many alternatives, splitting the agenda is
epistemically a good idea.
Note that that remains so, even in cases with a larger number of options.
Imagine our 303-person group faces twice as many options as before, k = 12.
The probability that the group as a whole will be correct in choosing over all the
dozen options at once is 0.18. The probability of the correct option emerging
from the pairwise round-robin procedure just described can be no lower than
0.63. (It is that in the worst-case scenario, in which the correct option is among
the first pair considered and therefore has to withstand ten more challenges
successfully, with the probability 0.9611.)
8.2.3 Subgroups Propose, Whole Groups Dispose
Consider next the possibility of letting subgroups vote on subsets of options,

with the whole group then voting among the options emerging out of each
subgroup’s subvote.
Suppose, once again, there are k = 6 possible options altogether. On the model
now under discussion, we might ask a third of the voters to choose among
options {x1, x2}, another third among {x3, x4}, and the last third to choose
among options {x5, x6}. The three winners of these subgroup subvotes would
then be pitted against each other in a vote among the electorate as a whole,
reconvened in a plenary session.
If individual voter competence is higher the fewer options there are under
consideration, then each individual voter will be more likely to choose cor-
rectly among that more restricted set of options than he would have been had
he had to choose among all six options at once. And that would be true at
both stages of the decision procedure, since each subgroup considers only the
smaller set of options assigned to it and the plenary group considers only
the smaller set of options that subgroups recommend to it. That is the epistemic
upside of this strategy.31
There are epistemic downsides to this strategy, however. The first is that each
subgroup is by definition smaller than the whole group. We know from the CJT
that smaller groups are epistemically less reliable than larger groups, all else
being equal. Of course things are not quite equal in this case: individual voter
competence is, ex hypothesi, higher as a result of dividing up the decision task
in this way. Still, the point remains: the epistemic costs that come from redu-
cing the size of the group could outweigh the epistemic gains that come from
reducing the number of options each subgroup has to consider.
Second, of course, the group as a whole will choose among the different
options recommended by the subgroups, and the chance of its erring in that
second-stage choice will further erode the epistemic gains made from assign-
ing the first-stage choice to subgroups that were more focused and hence more
competent. That is the second epistemic downside to this strategy—it is a two-
step decision process, with a possibility of error at both steps. The probability of
the correct outcome emerging from this two-step process is the probability of the
correct outcome emerging from the subgroup on whose agenda it appeared
times the probability that the group as a whole will vote for the correct outcome
when considering all the subgroups’ recommendations. Given that each of
those probabilities is less than one, the product of multiplying them together is
smaller than either alone.32
For a numerical example, consider the same one as before. Suppose the
group of n = 303 members faces a choice among k = 6 options. From Table 8.1
we know that if the group as a whole votes directly on all six options, the prob-
ability of their selecting the correct option is 0.40. Now suppose the group as a
whole first splits itself into three 101-member subgroups, each charged with
considering just two of those options (with each option being considered by
some subgroup, and none by more than one). The probability of each subgroup
sized n = 101 being correct over two options is 0.84. But the group as a whole has
then to come back together to vote on the three options thrown up by each
31 Voters in one subgroup may know (or suspect) that the globally best alternative lies within
some other subgroup’s portfolio. And they may vote strategically for some worse option within
their own portfolio, in hopes of further ensuring that that globally best alternative will emerge
as the clear winner as a result of the two-stage decision process. That would constitute a violation
of the CJT’s Sincerity Assumption. But that will make no difference to the overall outcome, just so
long as voters in the subgroup whose portfolio actually contains the globally best alternative do
not themselves mistakenly engage in this behaviour.
32 Taking the product assumes independence of these votes, of course. Another epistemic
downside of splitting the decision into many subvotes is that the same mechanisms that reduce
competence in the case of many alternatives might also decrease competence in the case of many
subvotes. For example, the voters might invest less time in researching each subvote if they have
to partake in many of them.
subgroup's subvote. The probability of a group sized n = 303 being correct over
three options is 0.77. So, per the procedure outlined above, we calculate that the
probability of the correct outcome emerging from this two-stage decision
process is 0.84 × 0.77 = 0.65.
That is considerably higher than the probability of 0.40 that the 303-member
group as a whole would have chosen the correct outcome, voting on all six
options directly and without splitting itself into subgroups with smaller agen-
das.33 But while this ‘splitting the group’ strategy for dividing epistemic labour
is clearly still beneficial, it is in this example not nearly so beneficial as the
‘splitting the agenda’ strategy discussed in Section 8.2.2.
8.2.4 Experts Propose, Whole Groups Dispose
Most of the time, the sheer numbers of moderately competent voters make reliance
on expert judgement unnecessary. But not always. Here is one way in which
experts might play a distinct role in winnowing down the options available.
Voters in the general population are likely to see their competence dwindle
quickly as the number of options increases, as per our assumptions in Section 8.2.1.
But experts may not suffer a loss of competence, anyway not to anything like
the same extent. There are several plausible reasons for supposing that this
might be so. First, experts are trained to deal systematically with many options.
Second, as professionals devoted to that task, they often have more time to consider
all options. Third, they tend to be less likely to be confused by the introduction
of incorrect options that look superficially like the correct option. Fourth, their
professional training might make experts less likely than the public at large to
be misled by the apparent ‘salience’ of certain incorrect options.
Here is a small numerical example to show what might be the effects of that
differential loss in competence. Suppose the experts are always 10 percentage
points better than a random choice in picking out the correct option, regardless
of how many options there are on the table. As the number of options increases,
the probability that experts will vote for the correct option still decreases—but
not as fast as that of the electorate at large, whose advantage over a random
choice diminishes with more options in the way described in Section 8.2.1.
We can make use of this special competence of experts when confronted
with many options to winnow the agenda for the general electorate.34 Here is
the idea. First let the experts vote on all options available. Then let the voters at
large choose between the two options that have the most and the next most
33 If k grows larger and n is relatively modest, however, the advantages of this strategy
are reduced.
34 Jeffrey (2017) advocates a version of this sort of model as a means of reconciling ‘limited
epistocracy and political inclusion’. See also Moore 2017.
votes among the experts.35 Supposing the general population is large and
reasonably competent in choosing among just two options, the majority vote
among them will almost certainly select the correct option, so long as the
correct option is among the two selected by the experts.
How epistemically reliable is the general electorate’s vote when experts
winnow its agenda in this way, compared to when they do not? Table 8.2 pro-
vides some examples for three experts and 303 voters, with voter competence
in the general electorate declining as a function of the number of options in the
way described in Section 8.2.1.
It is striking that, as long as the number of options is small (k = 3, in Table 8.2,
for example), holding a direct vote among the voters is epistemically superior.
Not by huge margins, to be sure: the general electorate voting on just the two
top options recommended by the experts is almost as good. But since the number
of experts is small, they occasionally fail to put the correct option onto the
agenda for the general electorate’s vote.
This pattern reverses itself, however, as the number of options increases. In
this specific numerical example, the preselection of options by the experts
improves results when there are four or more options on the agenda. By narrow-
ing down the choices before using the ‘wisdom of crowds’ the experts play an
important role in the epistemic performance of the group.
Of course, everything here hinges on the assumption that the experts are sig-
nificantly more competent than the population when the agenda is crowded—
an assumption that, while plausible, may not necessarily hold. But much also
depends on just how many experts there are and just how much more resilient
Table 8.2 Voting on multiple alternatives, either with three experts selecting their
top two options and 303 voters voting in the run-off, or a direct plurality vote among
population with voter competence as in Table 8.1.
k= Probability of each expert voting Probability of general Probability of general
for the: electorate selecting correct electorate selecting
option with expert correct option in
Correct Incorrect pre-selection, n = 3 direct plurality vote,
alternative alternatives experts and n = 303 voters n = 303 voters
3 0.4333 2 × 0.2833 0.74 0.77

4 0.35 3 × 0.2167 0.61 0.60
5 0.3 4 × 0.175 0.53 0.48
6 0.2667 5 × 0.1467 0.47 0.40
7 0.2429 6 × 0.1262 0.43 0.34
8 0.225 7 × 0.1107 0.40 0.29
12 0.1833 11 × 0.0742 0.33 0.18
16 0.1625 15 × 0.0558 0.30 0.13
20 0.15 19 × 0.0447 0.27 0.10
35 If ties need to be broken to select exactly two options, this is done by random choice.
their competence is than that of the general electorate in the face of larger
numbers of options. Were there 200 experts rather than three, or were they not
merely 10 but 20 percentage points better than random at choosing among
any number of options, the values in the next-to-last column in Table 8.2 would
be much higher.
Notice that, at least on the particular numerical values in the example above,
relying on experts to winnow the agenda is an epistemically less successful
strategy for dividing epistemic labour than either splitting the decision-making
group (as per Section 8.2.3) or splitting the agenda (as per Section 8.2.2). The
probability of the electorate choosing the correct outcome by relying on expert
preselection is, in the six-option case and on the assumptions in Table 8.2,
only 0.47. The probability of the electorate choosing the correct outcome through
the other strategies is 0.65 and at least 0.82, respectively. Obviously, this result
reverses when asking more competent experts: fifty experts winnowing down
the number of options leads to the electorate choosing the correct outcome with
probability 0.85 in the six-option case, for example. This demonstrates that,
even when relying on experts to winnow down the agenda, relying on more
experts is often better.
8.3 DEVOLVING CONTROL OVER SOME DIMENSIONS
In the model in Section 8.2.2, individuals were made more competent by assign-

ing them to subgroups that had fewer options to consider. There, what they are
more competent over is which of the options before them is better, tout court.
Next let us suppose that some (independently identifiable) voters are for
some other, exogenously given reason more competent at assessing certain
dimensions of a decision. What would happen if we devolved (in one way or
another) the whole group’s decision on each dimension to the subgroup of its
members who are particularly competent over that dimension?36
This approach obviously presupposes that the problem to be solved can be
broken down into wholly separable questions, and that different people have
differential competence on each of those questions. Gains from epistemic spe-
cialization thus require questions whose answer is determined by dimensions
that can be analysed separately by those who are particularly competent on
those dimensions.37
36 In the spirit of Ober’s (2012) REA II and REA III proposals, although the details differ. See
also List (2005a) for the link between specialization and judgement aggregation.
37 It must also be possible to derive the unique correct answer from correct decisions on the
different dimension. In the language of social choice theory, the agenda needs to be ‘minimally
connected’ as defined by Dietrich and List (2007).
Here is one way in which particularly competent subgroups might be

empowered positively (albeit partially) to make choices on the whole group’s
behalf. Suppose the alternatives on offer have three possible attributes: p or
not-p, q or not-q, and r or not-r. Everyone knows which alternatives have which
attributes. What needs to be decided is simply how the attributes (and therefore
the alternatives) should be evaluated:
1. Is it better to have p or not-p?
2. Is it better to have q or not-q?
3. Is it better to have r or not-r?
We stipulate that the correct alternative is the one for which the answers are p,
q, and r (but this is not known to voters in general). Overall there are eight
alternatives, constituted of all possible combinations of answers to how to
evaluate dimensions p, q, and r.
Suppose the population consists of four groups. Ten per cent of the population
are p-competent voters, having a 10 per cent better-than-random competence to
pick an alternative that has the right property with regard to p. Another ten per
cent of the population are q-competent voters with a 10 per cent better-than-
random competence to pick an alternative that has the right property with regard
to q. And yet another ten per cent are r-competent voters who have a 10 per cent
better-than-random competence to pick an alternative that has the right property
with regard to r.
Suppose furthermore that if voters are undecided between any alternatives
then they randomize. For example, a p-competent voter has (by stipulation) a
probability of 0.6 to vote for an alternative that has property p. Since there are
four alternatives with property p, when such a voter randomizes he chooses
each one with probability 0.6/4 = 0.15. And he chooses all other alternatives
with probability 0.4/4 = 0.1.
Finally, suppose that all the other voters—the remaining 70 per cent of the
population—choose purely at random among the eight alternatives on offer,
which means that they choose each alternative with probability 1/8 = 0.125.
These assumptions combined give rise to competence profiles for the eight
alternatives A1–A8 as stated in Table 8.3.
Table 8.3 Alternatives and voting competence.
A1 A2 A3 A4 A5 A6 A7 A8
p? p p p not-p p not-p not-p not-p

q? q q not-q Q not-q q not-q not-q
r? r not-r R R not-r not-r r not-r
10% p-voters 0.15 0.15 0.15 0.1 0.15 0.1 0.1 0.1
10% q-voters 0.15 0.15 0.15 0.15 0.1 0.15 0.1 0.1
10% r-voters 0.15 0.1 0.15 0.15 0.1 0.1 0.15 0.1
70% others 0.125 0.125 0.125 0.125 0.125 0.125 0.125 0.125
We can now explore three different arrangements for conducting the vote:
Decision Rule 1: All voters vote together and their plurality winner is the
group choice.
Decision Rule 2: The p-, q-, and r-competent voters vote together and the
plurality winner of that vote among competent people only is the group
choice.
Decision Rule 3: The p-, q-, and r-competent voters vote separately, only
on their respective attribute (so the p-voters vote on p versus not-p, the
q-voters on q versus not-q, etc.), and the group choice is the alternative
having the combination of attributes determined by the winners of those
three separate votes.
Figure 8.4 shows the results. The first thing to observe is that all three
methods produce good results where the number of voters n is large.38 That
is because, in the example as constructed, the p-, q-, and r-competent voters
are a fixed proportion of the total number of voters. So as the total number of
voters increases, so does the absolute number of specialists in the population.
Ex hypothesi, in this example, everyone else votes randomly; so on the basis
of the non-specialist 70 per cent of the electorate’s votes the outcome would
be essentially tied. And the votes of the 30 per cent who have particular com-
petence on one dimension or another will break that tie, increasingly likely in
1.0
0.8
0.6
PnPV
DR3: specialists vote by attribute

DR2: only specialists vote
0.4
DR1: entire population votes
0.2
0.0
0 10,000 20,000 30,000 40,000 50,000
n
Figure 8.4 Population and specialist votes in comparison.
38 Note that in Figure 8.4 n refers to the total population size. If only specialists vote, then only
30% of n vote.
the correct direction the larger the absolute numbers of such specialists there
are in increasingly large electorates.
The correct alternative will thus have the highest probability of winning,
even in a plurality vote among all voters (Decision Rule 1—DR1).39 However,
since the random votes of the 70 per cent are not making any epistemically
useful contribution to that result, it would be even better to defer to the special-
ized voters in a more focused way. If we take a plurality vote just among all
specialized voters (Decision Rule 2), the group tends to make the right decision
more often. But it is even better—much better, with smaller electorates—to
ask specialists specifically about their own area of expertise and combine
their partial votes on the different dimensions to form the group decision
(Decision Rule 3).
A variation on Decision Rule 3 would be to treat the votes of p-, q-, and
r-competent voters negatively, to veto alternatives that they perceive as incorrect
on the dimension on which they are particularly competent. In the example as
constructed, only one option would then remain; and it would be identified as
the group choice with the same probability as shown by the line associated with
Decision Rule 3 in Figure 8.4. But in other setups, several options might remain
unvetoed and eligible as potential choices of the group.
Note, however, that the epistemic advantages of deferring to specialists (as
per Decision Rules 2 and 3) are quite sensitive to the competence of the ‘others’.
In Figure 8.4, we assume that the ‘others’ vote completely at random. But if
voters in the general population are even just a little better than random, the
advantage of dividing labour and specializing quickly evaporates.
39 While we assume that the groups of experts are of similar size, that is not required as long as
they are all sizeable and all voters remain ‘neutral’ (in the sense of making a random decision)
when they are not competent.
Discussion and Deliberation
Empirical studies of political participation find that by far the most common
way for people to engage with politics is by talking to one another informally
about it.1 Over-the-fence chats and barroom arguments—much more than
televised debates and doorstep campaigns—are the real stuff of lived political
life. This chapter is devoted to tracing the contributions that such interpersonal
communication might make to the epistemic quality of group decisions.
We do so in the particular context of the Condorcet Jury Theorem, which
notoriously has ‘aggregation’ rather than ‘deliberation’ at its core. Many fume at
that fact and bemoan the CJT as being of strictly limited usefulness for that
reason.2 But while the CJT’s procedure is indeed aggregative at the end of the
day, there is (as this chapter will show) considerable scope for discussion and
deliberation preceding the vote—and there are various ways in which building
that into our political processes can improve the outcome that emerges from
the eventual CJT vote.3
1 Huckfeldt and Sprague 1995. Delli Carpini, Cook, and Jacobs 2004, pp. 323–4. Jacobs, Cook,
and Delli Carpini 2009.
2 Waldron (1999b, p. 136), for example, writes that ‘Condorcet’s result is maddeningly mechan-
ical. . . . [It] takes no account of discussion, deliberation and persuasion – the very processes that
are likely to produce a legislative record to which . . . intention-seeking judge[s] can appeal.’ See
similarly Estlund (2005, p. 611) and Anderson (2006). Ironically, just a decade earlier Waldron
(1989, pp. 1325–8) had himself suggested using deliberation in conjunction with CJT-style
aggregation—a suggestion that this chapter is dedicated to exploring.
3 Thus we strongly concur with Swift’s (2014, p. 225) judgement that, while
the Condorcetian and the deliberative approaches posit quite different mechanisms by
which democracy might tend to produce good decisions, . . . they are not mutually exclusive.
Condorcet says nothing about how individuals come to their political judgements. It is quite
consistent with his mathematics that citizens should formulate their views about how to
vote through a process of debate and critical reflection with one another (as long as each ends
up voting for what she really thinks, not simply toeing a party line). . . . If the deliberative
account is right, that process of deliberation will tend to improve our judgements, making it
more likely that the average person is more likely to be right than wrong.
Discussion and Deliberation 133
9.1 THE IDEAL AND PRACTICE OF DELIBERATION
One of the great insights of ‘social epistemology’ is this: virtually all of our
knowledge is second-hand knowledge. Just about everything we know we
have learned from someone else: our mother, our teachers, our barber, our
colleagues, the BBC.4 Of course we don’t automatically believe everything we
are told. When someone else asserts X, we sit back and decide whether that
testimony is credible.5 Still, even if we don’t believe everything we hear, most of
what we think that we do know we have gotten from someone else.
Even if they are only one-way, communication flows increase our informa-
tion base remarkably. And even there, we are not passive recipients of informa-
tion. We actively reflect upon the new information. We decide in the first
instance whether or not to believe it is likely to be correct.6 In the second
instance, we make the effort of figuring out how that new information fits with
the rest of our related beliefs, and what adjustments might be needed to them
in light of that new information supposing that it is credible. This ‘deliberation
within’ involves a lot of genuine cognitive labour even on the part of recipients
of one-way information.7
In the give-and-take of genuine interpersonal discussion, those effects are
intensified.8 There too, of course, you learn new facts from others who have a
different information base than your own. In the course of conversing together,
you jointly explore how the facts fit together, each bringing a slightly different
perspective to bear. In those conversations, you cannot count on sweeping hard
questions under the carpet. Typically, you are challenged to defend the way you
put the facts together; why you think these facts, rather than some others, are
the really crucial ones bearing on the choice before you collectively; and so on.9
Of course some discussions are of a higher quality than others. Some amount
to little more than Punch and Judy-style conversational equivalents of barroom
4 Goldman 1999. See similarly: Hardin 2009; Talisse 2009, ch. 3; 2013. As Steven Shapin (1994,
p. xxv) writes: ‘Knowledge is a collective good. In securing our knowledge we rely upon others,
and we cannot dispense with that reliance.’
5 Coady 1992. 6 Practising what Sperber et al. (2010) call ‘epistemic vigilance’.
7 Goodin 2000.
8 As Talisse (2009, ch. 4) writes:
if . . . the characterization I have offered of folk epistemology is accurate . . . then . . . only in a
democracy can an individual practise proper epistemic agency. . . . Folk epistemology . . . justi-
fies democracy: democracy is the political entailment—indeed, the political manifestation—of
the folk epistemic commitments each of us already endorses.
Misak (2000, 94) writes similarly: ‘What it is to assert, to make a claim, to believe, to judge is
also to be engaged in a process of justification. It is to commit oneself to giving reasons—to be
prepared, in the appropriate circumstances, to justify the claim to others, and to oneself.’
9 Mill 1859/1977, ch. 2. That is to say, we need to be talking about a ‘deliberative culture’ culti-
vating the ‘epistemic virtues’ (Talisse 2005, pp. 109–16 and ch. 7; Farrelly 2012), rather than thinking
of deliberation as a series of one-off deliberative events like Deliberative Polls. There is evidence
that people who have had to do this once interpersonally for real, as part of jury service, do it
internally for evermore (Gastil et al. 2010).
brawls. Theorists of deliberative democracy are undoubtedly correct in thinking

that it would be better—in epistemic as well as in many other respects—if
interpersonal interactions were governed by the high standards approaching
Habermas’s ‘ideal speech situation’.10 A raft of small-scale experiments trying
to do just that show that, after formal deliberations in which moderators
enforce such rules, people’s opinions are different in all sorts of ways that would
presumably make them more competent voters.11
Although those are highly stylized deliberative settings, they are not without
real-world political relevance. Many of those same standards are written into
manuals of parliamentary practice, after a fashion.12 Even where they are not,
they typically figure at least in manuals of good manners.13 Of course, both sets
of instructions contained are often honoured in the breach. Still, it may not be
beyond hope that those ideals might be approximated in the real world, at least
in certain settings.
Whether those experiences and experiments can be scaled up to the society-
wide level is an open question.14 Our point here is simply that they do not need
to be. The epistemic benefits that come from interpersonal interaction (two-way,
or even just one-way) do not completely depend upon realization of those
higher deliberative democratic ideals as a society-wide exercise—epistemically
better though it would no doubt be, if that were realized.15
9.2 THE MANY BENEFITS OF DELIBERATION
In the context of the CJT, it is natural to think about deliberation primarily as a

mechanism for improving individual competence.16 There are, however, multiple
10 Habermas 1984. Bohman and Rehg 1997. Dryzek 2000. There are myriad designs for organ-
izing formal deliberative events (Fung 2003; Gastil and Levine 2005; Smith 2009); there are many
different game-theoretic representations of what goes on within them (Dickson et al. 2008; 2015;
Hafer and Landa 2007); and outcomes can turn heavily on context and process (Myers and
Mendelberg 2013). Here we try to step back from those levels of detail to say something about
deliberation more generically.
11 Gastil and Dillard 1999. Luskin et al. 2002. Dryzek and Niemeyer 2006. Niemeyer 2011. List
et al. 2013. Myers and Mendelberg 2013. For a now slightly dated survey of the empirical literature,
see Delli Carpini, Cook, and Jacobs (2004).
12 Goodin (2008, ch. 9) demonstrates strong parallels between Habermas’s (1984) ideal and the
prescriptions of Robert’s Rules of Order (Robert 1876/1951).
13 Martin 1999. Washington 1746.
14 Lupia 2002. Ackerman and Fishkin (2004) make a fascinating, but almost certainly not-to-
be-implemented proposal for doing that through a nationwide Deliberation Day immediately
before US presidential elections.
15 Just look down the list of effects that are outlined in Section 9.2: all can come about in the
absence of any high Habermas-style deliberative ideal.
16 Of course it is logically possible that discussion leads someone who originally was right to
change his view in the wrong direction. Given the standard CJT assumption of better-than-random
individual competence, there will probably be more people who start out with the correct view
than any other. Were the effects of discussion purely random, it would typically be better to
pathways from group deliberation to better group epistemic performance.17

Here is a list of some possible pathways:
1. Deliberation can increase individual competence and hence group
competence by increasing individual competence pc in terms of the
classical CJT.
2. Deliberation can make the votes less positively correlated (or even
negatively correlated) in terms of the classical CJT.
3. Deliberation can lead to more sincere voting.
4. Deliberation can increase the probability that the circumstances or evidence
are truth-conducive (not misleading), and the probability that a best
responder would find the correct answer increases.18
5. Deliberation can change the decision problem:
(a) by adding new evidence;
(b) by adding premises and/or using logical relations between these
premises and the conclusion;
(c) by adding new and better alternatives.
We will now explore each of these different pathways in greater detail.
9.2.1 Deliberation to Increase Individual Competence
Perhaps the initially most obvious way that deliberation can benefit group
epistemic performance is by increasing (average) individual competence pc .
The additional flow of information, and the increased possibility of having
proceed directly to a majority (or plurality) vote without discussion, since (given there are more
correct people that discussion might randomly lead into error than the other way around) dis-
cussion would typically serve to reduce the probability of the correct option winning (Maurin
and Vidal 2012). Hence the importance of emphasizing the various ways that follow in which
discussion and deliberation can be truth-conducive.
17 Marti (2006, emphasis in original) offers a similar list:
(1) Deliberation increases the exchange of information. . . . (2) Deliberation permits the
expression of intensities of preferences. (3) Deliberation permits and improves the detec-
tion of factual and logical mistakes. (4) Deliberation permits the control of emotional factors
and helps to filter irrational preferences. (5) Deliberation makes the manipulation of infor-
mation and political agenda more difficult. (6) Deliberation is a filter of impartiality and
substantive justice.
Fearon (1998, p. 45) offers yet another:
1. Reveal private information. 2. Lessen or overcome the impact of bounded rationality.
3. Force or encourage a particular mode of justifying demands or claims. 4. Help render the
ultimate choice legitimate in the eyes of the group. . . . 5. Improve the moral or intellectual
qualities of the participants. 6. Do the ‘right thing’, independent of the consequences
of discussion.
18 As discussed in the Best Responder Corollary in Section 5.3.
one’s possibly erroneous views challenged and corrected in debate, can increase
one’s probability of voting for the correct alternative.19 That improvement in
individual competence is particularly relevant within small groups, where an
improvement of pc can increase group competence substantially.
One way in which deliberation can increase individual and hence group
competence is by making hitherto private information public. Making public
information that was previously distributed between different people is par-
ticularly important when knowing all or most of the evidence is necessary to
deduce the correct answer. Take a playful example. Some will recall the murder
mystery board game ‘Cluedo’ (‘Clue’ in North America). In that game, different
players have different pieces of evidence about a crime case the players aim to
solve. If all the information were public everybody could solve the crime case
immediately (and the game would be rather boring). The game Cluedo enter-
tains us precisely because the evidence is distributed among the players. Many
real-life epistemic challenges are like that.20 Expert committees, for example,
are often chosen because the different experts know different facts or different
‘parts of the puzzle’. Different people often know different things, and if they all
knew what they know separately (and all the logical entailments of that set of
propositions), they would be in a much better epistemic position. Deliberation
can tease out and publicize among the deliberators such private information,
because it invites participants to reveal what they know and it incentivizes
them to do so in order to underpin their arguments.21
Just how effective group deliberation is at pooling information to reach the
right conclusion depends very much on the rules governing those deliber
ations, of course. When participants are under instruction not to criticize one
another’s proposals, but merely add their own to those already under consider-
ation, the group generates quantitatively far fewer and qualitatively often worse
options.22 Various mechanisms drive that result.23 One is what psychologists
call ‘production blocking’ (in conversational settings you have to wait your
turn, and you forget some of your good ideas before your time comes to speak).
Another is ‘group polarization’ (groups talk themselves into more extreme ver-
sions of the position towards which they were initially inclined).24 Yet another
19 Fearon 1998, pp. 45–9. Fuerstein 2013.

20 Juries, however, are not (supposed to be) put in such a situation, as one basic premise of
court procedures is that all evidence is public knowledge, at least among the jury and the judges.
21 Mercier and Sperber 2011.
22 The ‘no criticism’ rule is the hallmark of Osborn’s (1948) original proposal for ‘brainstorm-
ing’. Social psychological studies have shown, time and again, that groups deliberating on that
basis do worse than an equal number of individuals deliberating on their own and then just
pooling their separate lists of proposals at the end of the same period of time. See: Diehl and
Stroebe 1987; 1991; Sawyer 2007, ch. 4.
23 See Sunstein (2006b; 2006c, pp. 75–102), in particular, for a survey and application to theories
of deliberative democracy.
24 Isenberg 1986. Sunstein 2002; 2009b. Mutz 2006.
is ‘common knowledge’ (information that is common across more members of

the group has more influence on group deliberations, and information that is
held by only a few members of the group is suppressed).25
When deliberators are instructed to engage critically with one another,
however, things are very different.26 Debates create competitive environments
in which all participants are more tempted to present the best arguments and
evidence they know of to make the best possible case for their point.27 Whereas
under the more collaborative ‘no criticism’ rule information that is not widely
shared across the group gets suppressed or ignored, under the more competitive
rules of debates participants are incentivized to reveal any private information
they have bearing on the group choice—thereby garnering credit for leading
the group in the right direction.28
Deliberative groups operating under rules that encourage the critical
engagement of participants with one another are likely to do three things.
One is to get new and better options onto the table, and thus transform the
decision situation; we will discuss that sort of contribution in Section 9.2.5.
A second is to improve the evidence base in such a way as to increase the
probability that the ‘best responder’ to that evidence can reach the correct
conclusion; we discuss that sort of contribution in Section 9.2.4. The third is
25 Stasser and Titus (1985) is the seminal paper on the ‘hidden profile’ paradigm. See further
Gigone and Hastie (1993).
26 They are, when deliberators actually act upon the instructions given. What Fricker
(2007; 2013) calls ‘epistemic injustice’ arises from certain other people’s statements being ignored,
or anyway heavily discounted—and insofar as that discounting goes beyond what would be war-
ranted by the objective probability that the statements are not true, that constitutes an epistemic
cost for the (non)listener as well as an epistemic injustice to the ignored speaker.
27 That is the claim of Mill (1859/1977), elaborated by Talisse (2009, p. 107) thus:
[I]nsofar as they see themselves and each other as engaged in a common epistemic enter-
prise, their disagreements, though often intense and heated, can be . . . driven by the shared
aspiration to follow the best reasons. Hence, proper discourse . . . sees dissensus as a condi-
tion that could enable proper believing. In this way, proper discourse should be expected to
be agonistic, not always and not necessarily calm, pleasant, and cooperative.
Lindsay (1929, p. 37), one of the earliest defenders of ‘government by discussion’, writes similarly:
‘Every scientific discoverer knows that what he most wants to know is not what can be said for, but
what can be said against his theory. What he most wants is an opposition’; and ‘any one with
experience of the effectiveness of discussion in a small . . . society must recognize how valuable is
the contribution of those who are not easily convinced but can stand up resolutely for their own
point of view.’ Among contemporary deliberative democrats, Bächtiger (2011) has most clearly
seen the epistemic importance of ‘contestatory deliberation’.
Whether the competitive exchange of arguments leads to epistemically good results depends
on the argumentative setting. Mercier and Sperber (2011) argue that the main function of reason-
ing and argumentation is not to improve individual cognition but to avoid being duped. The
biggest epistemic advantages emerge not from individual but from group reasoning under suit-
able conditions, as they show with reference to many recent experiments. See further Mercier and
Landemore (2012).
28 For evidence from experimental psychology see: Nemeth et al. 2004; Mercier and
Landemore 2012.
to increase the individual competence of participants; that is the topic under

discussion in this section.
That latter sort of contribution will, in and of itself, typically make little
difference to the overall epistemic performance of the group as a whole. The reason
is simply that, in large groups, the collective competence of the group as a whole
(Pn) is not particularly sensitive to an increase of individual competence pc , or
anyway it is not once pc is already appreciably better than random. For a large
group, any mean individual competence level appreciably above random will have
almost ‘maxed out’ the ‘wisdom of crowds’ effect already. Increasing mean individ-
ual competence further does not, therefore, have much effect for large groups.
9.2.2 Deliberation to Reduce Dependence
So far we have been focusing on the obvious advantages of discussion and

deliberation for voter competence. Before continuing in that optimistic vein,
however, we need to face a theoretical complication touched upon briefly in
Section 4.5 and at length in Chapter 5: voter deliberation not only affects voter
competence; it also affects the independence of votes. And while there are good
reasons to assume that voter competence is increased by discussion, the nature
and consequences of its effects on independence are more ambiguous.
Rousseau famously recommends that assemblies proceed directly to a vote
without any preceding discussion. Condorcet sometimes agrees that would be
best—even whilst recognizing that that realistically is not an option.29 The
reasons Rousseau and Condorcet offer for eschewing discussion are mostly
first-order political reasons, having to do with the way in which discussion can
foster factionalism.30 But in addition to that, there is often thought to be a more
formal CJT-style reason as well.
The CJT requires votes to be independent of one another. On one naive
understanding, that amounts to a ‘no contact’ rule. On that understanding of
what Independence requires, any interpersonal interaction among people—
and most especially any sharing of information among them—violates the CJT’s
Independence Assumption and prevents the theorem from being applicable at
all. That is a mistaken understanding of what the Independence Assumption
requires, as we have shown in Section 5.1.1. Still, some of the very best political
philosophers, from Rawls down, have been misled in this way.31
29 Rousseau 1762/1997, bk. 2, ch. 3. Condorcet 1789/1994, pp. 178–9.

30 Although that may be epistemically no bad thing, as we argue in Chapter 14.
31 Rawls (1971, p. 358) dismisses the CJT cursorily, saying: ‘it is clear that the votes of different
persons are not independent. Since their views will be influenced by the course of the discussion,
the simpler sorts of probabilistic reasoning do not apply’ (see similarly Cohen 2010, p. 79). While
that is technically correct, those not au courant with the technical details are all too tempted to
read this as a dismissal of the CJT on grounds that discussion violates Independence, and that’s
the end of the matter.
With regard to independence, deliberation can cut both ways. On one hand,
the exchange of information between voters might make their votes more
interdependent: rumours spread through the group; opinion leaders emerge;
voters might be led to develop a more homogeneous view of the evidence,
theories, heuristics, and cues to interpret it; and so on.32 In the worst case,
deliberation can cause positive correlation among the votes, without making
the voters more competent at all. If increased dependence is not compensated
for by higher individual competence, then deliberation will lower group
competence—perhaps severely.33
On the other hand, deliberation might also make the votes less interdepend-
ent. By deliberating, for example, the voters might find out that they paid too
much attention to Fox News and begin to follow a more diverse mix of news
sources. Or they might discover that their hitherto shared world view is not the
only way to see the world, introducing a greater diversity of common causes
pulling in different directions.34 Finally, deliberation might motivate voters to
make up their own minds, rather than blindly following what someone else tells
them. Such forms of deliberation would increase the independence of votes.
Groupthink and group conformism—the tendency for people to hesitate to
express their own true views if they are contrary to the views of many around
them35—is a serious threat to the CJT result. If people vote the way others vote,
just because that is the way they vote, that violates the CJT’s crucial Independence
Assumption in a most fundamental respect.36 Evidence of tendencies toward
‘group polarization’ within deliberating groups is sometimes taken—arguably
wrongly—as evidence of just that.37
32 Deliberations within juries are sometimes criticized on such bases (Hedden 2017).
33 This is why Francis Galton (1907b) proposed that juries should at least prevent the direct
causal influence between votes: ‘I suggest that the process for a jury on their retirement [to reach
a verdict] should be (1) to discuss and interchange views; [and then] (2) for each juryman to write
his own independent estimate on a separate slip of paper. . . .’ This would be a form of what
Vermeule (2015, p. 223) calls ‘open-secret voting’: ‘The hope is that the open vote will induce
maximally responsible judgments while the secret vote induces maximally autonomous judg-
ments, and that the combination of the open and secret votes in succession will prove superior to
either taken alone.’
34 See our discussion of diversity in Chapter 7 above.
35 The tendency has long been discussed. In Federalist no. 49, Madison (1788/2003, p. 247)
writes, ‘The reason of man, like man himself, is timid and cautious when left alone, and acquires
firmness and confidence in proportion to the number with which it is associated.’
36 As we have discussed in Section 4.5.1—although, as we show in Chapter 5, there are ways to
cope even with violations of Independence of that sort.
37 In line with social psychological conventional wisdom (Myers and Mendelberg 2013,
pp. 714–20), Cass Sunstein (2000; 2002; 2006c, pp. 92–8; 2009) offers a raft of examples and evidence
of group polarization. But even he admits that group polarization might have a perfectly rational
explanation: ‘most people listen to the arguments made by other people . . . [and as] a statistical
matter, the arguments favoring [the] initial position will be more numerous than the argu-
ments pointing in the other direction.’ That, Sunstein (2006c, p. 94) admits, is the ‘most import-
ant’ source of group polarization. For another rationalization of seeming ‘group polarization’
see Goodin (2009, pp. 7–9). Both explanations seem to fit with the finding of Ambrus et al. (2015)
that ‘preferences tend to shift towards the choice of the individual’s previous group’. We hasten to
If, however, the group proceeds under rules of ‘free and open discussion’,
encouraging people to speak freely and to express diverse views, that result can
often be avoided.38 One person’s expression of dissent might then liberate
others to voice their own doubts as well, thus reducing any perceived pressure
to conform to the majority or the perceived consensus opinion. It may reduce
positive correlation among votes by exposing and then reducing the influ-
ence of common causes, making the votes more independent. In the best pos-
sible case, discussion might conceivably induce a level of diversity that leads to
negatively correlated votes. Those negative correlations will actually improve
the epistemic performance of the group as a whole, as Section 7.2 has shown.
Deliberation in diverse groups might similarly help to remove shared judge-
mental biases arising from people’s shared values via ‘motivated reasoning’ of
the sort discussed in Section 4.4. Motivated reasoning can persist within
enclaves where everyone shares the same values and is motivated in the
same direction to believe the same things. But when you are conversationally
required to justify your belief to someone who does not share your same motiv
ational structure, your own reasoning is improved. You are driven back to the
evidence, as it intersubjectively appears and not as you wish it to be. And, epi-
stemically, that must surely be a good thing. Experiments have shown this to
be the case, both with groups reasoning together in general and in various
countries involving deliberation across deep divides.39
9.2.3 Deliberation Induces Sincerity
Another crucial assumption of the classical CJT is that people vote sincerely for
what they genuinely regard as the correct alternative, rather than voting for
some other alternative in hopes of strategically manipulating the outcome of
the vote.
add that rationalizing ‘group polarization’ in these ways does not make the violation of the classic
CJT Independence Assumption disappear. It merely changes the common cause leading to the
violation, away from votes depending on other people’s votes, and (more innocuously) towards
everyone’s votes depending on the same stock of evidence.
38 This is Ladha’s (1992, pp. 630–1) epistemic argument for free speech.
39 Mercier (2011, p. 318) summarizes the former literature as follows:
When people who disagree argue with each other, reasons that would be only poorly evaluated
internally become the object of intense scrutiny. The poorest arguments are weeded out
and the group can proceed to a better solution. In such a context, the confirmation bias, an
impediment to individual reasoning, becomes a form of division of cognitive labor, with
opposite biases balancing out.
On deliberation across deep divides, see: Caluwaerts and Deschouwer 2014; Luskin et al. 2014;
Flynn, Nyhan, and Reifler 2017.
We have offered various reasons in Section 4.3 for thinking that, in information-

pooling settings of the sort envisaged by the CJT, people would ordinarily
employ sincerity as their default rule. The most reliable way to true belief is to
test your provisional beliefs in the cauldron of deliberation and debate with
others. So says everyone from John Stuart Mill40 to American pragmatists like
C. S. Peirce41 and discourse ethicists like Jürgen Habermas.42 But if you believe
that, and that is your motive for engaging in discussion and debate with others,
then you will have no motive for misrepresenting your beliefs or unfairly char-
acterizing theirs.43
Experimental evidence reported in Section 4.3.4 suggests, however, that
among deliberative groups truth telling is more than a default rule—more than
merely a defeasible presumption. That evidence suggests that, if there is com-
munication within the group before it reaches a decision, even people who might
be inclined to behave strategically without any prior communication will sincerely
reveal whatever information that they have.44 That, then, is another advantage
of deliberation: it induces sincerity.
9.2.4 Deliberation to Improve Best Responder Performance
So far we have discussed the problem of dependence from the perspective of

the classical CJT. As Chapter 5 showed, however, there are more technically
refined ways to deal with the dependence-by-deliberation issue. These deal
with common-cause dependence by carefully conditionalizing on all the com-
mon causes. As we have shown with the Best Responder Corollary in Section 5.3,
such a jury theorem does not necessarily have the group competence converge
to 1. Instead, the convergence is upper-bounded by the probability that a ‘best
responder’ will get it right, given the evidence or the decision problem more
generally. ‘Garbage in, garbage out’, as the old saying from computer science
goes—if the evidence is poor or the circumstances misleading, neither the best
responder nor the masses can find the right answer.
40 1859b, ch. 2.

41 As analysed by Misak (2000, p. 94) and Talisse (2005, pp. 103–4; 2009, p. 123; 2013).
42 Habermas 1990.
43 As Talisse (2009, p. 106, emphasis in original) says:
[I]f we aim to have true beliefs, and if this aiming requires us to exchange our reasons with
others, we must avoid adopting attitudes and habits that obstruct or frustrate the dialectical
process of examining and exchanging reasons. So, with regard to our interlocutors, we must
be open-minded, attentive, honest and charitable . . . . It bears emphasizing that this . . . is
emphatically not a moral conception; it is thoroughly epistemic.
44 Fearon (1998, p. 48) points to one mechanism that might be at work: ‘social conventions might
just entail discomfort for people caught publicly lying, so that if there is some chance of being caught,
discussion might be rendered somewhat informative while private voting would be less so’.
The probability of the best responder giving the right answer, which is the
probability that the circumstances are truth-conducive, determines the upper
boundary of group competence. It becomes the crucial factor for epistemic
success. In our previous discussions, we have been thinking in terms of what
the best responder can do, epistemically, as being fixed. For example, if a jury is
confronted with misleading evidence in 10 per cent of all cases it decides, then
the best responder will also fail in that proportion of cases, but will get the other
90 per cent of cases right. The jury, by pooling their information, can, if large,
approach that performance benchmark set by the best responder—but it can
never be better.
Assume now, however, that we teach the jury some ‘new tricks’ to lift this
upper bound. For example, let the jury members attend the ‘Inspector Columbo
School of Critical Investigation’, being led in the course of deliberative engage-
ment with one another to distrust evidence that presents itself as too obvious.
Employing critical reasoning and more persistent investigation (‘One more
question, sir . . . ’), the best responder relative to that situation might be misled
in a lower proportion of cases, for example down to 5 per cent from 10 per cent.45
Since the jury is now less often systematically misled by the evidence, its group
competence is now upper-bounded at 95 per cent—the ‘new tricks’ have
reduced the incidence of misleading cases by 50 per cent!
The upshot is that an increase in competence can come in two ways. One is
in the classical CJT framework, as an increase in the probability of voting for
the correct alternative. The other is in the more refined Best Responder
Corollary, as an increase in the probability of not being systematically misled
by the decision problem.
9.2.5 Deliberation Can Change the Decision Problem
The previous section showed how the decision problem can change in one
specific sense: making it less misleading by increasing the abilities of the
voters. Now we turn to ways in which deliberation can lead to a more funda-
mental change of the decision problem itself. If done well, deliberation
can not only provide additional evidence; it can also add new alternatives
to the agenda; or it can recast the decision problem so the group is reflecting
not only on the question overall but also on certain premises logically
related to the correct answer.46 By changing the decision problem in either
45 See further Section 5.5.

46 Fearon (1998, pp. 49–50) speculates that, faced with
the fact that our imaginations and calculating abilities are limited and fallible, . . . individuals
might wish to pool their limited capabilities through discussion and so increase the odds of
making a good choice. Discussion might lessen the impact of bounded rationality for two
reasons. First, it might be ‘additively’ valuable in that you might think of some possibility that
of these ways, the probability that the group will choose the best outcome
can increase.47
As regards the first mechanism, remember that all that the CJT proves is that
the majority of a large group of independent voters who are better-than-
random is almost certain to choose the best alternative from among the alterna-
tives presented to them. The classic CJT model requires not only that the agenda
has to be well defined; it must also contain within it the correct alternative. If
the correct alternative is missing from the agenda, the correct alternative cannot
be chosen.48 For groups too large to deliberate together as a whole, perhaps the
agenda simply must be taken as given.49 But among a smaller and more discur-
sive group, discussion of what options ought to be on the table is genuinely
possible, in a way that it is not among larger and less discursive groups. And at
least sometimes that will lead to an improved set of alternatives being on the
agenda for deliberation and decision.50 Insofar as one of those alternatives is
correct, the group is able to select it. That is another way in which discussion
and deliberation among small groups can cause the collective competence of
the group to increase.51
It is no mere theoretical speculation that deliberation in small groups can
generate more and better alternatives for the agenda. There is evidence from
empirical social psychological experiments to that effect as well. Deliberating
hadn’t occurred to me, and vice versa. Second, it might be ‘multiplicatively’ valuable in that
in the course of discussion we might think of possibilities or problems that would not have
occurred to either of us by ourselves (this is brainstorming, I suppose).
In short, he suggests that discussion is ‘a means of learning things that you never even conceived
of ’, and in that way remedying ‘failure[s] of the imagination’.
47 As Condorcet (1793/1994, p. 193) clearly foresaw:
Discussions in a debating assembly clearly have two main concerns. . . . In [the] first stage,
the opinions are . . . all different and none obtains the majority of votes. . . . As the question
becomes clearer, the opinions become less diverse and begin to combine into a small number
of more general opinions. Soon, the question can be reduced to a number of clear and simple
questions about which the assembly can be consulted.
48 Goodin 2008, pp. 122–4.
49 As has long been familiar: see e.g. Aristotle’s Politics (350 bce/1997, 1268 b). But perhaps not:
consider for example the recent attempt at crowdsourcing a new constitution for Iceland
(Landemore 2015).
50 Mercier and Sperber 2011, p. 65. See Goodin (2017) for a worked example illustrating this, by
reference to discussions within the small Executive Committee of the National Security Council
convened by President Kennedy to help him cope with the 1962 Cuban Missile Crisis. This example
puts paid to Condorcet’s (1793/1994, p. 193) speculation that the generating and winnowing of
options ‘does not require everyone to come together in the same assembly, and can in fact be
conducted just as well, and maybe better, in writing.’ It is clear from the transcripts of those
ExComm deliberations how important conversational back-and-forth within the assembled
group was for originating and refining new options.
51 This point is related to observation discussed in Section 4.2.2 above that the quality of col-
lective decisions depends on the agenda (Fuerstein 2008). Note that the effect of deliberation can
cut both ways. It can improve the definition of the alternatives so that the CJT applies (Fearon 1998,
p. 49), but it could also render the decision problem less tractable.
together in a critically engaged manner has been shown to lead to better

solutions, particularly where the problem under discussion is a complex one
and creativity in solving it is required.52
A second effect of discussion and deliberation among a smaller and more
discursive group might be on processes of reasoning. Among large groups,
decisions are inevitably taken by aggregating people’s votes on the ‘bottom line’.
Among smaller groups, there are more discursive opportunities for ‘premise-
probing’ (‘what makes you think that?’).53 The correct answer at the ‘bottom
line’ might be the result of a reasoning process involving many premises.
Smartly aggregating what the individual voters know about the premises
increases the probability of correctness regarding the bottom line.54
That can obviously improve individual-level decision-making, insofar as
each internalizes the lessons learned from those probes of her own and others’
premises and revises her chain of reasoning accordingly. It might also improve
group-level decision-making, insofar as the decision process of the group
involves pooling people’s views on premises and not just votes on the ‘bottom
line’.55 Incorporating premises is often (but not always) a better way to pool
information and therefore reach better results.
9.3 THE DELIBERATION EFFECT
At various points in the Federalist Papers’ defence of the new Constitution, the
American Founders were counting heavily upon the capacity of people, when
deliberating together, to reach more nearly correct decisions. That, for example,
is one of the grounds they give for thinking that the smaller and more convers-
able Electoral College would be more likely to choose a good president than the
electorate as a whole.56 That is also one of the reasons they give for thinking that
the president’s qualified power of veto—‘returning [vetoed] bills with objec-
tions . . . for re-consideration’ and further discussion57—would improve the
chances of only the correct legislation being enacted.
52 Nemeth et al. 2004. Sawyer 2007, ch. 4. One writer on ‘innovation in organizations’ literally
defines ‘creativity’ in these terms: as ‘bringing something new into being’ (Mohr 1969, p. 112).
53 Goodin 2008, pp. 87–92.
54 Betz (2013) provides sophisticated simulation modelling revealing the various mechanisms
by which this might occur.
55 Bovens and Rabinowicz 2006. List 2005a; 2006; 2011. Pigozzi 2006. Spiekermann 2010.
56 Hamilton in Federalist no. 68 (1788/2003, p. 331) writes, ‘It was . . . desirable that the imme-
diate election should be made by men most capable of analyzing the qualities adapted to the station,
and acting under circumstances favourable to deliberation.’ For a worked numerical example,
see Section 16.1.3.
57 Hamilton, Federalist no. 73 (1788/2003), pp. 357–8, 360.
We have identified mechanisms by which discussion and deliberation as

such might make a group (particularly a small and hence more discursive
group) collectively more competent. Taken together, we will refer to them as
‘the Deliberation Effect’. We do not know how big an effect those features of
discussion and deliberation have, taken either individually or together. We
have no empirical evidence to offer on that issue. Nor can we think of any good,
non-arbitrary way of modelling those effects for realistic social settings.58
What we will do instead is to proceed on a deliberately low estimate. In our
worked numerical example in Section 16.1.3, we assume discussion and delib-
eration improve smaller and more discursive groups’ competence by a small
amount. We would expect (but as we say, with no particular evidence) the
actual effect almost certainly to be greater. But we start with a conservative
estimate. And we then proceed to demonstrate in that section that, given this
small group competence boost from the Deliberation Effect, members of vastly
smaller groups (such as the 539-person US Electoral College) have to be only
very slightly and utterly achievably more individually competent for that
smaller group epistemically to outperform a vastly larger group (such as the
131-million-strong US electorate). To foreshadow the results that we will present
in Section 16.1.3, the Deliberation Effect can make a huge difference, in reducing
the individual competence levels required for smaller and more discursive groups
to be collectively as competent as very large groups.
58 However, for less realistic social settings, experiments conducted by social psychologists
suggest that deliberation can be beneficial for problem solving. That is particularly true when the
correct solution is difficult to find, but once it is found its correctness is easy to demonstrate. In
such instances, groups are at an advantage because even if just one member is able to find the
correct solution, that member can convince everyone else in group discussion. See Larson (2010,
chs 4–6) for a review.
Part III
Political Practices
In this part of the book we will focus on the implications of the Condorcet Jury
Theorem for four familiar political practices.
The first two correspond to two of Max Weber’s classic sources of authority:
respecting tradition (Chapter 10) and following leaders (Chapter 11).1 The
first, we show, ordinarily carries high epistemic costs. The second, about which
commentators on the CJT have worried more, may or may not be epistemically
costly for the group, depending upon just how diverse the leaders are and how
much power they exercise over the electorate.
A more general version of the latter practice is ‘taking cues’ (Chapter 12).
Taking guidance from competent cues can be particularly epistemically helpful
in the case of non-competent (and all the more so in the case of positively
incompetent) voters. Even in the case of competent voters, cue-taking can be
relatively epistemically innocuous, particularly (but not exclusively) if there are
several competent and independent cues that are being followed.
The final pair of chapters in this part of the book refer to practices arising from
the fact of pluralism in contemporary political communities. People have differ-
ing values and priorities and group-specific interests. In Chapter 13 we discuss
what we have called ‘moral majoritarianism’—the proposition that, assuming the
other conditions of the CJT are satisfied, democratic majorities will in such cir-
cumstances favour the correct outcome from the perspective of the values or
priorities that the largest segment of the community favours. In Chapter 14 we
discuss ‘factionalism’ as a way in which people with group-specific interests can
use CJT-style information-pooling strategies to secure reliable assessments of
what their group interests truly are. As we there show, factionalism can enable
less-well-informed masses to overcome the epistemic advantages of better-
informed elites in distributional struggle.
1 Weber 1947.
10
Respecting Tradition
There are many reasons to value—and indeed, to invent—traditions in terms of

giving people a common sense of identity and purpose.1 But for the purposes
of this book, it is the epistemic claims made in favour of respecting tradition
that will concern us.
Consider as an example of this Hayek’s claims on behalf of what we will call
(though he did not) ‘the wisdom of the ages’:
We would destroy the foundations of much successful action if we disdained to
rely on ways of doing things evolved by the process of trial and error where only
the superior manner, but not the reason for adopting it, has been handed down to us.
The appropriateness of our conduct is not necessarily dependent on our knowing
why it is so. Such understanding is one way of making our conduct appropriate,
but not the only one. . . .
This estimation of tradition and custom, of grown institutions, and of rules whose
origins and rationale we do not know, does not, of course, mean—as Thomas
Jefferson believed with a characteristic rationalist misconception—that we ‘ascribe
to men of the preceding age a wisdom more than human, and . . . suppose what
they did beyond amendment’.2 Far from assuming that those who created the
institutions were wiser than we are, the evolutionary view is based on the insight
that the result of the experimentation of many generations may embody more
experience than any one man possesses.3
One way to understand such claims on behalf of traditionalism is to subsume
them under the CJT’s ‘wisdom of the multitude’ rubric, pointing out that over
the ages there have been many more people than are presently alive. Where
Edmund Burke conceives of society as ‘a contract between the living, the dead
and those who are yet to be born’, we might in CJT spirit conceive of inherited
1 Pocock 1957. Shils 1981. Hobsbawm and Ranger 1992. 2 Jefferson 1816.

3 Hayek 1958, pp. 237, 235; see similarly Hume 1754/1760. Cf. Holmes (1897, p. 469):
It is revolting to have no better reason for a rule of law than that so it was laid down in the time
of Henry IV. It is still more revolting if the grounds upon which it was laid down have vanished
long since, and the rule simply persists from blind imitation of the past.
tradition instead simply as an information pool that is shared in common

among everyone past and present.4
There are other ways of reading Burke, to be sure. But that is a perfectly
plausible way of reading passages such as this, from his Reflections on the
Revolution in France:
The science of government being . . . practical in itself . . ., a matter which requires
experience, and even more experience than any person can gain in his whole life,
however sagacious and observing he may be, it is with infinite caution that any
man ought venture upon pulling down an edifice which has answered in any
tolerable degree, for ages the common purposes of society. . . .5
Similarly, he thinks of the ‘common law as a form of customary law’—as ‘the
collected reason of ages, combining the principles of original justice with
the infinite variety of human concerns’.6 In Cass Sunstein’s gloss on the point,
‘the best account of why traditions . . . deserve respect . . ., coming from Burke
himself, emphasizes that many minds have contributed to long-standing
practices, and thus give them a kind of epistemic credential’.7
10.1 TRADITIONALISM IN PRACTICE:

PRECEDENT IN THE COURTS
In modeling the epistemic consequences of respecting traditions, we will set

our story in the courts. Courts are a useful example because they exist for a long
time, have to make repeat decisions about similar or related matters, and often
decide on matters to which a standard of correctness applies. In addition, courts
typically follow a strong norm of stare decisis—following previous precedents,
unless there is compelling reason to overturn them in the case at hand.
There are many reasons for the rule of stare decisis. In part it is a matter of
fairness and predictability and, indeed, of the ‘rule of law’ itself.8 In attempting
to craft their action in such a way that it is in accordance with the law, people
inevitably rely upon past judgments to surmise what the law is to which they
will be held.9 Upsetting precedents wrong-foots citizens who are, in good faith,
trying to do as the law requires.
4 As suggested in Landemore’s (2013b, 239–40) cautious remarks on ‘the wisdom of the past many’.
5 Quoted in Sunstein 2009, p. 50.
6 Quoted in Sunstein 2009a, p. 52. See similarly: Moore 1996, pp. 266–8; Strauss 1996, pp. 891–4.
7 Sunstein 2009a, p. 88.
8 Justices O’Connor, Kennedy, and Souter (1992, pp. 699–700) write, ‘the very concept of the
rule of law underlying our Constitution requires such continuity over time that a respect for pre-
cedents is, by definition, inescapable’. See further Douglas 1949.
9 For example, Justice Rehnquist (1991, p. 737) writes that ‘considerations in favor of stare deci-
sis are at their acme in cases involving property and contract rights, where reliance interests are
involved’. See further: Douglas 1949; Schauer 1987, pp. 595–602.
Respecting Tradition 151
But in part there is an epistemic rationale underlying the rule of stare decisis,
as well. This is the thought that ‘the judgements embodied in long-standing
practices’ are simply more likely to be correct than ‘each person’s “private stock
of wisdom” ’. Pragmatically, ‘those who follow entrenched practices, or who
attempt humbly to build on them, will do much better than those who abandon
traditions or evaluate them by reference to an abstract theory’.10 There is evi-
dence aplenty of that thought, too, being at work among ‘minimalists’ on the
US Supreme Court.11
A note on the empirical facts of the matter. When we say ‘there is a strong
norm of stare decisis in the courts’, we merely mean to say that that norm is
often invoked and indeed often followed in the sense that judges feel compelled
to offer precedents in support of their decisions. Whether or not judges decide
as they do because of those precedents, or whether they are just paying lip service
to the precedents, may well be another matter. Scholars of judicial behaviour
adduce strong evidence that the latter might well be the case.12 If that evidence
is correct, and judges typically do not follow precedents that go contrary to
their own views, then they will be similar to the ‘stubborn judges’ we will discuss
in Section 10.4.2—and so much the better for the epistemic performance of the
court as a whole, as we shall show there.
10.2 THE EPISTEMIC COSTS OF COMPLETE

DEFERENCE TO PREVIOUS DECISIONS
The original CJT setup tends to assume that voters are making decisions sim-
ultaneously with one another, or anyway in ignorance of (or indifference to)
what other voters have done. The traditionalist approach to constitutional
interpretation envisages something very different. There, judges in courts are
making their decisions sequentially, not simultaneously. Furthermore, when
being traditionalists, subsequent judges make their decisions not only in
knowledge of but also in deference to earlier judges’ decisions. That changes
things dramatically, from a CJT perspective.
Those issues are standardly posed as ones of ‘information cascades’, as we
shall go on to discuss shortly. They can also be described in terms of the sorts
of causal diagrams we introduced in Chapter 5—for at root what is at stake here
is a violation of the CJT’s Independence Assumption.
In connection to this, let us consider two scenarios, one representing the
best case and the other the worst. For the best-case scenario, suppose your
10 Sunstein 2009a, p. 91.

11 Provided in Sunstein (2006a), which is the longer article upon which the relevant chapters
of Sunstein’s (2009a, chs 2–3) book draws.
12 Segal and Spaeth 1996. For more strategic versions, see e.g.: Knight and Epstein 1996; Bueno
de Mesquita and Stephenson 2002.
(a) State (b) State
Evidence
Evidence
Vote 1
Vote 1 Vote 2 Vote 3 ... Vote n-1

Vote 2
Vote n Vote 3
Vote 4
...
Figure 10.1 Differing points at which deference might set in.
court is the first court to take the traditionalist approach to constitutional

interpretation; and suppose it does so only after many other courts have already
ruled independently of one another on just the same constitutional question.13
That is the situation represented (in truncated form) in Figure 10.1(a). Vote n is
dependent on Votes 1 through n − 1; but at least Votes 1 through n − 1 are them-
selves independent, conditional on the Evidence. And that can be epistemically
very good news.
Here is a numerical illustration of that fact. Suppose that before you there
have been 999 decisions by judges (suppose each court consists of nine judges,
so there have been 111 court judgments with nine judges each) who have inde-
pendently judged the same matter. Then even if each of those judges had been
only pc = 0.55 likely to be correct, the majority among the 999 decisions would
(by standard CJT calculations) be Pn = 0.999 likely to be correct. The CJT’s ‘many
minds’ argument would have real epistemic bite, in those circumstances.
But why on earth suppose that your generation is the first to implement
Burkean principles? Burke has been a famous figure for over two centuries, and
principles of stare decisis go back long before that.
Consider, then, the opposite extreme, which constitutes our worst-case
scenario. Suppose that judges on the very first court to consider the question
were the only ones to exercise any independent judgement at all. Suppose that
thereafter all subsequent judges who have considered that same question have
13 What constitutes the ‘same question’ is of course an important matter of interpretation in
and of itself, allowing scope for independent judgement (Sunstein 2009a, pp. 63–7; see also Schauer
1987, pp. 577–88). To simplify modelling, we ignore that.
simply deferred to that same initial court’s decision (or deferred to previous
judges who were deferring to that initial court in turn). Then we have a situ-
ation similar to that represented by Figure 10.1(b). Here (for parsimonious
visualization) Vote 1 is the only vote that is influenced directly by the Evidence.
Vote 2 already depends on Vote 1; Vote 3 depends on Votes 1 and 2; and so on.
These votes are highly interdependent. The epistemic consequences of such a
situation can be quite deleterious.
For a numerical illustration similar to that, go back to our example with
nine judges, as above. Assuming now each of the nine judges on the court is
individually pc = 0.55 competent, the probability that the initial court reached
the correct solution would be Pn = 0.621. But given that each subsequent court
follows the ones before it with probability π = 1.0, the probability of any subse-
quent court reaching the correct solution is exactly the same as the probability
that the initial court did so. No matter how many copies of that initial court’s
opinion there have subsequently been, the fact that they are mere copies with
no independent content means that they add nothing epistemically to the ini-
tial court’s performance.
Where each member of every subsequent court votes exactly the same way
as the initial court, precisely because that was the way the initial court voted,
there is no independent judgement being exercised by any members of the sub-
sequent courts. What we would then have is a case not of ‘many minds,’ but
rather of ‘few minds, many mimics’. On that scenario, a court would do no
better, epistemically, following many previous courts’ precedents than it would
by simply deciding the matter for itself.14 The probability of that court’s being
collectively correct is Pn = 0.621 either way, on the assumptions just sketched.
The rest of this chapter will be devoted to modelling scenarios in between
those two extreme cases of complete independence and complete lack of
independence.
10.3 MODELLING PARTIAL DEFERENCE
In all of the models of traditionalism that follow, we will assume that judges
have to decide on the same dichotomous question at various different points
in time. We assume that each judge has an independent and symmetric private
signal as to which of the two alternatives is the correct one; and we assume
that that signal is more likely to point to the correct alternative than the incor-
rect alternative (the CJT’s standard Competence Assumption). We assume
that each judge is attempting purely to make a correct decision in the current
case (the CJT’s standard Sincerity Assumption that agents are not behaving
14 Assuming average competence within their court is the same as that within the initial court.
strategically).15 We further assume that, in doing so, each judge forms her
decision on the basis of (a) her own private signal and (b) the history of votes
in previous courts (which we assume to be common knowledge).
It is well known that any setup of this kind can easily give rise to informa-
tional cascades.16 That is to say, the history of votes can provide such strong
evidence in favour of one alternative that all subsequent judges will always fol-
low the evidence derived from that history, and never vote according to their
own private signals.17 If that happens, an informational cascade will have begun.
Thereafter, judges will have stopped learning from their own signals; instead
they will follow blindly the judgment suggested by the voting record of previous
courts. An informational cascade is problematic in that, once it begins, the
informational base on which all subsequent judgments are based can be very
thin, and all information from private signals is thereafter systematically ignored.
This is what Vermeule calls the ‘Burkean paradox’.18
One natural way to model sequential decision-making would be for a judge
to take all historic votes, add her own private signal as yet one more vote, and
then vote for whichever alternative has a majority among those pooled votes.19
(After all, if all judges are assumed to be equally competent, it would seem only
natural for a judge to count each of every other judge’s opinion equal to her
15 We assume she is doing so purely for the sake of the correct decision itself, although much
the same story would emerge if she were doing so to avoid acquiring a reputation as an unreliable
judge with a history of voting for the incorrect outcome (Ottaviani and Sørensen 2001).
16 The literature on informational cascades has grown rapidly and has by now resulted in all
sorts of technical refinements that we will not address in this chapter. Seminal papers include:
Bikhchandani, Hirshleifer, and Welch (1992, henceforth: BHW), Banerjee (1992) and Smith and
Sorensen (2000); a useful introduction is Chamley (2004); for applications to the courts see
Daugherty and Reinganum (1999) and Talley (1999). A very simple setup, roughly in line with the
model introduced by BHW, suffices to clarify the problem with ‘many minds’ arguments based on
sequential judgments. In our treatment, we will not always model judges as fully Bayesian rational,
unlike BHW and others. In particular, we want to maintain the possibility that judges can be
irrationally overconfident about their own private signal, or that they vote only according to their
private signal as a matter of principle. This assumption of ‘bounded rationality’ makes room for
the assumptions that may ultimately be more realistic than full Bayesian rationality for the phe-
nomena at hand. Of course this could be done in a Bayesian framework by changing the utility
functions of the judges: they might prefer a process in which they vote only according to their
private signal, so that they do not (only) care about being right. But the Bayesian treatment comes
with some algebraic costs and little gain for the purposes of this exploration.
17 Sunstein (2009, p. 105) writes, ‘Suppose that people engage in certain behavior or accept
certain beliefs solely on the ground that other people have engaged in that behavior or accepted
those beliefs. Once private information begins to emerge, it should defeat the cascade.’ But while
what Sunstein says is true of ‘pluralistic ignorance’, which is indeed fragile in that way (Kuran 1995),
it is not necessarily true of cascades, precisely because once the cascade has set in there are no
occasions for people’s private signals to emerge in the public record.
18 Vermeule 2009a, pp. 75–7.
19 This is the standard setup in the ‘informational cascade’ literature: see BHW, for example.
own.) But some judges might be more stubborn than that, weighing their own
views more heavily than others’.
Our model will therefore introduce one further parameter. (For the moment
we will assume this parameter to be identical for all judges, but we will relax
that assumption in Section 10.4.2.) This parameter reflects the weight, w, that
each judge gives to his own private signal relative to the votes of other previous
judges. Operationally, w is equal to the number of ‘votes’ that the judge allo-
cates to himself in reckoning which alternative wins after pooling all previous
votes with his own private signal-cum-vote.20
If w is large (larger than the number of individual decisions in the past) the
judge will always vote on the basis of her own private signal. Otherwise the
judge’s vote depends on the content of the judge’s private signal, the weight w
and the margin in the previous votes considered. If the margin is larger than w
then the judge’s own signal is overruled by past decisions; if it is not, the judge’s
signal is decisive.21
Suppose the competence of each judge (i.e. the reliability of her private signal)
is pc* = 0.55. Suppose that each judge weighs her own private signal equally
with every other previous judge’s vote (w = 1). Each court has nine judges,
deciding simultaneously.22 The results of a simulation employing those settings
for twenty sequential cases are shown in Figure 10.2.
What we see in Figure 10.2 is the effect of a classic cascade. The probability
that the current court’s decision is correct plateaus well under Pn = 0.7; and it
reaches that plateau after only a handful of previous decisions. After at the
most half a dozen previous decisions, subsequent judges are merely playing
‘follow-the-leader’ rather than revealing any independent judgement (private
signal) of their own through the process.23 Given the number of independent
judgments that the later courts could have relied on without the cascade, if not
following the precedent of earlier courts, the result is surely suboptimal from
an epistemic perspective.
20 This is one way of formalizing Vermeule’s (2009a, p. 76) suggestion that ‘individual judges
might adopt an intermediate approach, according to which they give some but not complete
deference to the views of the past, and correlatively think for themselves to some degree or in
some circumstances’. Gersen and Vermeule (2007) offer a model similar in spirit, but different in
operationalization, of how different judicial doctrines would lead different judges to give more or
less weight to the expertise of executive agency policymakers, compared to their own, when passing
judgment on administrative actions.
21 We assume that in case of a tie the judges vote on the basis of their private signal.
22 And ignoring any possible strategic incentives with regard to what the other judges on the
court might do.
23 In most cases a single round suffices to start the cascade. Only very rarely will the result be
so tight over several rounds that a cascade is delayed.

1.0
0.8
0.6
Pn
0.4
0.2
0.0
0 5 10 15 20
t
Figure 10.2 Probability of correct majority decision from a court with nine members,
each pc* = 0.55 likely to be individually correct, with weight w = 1.
10.4 SOLUTIONS
Now let us canvass some possible ways of averting the epistemic damage that
can thus come from judges being traditionalists, relying on precedent in decid-
ing how to vote on cases before them.
10.4.1 Hiding Precedents
One obvious way of preventing a cascade from occurring among traditionalist

judges who defer strongly to precedent would be to prevent them from discover-
ing the precedent. Suppose all cases were heard in closed session, all participants
were prevented from discussing the proceedings with anyone else, and no written
record of the proceedings is kept (or if kept, it is kept secret).24 Then, obviously,
however strongly predisposed a traditionalist judge may be to follow precedent,
he cannot do so because he has no way of knowing what the precedent is.
Completely preventing any communication of precedents might be difficult,
of course. Even before there were any extensive written records forming the
basis of English common law, ‘judgments . . . of common justice carr[ied] a con-
stancy, congruity and uniformity to one another’ as a result of ‘justices hav[ing]
24 Out-of-court settlements are often favoured by some parties to litigation precisely in order
to avoid having the case decided by the court and thus setting a precedent for future litigation.
had a common education in the study of the law . . ., sit[ting] near one another in
Westminster Hall, . . . and daily in term-time convers[ing] and consult[ing] with
one another’ there.25
Nevertheless, studies of the advancement of science suggest that less dense
networks of communication can help prevent premature closure. Suppose there
are some preliminary but erroneous results that seem to tell decisively in favour
of one particular hypothesis. Scientists embedded in a dense communication
network would all learn of those results promptly, and they might well all be led
to abandon exploration of alternative hypotheses in consequence. But if the
communication network is less dense, and some scientists do not learn of the
misleading results for some time, they will continue pursuing other hypotheses
in the meantime, and may in that way compile a body of evidence that will
expose the errors in the original misleading findings.26
Similar advantages might obtain from reducing the scope for communica-
tion of precedents among precedent-following traditionalist judges. If they
have not heard about any precedent for the case before them, they will have
to adjudicate it on its own merits—thus offering an ‘alternative possible pre-
cedent’ to the already-existing one once they come to hear of it. And subsequent
precedent-respecting traditionalist judges will then have two alternative
precedents between which to choose, and they will have to exercise their
own independent judgement as to which is the better of the two, especially if
they do not know how many previous judges have supported the one or
the other precedent. Assuming there are a lot of judges independently and
simultaneously choosing between those options (and assuming the other
assumptions of the CJT are met), the better of the two is likely to be chosen
by a majority of those judges, and thus emerge as the preferred precedent for
future adjudication.
10.4.2 Resisting Precedent: Stubborn Judges
Now consider the case of ‘stubborn’ judges, defined as judges who do not accord
much weight to precedents as compared to their own private signals in decid-
ing how to vote.
Increasing the weight that each judge gives to his own private signal relative
to previous judges’ votes has two effects. First, it delays the onset of a cascade.
Second, in the long term it increases the probability of the current court’s deci-
sion being correct as long as the judges do not give complete weight to their
own private signal and eventually succumb to the weight of accumulated past
25 Matthew Hale, quoted in Simpson (1973, p. 96).

26 Zollman 2010. Grim et al. 2013.

1.0
0.8
0.6
Pn
0.4 ω=3
ω = 10
0.2 ω = 50
0.0
0 50 100 150 200
t
Figure 10.3 Probability of correct majority decision from a court with nine members,
each pc* = 0.55 likely to be individually correct.
evidence.27 Both of these tendencies are shown clearly in Figure 10.3, which

displays the results of simulations for w = 3, w = 10, and w = 50.
These findings admit of easy explanation. Being ‘stubborn’ in the sense of
putting great weight w on your own private signal has two effects. First, when
there are not that many previous judgments available, stubborn judges are not
at all swayed by the (limited) evidence from the past and vote according to their
own signal, thereby accumulating evidence to be used for future judges. Second,
once the evidence has accumulated it will eventually outweigh the importance
stubborn judges give to their own signals. At that point, courts of stubborn
judges start benefiting from the informative judgments of the past. After a cer-
tain point, reached very soon by judges who treat others’ judgments as even
roughly on a par with their own (e.g. w = 3), an informational cascade of the
same sort as before arises. Then instead of ‘many minds’ we merely have ‘many
mimics’ who confer little epistemic advantage—but since that cascade was
delayed, the courts do better with higher w.
The upshot of this analysis is clear. The only way in which courts composed
of homogeneous judges will be able to achieve any substantial epistemic success
is by judges attaching very little importance to the judgments of previous
judges, relative to their own. That is to say, judges would have stubbornly to
stick with their own views in the face of a very substantial body of traditional
27 Although in the short term it might actually reduce that probability, compared to judges
who attach relatively less weight to their own private signals, compared to precedents. This is seen
in Figure 10.3, in the first fifty sets of cases.
evidence in the opposite direction. In short, in this scenario, epistemic success

would require judges largely to resist tradition rather than bowing to it.
10.4.3 Discerning Traditionalists: Picking

Informative Precedents
In the previous setup, we assumed that judges can only ascertain how previous
judges voted, not why. Specifically, that setup assumes that they cannot deter-
mine whether a previous judge has voted on the basis of her private signal or on
the basis of the history of votes of prior judges.
That is not a very realistic assumption. After all, judges can and almost
invariably do read the opinions of previous judges.28 From that, they can usually
(albeit imperfectly) surmise the extent to which any given previous judge was
following her own judgement or that of prior courts. Let us now vary our setup
to reflect that fact.
For this setup, we hypothesize a heterogeneous court.29 Some judges on that
court always reveal their private signals. Other judges on that court (indeed, we
hypothesize, a majority of judges on the court) make their votes as in the previ-
ous setup, by pooling their private signal with the votes of judges on previous
panels.30 But—and this is the second crucial difference between this setup and
the last—we assume that in so pooling current judges take into account the
votes only of judges on previous panels who voted on the basis of their own
private signals and not on the basis of tradition. That is to say, these judges vote
on the basis of previous judges’ votes only when those votes are truly ‘inform-
ative’ and not merely the product of an informational cascade.
Figure 10.4 provides a diagrammatic representation of that scenario in its
most stark form.31 In that figure Votes 3 and 4 represent traditionalist judges.
Those judges base their own decisions purely upon the decisions of previous
judges (Votes 1 and 2 in the case of Vote 3; Votes 1, 2, and 3 in the case of Vote 4),
28 Talley (1999, pp. 107–10) discusses how this will help overcome informational cascades in
the courts.
29 Courts cannot be composed entirely of ‘discriminating traditionalists’ of the sort we describe
here, because ‘discriminating traditionalists’ require the existence of independent judges stub-
bornly exercising their own judgement to discriminate in favour of. But on the present account,
nonconformism on the bench is of value just because there are others on the bench (‘discriminating
traditionalists’) who recognize the nonconformists and ignore the opinions of the rest.
30 Vermeule (2009a, p. 76) anticipates this part of our model, but not the next, when writing,
‘Perhaps some judges in the stream of precedent or tradition have contributed independently,
while some have not.’
31 ‘Most stark’, because (not having any causal arrows from Evidence to the Votes of any
traditionalist judges, i.e. Votes 3 and 4), this diagram implicitly assumes tradition-following judges
follow it slavishly and assign w = 0 weight to their own private signals. In the numerical exercise
that follows we will relax that assumption. But we retain it in the causal diagram to make difference
between traditionalists and discerning traditionalists stand out as starkly as possible.
State
Evidence
Vote 2 Vote 1
Vote 3
Vote 4 Vote 5
Figure 10.4 Discerning traditionalists.
without independently assessing of the evidence for themselves. Vote 5 represents

what we call a ‘discerning traditionalist’. Vote 5 is not itself directly influenced
by the Evidence, as are Votes 1 and 2. But Vote 5 is at least influenced only by
other Votes that are themselves directly influenced by the Evidence. And that
can make a big epistemic difference, especially if one were to extend Figure 10.4
to include many more votes. Following judges who have no independent access
to evidence is largely pointless, and can be severely misleading.
Here is a numerical example to illustrate how much difference discerning
traditionalists might make. For this setup, we assume courts consisting of four
maximally ‘stubborn’ judges who always vote exclusively on the basis of their
own private signal and five judges who give the same weight (w = 1) to their own
private signal as to all such previous informative votes when deciding how to
vote. Figure 10.5 shows the epistemic performance of courts of this sort. As we
see there, the probability that such a court will reach the correct decision does
not plateau in this case. Instead, it continues to increase the more previous courts
there have been.
That probability approaches 1 slowly, to be sure. As Figure 10.5 shows, after
taking into account fifty previous courts the probability of the majority of the
current court reaching the correct decision is around Pn ≈ 0.90, and after one
hundred around Pn ≈ 0.95. By the time 200 previous courts have been taken
into account, a correct decision is virtually certain. And with courts that are
heterogeneous in this way, the probability of a correct decision is a much more
rapidly increasing function of the number of previous courts than it is with
homogeneous courts with very stubborn judges: the curve representing very
stubborn judges (w = 50 in Figure 10.3) does not catch up with that represent-
ing discerning traditionalists (in Figure 10.5) until there have been almost a

1.0
0.8
0.6
Pn
0.4
0.2
0.0
0 50 100 150 200
t
Figure 10.5 Probability of correct majority decision from a heterogeneous nine-member

court (each judge pc* = 0.55 likely to be individually correct), where four judges vote on
the basis of their private signal and the rest decide by weighting on the basis of their
private signal w = 1 and only take into account informative votes.
hundred previous courts. In addition, the heterogeneous courts converge to 1,

while courts of homogeneously stubborn judges plateau at a level close to 1 but
do not converge to 1.
This, then, is a second way a revisionary traditionalist approach might succeed
epistemically. Judges on heterogeneous courts can improve their chances of
being correct by taking into account the decisions of previous judges, provided
they do so in this very particular way. But note well the irony. Traditionalist
courts of this sort benefit epistemically only from judges taking account of the
votes of previous judges who were not themselves traditionalists, and who voted
purely on the basis of their own private signal rather than on the basis of the
history of votes before them.
10.4.4 Lots of Precedents to Choose Among
Finally, and more impressionistically, we consider another variation on the het-

erogeneous court setting. Rather than one group of judges being simply ‘stub-
born’ when following their own private signal, let us now assume that there are
simply a lot of different precedents for the judges to choose among.32 If there are
32 Higher courts are led to hear a case by conflicting rulings among lower courts for example.
multiple precedents available which point towards different outcomes in the case
at hand, then such judges exercise their own judgement and use their own private
signals in picking among them and deciding the case.33
‘Following precedent’ would of course then be a mere pretence—the judges
fulfill the same function as the ‘stubborn’ judges. In the scenario just outlined,
what the judge would really be following is his own private signal. The precedent
to which he appeals in support of his decision would be pure pretext, rather
than exerting any independent power over his decision. Such judges would be
traditionalists in their words but not their deeds: they would be appealing to
precedent, without actually following it.
But let us imagine a less duplicitous judge. Suppose he is genuinely sensitive
to precedent, in the sense that if no precedent can be found for deciding the case
in such-and-such a way (or if literally all precedents militate against doing so),
then he would not decide the case in such-and-such a way. But suppose that, as
it happens, the judge is faced with a large set of possible precedents which are
more-or-less applicable to the case at hand many of which point towards differ-
ent ways of deciding the case at hand. And suppose that in choosing which pre-
cedent to follow, the judge lets that choice be determined at least in part by what
his private signal tells him is the correct resolution of the case at hand.
There, again, the force of the threatened cascade of precedent would be blunted.
There, again, the judge would be exercising his own independent judgement
and acting in large measure on his own private information about the correct
outcome in the case at hand. But there again, that outcome is achieved by
weakening the grip of precedent.
As in the previous section, there is a second type of judge required: when one
group of judges reveal their private signals, another group must aggregate all
these private signals, vote accordingly and win most of the time to make use of
all that information. But a combination of precedent-selecting judges (who
effectively reveal their private signal) and faithfully aggregating judges can lead
to results similar to those shown in Figure 10.5.
All three of the ways of solving the problem of traditionalism inducing a
cascade that undermines the epistemic power of decision-makers thus achieve
that result in basically the same way: by abandoning or weakening our defer-
ence to tradition.
10.5 BEYOND TRADITIONALISM IN THE COURTS
We have taken the courts as the primary focus for this chapter’s discussion of
traditionalism for several reasons. One is that the courts are where it is most
33 Schauer 1987. Some of the empirical evidence suggests that judges do precisely this (Segal
and Spaeth 1996).
commonly manifested in contemporary political practice, through the doctrine

of stare decisis. Another is that because courts confront broadly the same issues
in successive cases, this gives the notion of ‘respecting tradition’ and ‘following
precedent’ a more determinant meaning there than in many other realms. And
of course judges are supposed to be ‘judging correctly’, according to some external
(albeit hard to access) standards of correctness, rather than merely exercising
their preferences.
In closing our discussion in this chapter, we ought also to point out that the
institutional structure of courts makes them the ‘best case’, when it comes to the
epistemic merits of traditionalism. The traditionalist thought, when it comes to
the courts, is that any given panel of judges (we have been talking in terms of
nine-member panels, here) has within itself limited epistemic resources; and
that by following precedent, judges, instead of simply acting on their own private
judgement, are able to draw on a much wider array of epistemic resources.
We have, over the course of this chapter, shown various ways in which that
thought is problematic, even when it comes to small panels of judges. But note
that it is a thought that has only limited plausibility, even where (such as in
courts) the group of current decision makers is small relative to the number of
past decision makers on whom they can draw. If the current decision-making
group is itself large—numbering just in the hundreds found in legislatures,
much less the millions found in electorates—there is little reason at all to
suppose that today’s hundreds or millions would do epistemically better by
deferring to yesterday’s many more.34
In short, if traditionalism does not have much to recommend it epistemically—
and we have shown that it does not—even when it comes to the small-scale
setting of the courts, it will have even less to recommend it when it comes to
other arenas of political decision.
34 True, there will typically be more hundreds or millions of past legislators or voters than
found in today’s legislature or electorate. But remember, in the classic CJT group competence rises
very quickly with increasing numbers, so those extra hundreds and (all the more so) millions will
bring only vanishingly small epistemic gains.
11
Following Leaders
‘The function of leadership is to lead’, we are often told—often by leaders who

are themselves pretty unsavoury. Still, leadership is a ubiquitous and important
feature of social and political life.1 Many of the most fundamental questions
in normative and positive political theory are intertwined with the concept of
leadership: authority, legitimacy, power, influence, delegation, and so on.
From the perspective of democratic theory, we can draw a distinction con-
cerning popular sovereignty parallel to one that (radical) economists draw
concerning consumer sovereignty.2 Sometimes suppliers (firms, political lead-
ers) strive to satisfy the population’s preferences, taking the population’s exist-
ing preferences as given. That is the neoclassical model of the market economy
and the Schumpeterian model of liberal democracy.3 But other times suppliers
(firms, political leaders) strive to shape the population’s preferences, changing
these preferences and bringing them more into line with what those suppliers
are selling. Within economics, that is the Vance Packard model of The Hidden
Persuaders and John Kenneth Galbraith’s model of ‘demand management’;
within politics, it is Steven Lukes’s ‘third face of power’.4
In practice, of course, it may not always be easy to determine who is follow-
ing whom. Are citizens following their leaders? Or are leaders following their
citizens? Or might the concurrence in beliefs arise from the fact that, in choos-
ing their leaders, citizens choose people with beliefs similar to their own? No
doubt some of each is going on. Some evidence suggests that the latter two
processes are stronger than the first.5 So the jury is still out on the empirics, and
it may well be that we have less to fear from leaders manipulating the public’s
views than might be supposed.
1 Political leadership is the subject of a huge 784-page Oxford handbook (Rhodes and Hart
2014). For the most sophisticated recent treatments of it within political theory see White and Ypi
(2011; 2016) and Beerbohm (2015).
2 Gintis 1972a, b. 3 Schumpeter 1942.
4 Packard 1957/1981. Galbraith 1967/1972. Lukes 1974/2005.
5 On the former point see Verba and Nie (1972, ch. 18 and esp. pp. 331–2); on the latter, Miller
and Stokes (1963).
Following Leaders 165
Still, in the context of this book we need to explore just how political leaders
(or opinion leaders more generally) might pose a threat to the operation of the
CJT, if they are indeed a powerful force in shaping voters’ views. Since our
concern here is purely the epistemic quality of democratic processes, our focus
will be purely upon the impact that following leaders may have on the epi
stemic content of people’s votes. That is to say, we will here be concerned with
opinion leadership only in the realm of beliefs, not desires.6
One of the crucial assumptions of the CJT, recall, is that each vote is inde-
pendent of every other. If the voters are subject to some common cause that
leads all of them to vote as they do, then that Independence Assumption is
violated. According to the conventional wisdom, one sort of common influence
that is thought to pose particularly grievous risks for the CJT is the existence of
‘opinion leaders’ who dictate the votes of many voters at one and the same time
and in the same direction.7 Just how correct is that conventional wisdom?
Answering that question is the aim of this chapter.
11.1 A SINGLE OPINION LEADER
Let us first explore the epistemic consequences of having a single opinion

leader.8 When all voters follow the same opinion leader, that constitutes a vio-
lation of the CJT’s Independence Assumption. In terms of the causal diagrams
in Chapter 5, the situation is then as represented in Figure 11.1.
When we say ‘OL is an opinion leader with respect to voter V’ we mean that,
once OL has announced his position, then voter V will (with some ‘probability
of following’, which we will denote π) adopt the same position as OL without
voter V considering her own private signal. In Figure 11.1(a) the only causal
path from the Evidence to the Votes runs via the opinion leader OL. This means
that the only source of information for voters is what they learn from the opin-
ion leader—if the OL provides false or misleading information, their only
options are either (at best) ignore him and cast a random vote or else (worse)
follow him. In Figure 11.1(b), by contrast, the Votes are also influenced by the
Evidence directly, so that the voters can be individually competent without
relying on the opinion leader.
6 ‘Preferences’ are typically underpinned by both, of course. Hence we will not speak further in
those terms over the rest of this chapter (even if that is the way the literature to which our opening
discussion alludes has traditionally been cast).
7 Sunstein (2009, pp. 125, 169, 171–3, 175) identifies this as the greatest epistemic worry
with what he calls the Populist model. See also: Estlund 1994; Ladha 1992; Boland 1989; Boland,
et al. 1989.
8 See Boland et al. (1989) for a technical treatment of this setting.
(a) State (b) State
Evidence Evidence
OL OL
Figure 11.1 Multiple votes influenced by the same Opinion Leader, either without
(a) or with (b) direct influence of Evidence on Votes.
For a numerical example of this latter setting, imagine a set of voters, each of
whom is individually pc* = 0.55 competent when not following the opinion
leader, and each of whom is attached to an opinion leader who is herself also
pOL = 0.55 competent. Figure 11.2 displays curves representing the probability
that a majority among those voters will be correct, for various values of π (the
probability of each voter adopting the position of the opinion leader without
considering his own signal) and various values of n (the number of voters).
Figure 11.2 shows that with suitably low probabilities-of-following, the basic
CJT effect still occurs.9 It just occurs a little more slowly than without any fol-
lowing at all. In Figure 11.2, for example, the curve for π = 0.05 is just a little
lower than for π = 0.
Notice that a similar result would emerge if the opinion leader were gener-
ally incompetent. Suppose pOL = 0.40, for example, with all else remaining
the same. Then the curves for high probabilities-of-following would, with
increasing numbers of voters (n), asymptotically approach 0.40 from above,
as in Figure 11.3. But for low probabilities-of-following, such as π = 0.05, the
curve would still asymptotically approach 1. In short, no great epistemic dam-
age is done even if voters follow bad leaders, just so long as they do not do so
too slavishly.
The crucial cutting point is, as we have said in Section 4.5.3, π = (pc* − 0.5)/pc*.
In the scenario represented in Figures 11.2 and 11.3, that is just over π = 0.09. If
everyone follows the same opinion leader less than that, then the standard CJT
result obtains: a majority of them will be more likely to be correct than any one
of them would be individually, and that probability will asymptotically
approach 1.0 as the number of voters approaches infinity. That will happen
9 As already discussed in Section 4.5.3.


1.0
0.8
0.6
Pn
0.4 π=0
π = 0.05
0.2 π = 0.1
π = 0.2
0.0
0 100 200 300 400 500
n
Figure 11.2 Probability of correct majority decision among voters with individual
competence pc* = 0.55, given a single opinion leader of competence pOL = 0.55 followed
with probability π.
more slowly the nearer π is to that crucial cutting point, to be sure—but as long
as π is below that cutting point that will still happen. (For an example of that,
consider the case of π = 0.05 in Figures 11.2 and 11.3.) Where the probability of
voters following an opinion leader is higher than that crucial cutting point,
however, the probability that the majority among them is correct asymptotically
approaches instead the probability that the opinion leader herself is correct.
(For an example of that, consider the case of π = 0.2 in Figures 11.2 and 11.3.)
Just how much does following an opinion leader slow down the emergence
of the basic CJT result? Here is a numerical example, based on the Figure 11.2
scenario. If voters there do not follow an opinion leader at all (π = 0), it takes
only around 180 members for the group to attain collective competence (the
likelihood of a majority of then being correct) of Pn = 0.9. If each voter follows
the opinion leader with probability π = 0.05, the group can still achieve that
same collective competence level, but doing so will require something like
double that number of voters.
What happens if voters’ probability of following the opinion leader exceeds
the crucial cutting point? Suppose in the Figure 11.2 scenario they exceed it by
only just a little—following the opinion leader just ten per cent of the time
(π = 0.10), for example. Then the probability of the majority among them being
correct peaks just under Pn = 0.75 with 171 voters. At that point, it begins declin-
ing, asymptotically approaching Pn = 0.55, (corresponding to the probability
pOL = 0.55 that the opinion leader herself is correct) as the number of voters

1.0
0.8
0.6
Pn
0.4
π=0
π = 0.05
0.2 π = 0.1
π = 0.2
0.0
0 100 200 300 400 500
n
Figure 11.3 Probability of correct majority decision among voters with individual
competence pc* = 0.55, given a single opinion leader of competence pOL = 0.4 followed
with probability π.
thus following the opinion leader increases.10 Something similar happens, of

course, if the opinion leader is incompetent, as in Figure 11.3, but the decline
will bottom out at 0.4, which is the opinion leader’s competence in that case.
That last-described result indicates not only why we should worry, but also
just how much (or how little) we should worry, about the effects of a single
opinion leader on the epistemic performance of electorates. If voters follow
a single opinion leader too faithfully, the epistemic power of ‘many minds’ con-
tracts quickly to the epistemic power of the single mind of the single opinion
leader. But just how worried we should be depends, of course, on just how
competent that single opinion leader herself actually is. If the opinion leader is
pOL = 0.55 likely to be right, then the majority of a group following her faithfully
(π = 1) is Pn = 0.55 likely to be right.
11.2 MULTIPLE CORRELATED OPINION LEADERS,

BUT SOME INDEPENDENT VOTERS
Just how realistic is it that there will be only one single opinion leader? That
may happen in cases of really successful totalitarian regimes. Maybe it was the
case (or maybe not) in Italy when Berlusconi controlled virtually all mass
10 Which is just a special case of the Best Responder Corollary in Section 5.3.
media outlets.11 But in most places most of the time there are multiple opinion
leaders. What difference might this make to the epistemic competence of
electoral majorities? That is the issue to which we next turn.
We begin, in this section, by discussing cases of two opinion leaders that
are correlated with one another, either positively (Section 11.2.1) or negatively
(Section 11.2.2). We then proceed, in the next section (Section 11.3), to discuss
the epistemic effects of having many opinion leaders each of whom is inde-
pendent of every other, conditional on the state of the world and the evidence.
11.2.1 Positively Correlated Opinion Leaders
Having multiple opinion leaders instead of just one will not make any difference,
of course, if they are all merely clones of one another. For a variation on the
previous scenario along those lines, suppose that there are two opinion leaders,
each of whom is followed by one half of the electorate with probability π = 1.0.
But suppose the second opinion leader follows the first one with probability
π = 1.0, in turn. Then the epistemic consequences would be no different from
having a single opinion leader who is followed with probability of π = 1.0 by the
whole electorate. Even with a very large number of voters, the probability that
a majority of them is correct will just be the same as the probability that the first
opinion leader is.
Figure 11.4 varies that basic scenario by introducing, for more realism, a
third group of voters who follow neither opinion leader. Those independent
voters are going to become important in our analysis in the next section. But
in the context of positively correlated opinion leaders, they do not make
much difference. Suppose that in the Figure 11.4 scenario Votes 1 and 2 f ollow
OL1 with probability of π = 1.0; Votes 3 and 4 follow OL2 with probability of
π = 1.0; and OL2 follows OL1 with probability of π = 1.0. Then there will
always be a majority of four votes in favour of whatever OL1 says, and the
probability that that majority is correct is once again just the probability that
OL1 is correct.
We can loosen up the model somewhat by setting the probabilities of follow-
ing at something less than π = 1.0 and giving voters some better-than-random
probability of voting correctly when they do vote independently of the influ-
ence of any opinion leader. But the results are still going to be qualitatively
pretty similar to those already shown in Figure 11.2.
11 Habermas 2006. But as we have said in the introduction to this chapter, to be successful,
politicians and businessmen (perhaps even Berlusconi) need to be attentive to public opinion at
the same time as they are shaping it.
State
Evidence
(+)
OL 1 OL 2
Figure 11.4 Perfectly positively correlated Opinion Leaders and some independent Votes.
11.2.2 Polarization: Negatively Correlated Opinion Leaders
Things are very different, however, in the case of multiple opinion leaders who
are negatively correlated with one another. In Section 7.2 we discussed one way
in which negative correlation among voters might improve a group’s epistemic
performance. Here we shall show how negative correlation among opinion
leaders can, at least under certain circumstances, do likewise.
For the limiting case, imagine a highly polarized polity, with two opinion
leaders who are diametrically opposed to one another. This setup resonates
with recent US history, of course. But it also captures the long-standing con-
ception of the role of the loyal opposition in the UK, where it is standardly said
that ‘the duty of an opposition is to oppose’.12
Figure 11.5 depicts the case where there are two opinion leaders, one of whom
takes exactly the opposite position on any given issue to the other (as indicated
by the minus sign over the arrow connecting them).13 Some voters strictly fol-
low opinion leader OL1, some voters strictly follow the other opinion leader
OL2, and some voters vote independently of any opinion leader and base their
votes on the Evidence alone.
In his treatise, The American Commonwealth, Lord Bryce speculated that
that sort of situation might have happy epistemic consequences. He wrote:
[T]he educated and reflective class in America . . . may be numerically a small
minority of the voters, but as in many states the two regular parties command a
12 It is a slogan Winston Churchill claimed to have learnt from his father. Safire (2008, pp. 403–4)
also reports Fred Vinson, the mid-twentieth-century Chief Justice of the US Supreme Court,
recalling ‘the story of a Kentucky politician who was asked whom he was going to support in a
primary election. The politician’s answer, “I don’t know yet. I’m waiting to see what the opposition
does, so I can take the other side.” ’
13 Obviously, in a binary choice situation that necessarily means that one of the opinion leaders
is correct, and the other is incorrect.
State
Evidence
(–)
OL 1 OL 2
Figure 11.5 Negatively correlated Opinion Leaders and independent Votes.
nearly equal normal voting strength, a small section detached from either party
can turn an election by throwing its vote for the candidate, to whichever party he
belongs, whom it thinks capable and honest. Thus a comparatively independent
group wields a power in elections altogether disproportionate to its numbers.14
Let us explore Bryce’s speculation a little more formally through the following
model. Suppose the electorate is divided into three equal-sized groups. One-
third of the electorate follows opinion leader OL1 with certainty (π = 1). Another
third follows the other opinion leader OL2 with certainty (π = 1). The remain-
ing third of the electorate does not follow either opinion leader, and each of
those independent voters is individually pc = 0.55 likely to be correct when
casting his vote.
In that sort of scenario, it is clear what would happen. If the two opinion
leaders lead their followers to vote in diametrically opposed directions, and
each leader commands an equal-sized following, the votes of those two groups
of voters simply cancel each other out. The election is then decided by the votes
of the remaining independent voters. In the scenario sketched above, the prob-
ability of the correct option winning a majority of votes among the whole elect-
orate consisting of n voters is simply the probability of the correct option
winning a majority of votes among the n/3 independent voters.15 If the whole
electorate consists of n = 999 voters, then following the standard CJT calcula-
tion the probability of the majority vote being correct on the present scenario
is Pn = 0.966.
In short, the effect of diametrically opposed opinion leadership, where each
leader has an equal number of equally faithful followers, is simply to reduce the
14 Bryce 1888, vol. 3, p. 119, quoted in Vermeule 2011b, pp. 20–1.

15 Of course, the size of the majority for the correct option would be larger than that, since it is
boosted by the n/3 voters who followed the opinion leader who led them in the correct
direction.
‘epistemically effective’ size of the electorate. But recall that, in the classic CJT,
the probability of the correct option winning a majority is a rapidly increasing
function of the number of voters. So in any moderately large electorate, with
even just moderately numerous independent voters who are even just moder-
ately competent, the probability of the majority being correct even in a pretty
polarized polity is still really pretty high.
Just how heavily does this outcome depend on the two diametrically opposed
opinion leaders having roughly equal numbers of equally faithful followers? If
they do, then we can clearly count on their influence cancelling each other out.
But the numbers of followers need not be literally equal. What matters for the
cancelling effect we here associate with negatively correlated opinion leaders is
the expected margin. If the expected margin created by the opposing opinion
leaders is smaller than the expected margin of correct to incorrect votes in the
group of independent voters, then the correct votes of the independent voters
are not expected to be overturned by the opinion leaders’ influence. That effect
becomes more reliable as the group grows large (while holding the propor-
tional influence of the opinion leaders fixed). In that case the group becomes
more competent as its size increases.
Similarly, we can weaken the assumption that voters who follow opinion leaders
follow them with certainty (π = 1.0) and the same sort of cancelling may still occur.
Given a large electorate, that is still likely to occur if for example a third of the
electorate follows opinion leader OL1 with probability π = 0.5 and another third
of the electorate follows diametrically opposed opinion leader OL2 with that same
probability, while the remaining third consists of competent independent voters
as before. And differential rates of following can of course compensate for differen-
tial sizes of followerships for different opinion leaders.16
11.3 EVERYONE PARTIALLY FOLLOWS

UNCORRELATED OPINION LEADERS
The results reported in Section 11.2 all assume there to be a certain number of

people who always vote independently of the influence of any opinion leader.
Now let us drop that assumption, and assume instead that every voter follows
some opinion leader, to some extent but not perfectly (0 < π < 1).
Let us further assume for the purposes of the subsequent analyses that
the opinion leaders are wholly independent of one another, conditional on the
state and the evidence.17
16 All these considerations require a very good understanding of the causal factors influencing
voters—an understanding unlikely to be present in real-life settings. The point of these explor-
ations is therefore not to predict but to show what could happen in polarized societies.
17 In the terminology of Section 5.2.
11.3.1 Opinion Leaders of Purely Random Competence
For an initial model along those lines, let us assume that each of those opinion
leaders is of purely random competence (so for every opinion leader, pOL = 0.50).
And, purely for convenience, let us revert to our original assumption that each
opinion leader has the same number of voters following him, and each voter
follows her respective opinion leader with the same probability π.
Table 11.1 offers some sample calculations to show what might be expected to
happen on this scenario. For the purposes of those calculations, we assume an
electorate of 1,000 voters, who follow between one and twenty opinion leaders
to varying extents and are pc* individually competent when not doing so.
From Table 11.1 we can infer it does not much matter epistemically even if
everyone follows some opinion leader or another, even if all of the opinion
leaders are of purely random competence, just so long as people do not follow
their opinion leader too closely and they are individually of better-than-random
competence when they do not. If, for example, the rate of following is π = 0.10,
the probability that the majority will be correct exceeds Pn > 0.90 whenever
there are three or more opinion leaders.
What is driving these results is, of course, simply the fact that something
akin to independence is here being secured in other ways than before. In
the previous section’s models, independence was achieved by having voters
who vote independently of any opinion leader on all occasions. In this
section’s models, where the rate of following π is low, we have many voters
who vote independently of their opinion leader on a great many (if not all)
occasions.
Furthermore, observe in Table 11.1 that increasing the number of independ-
ent opinion leaders can compensate for higher rates of followership. For example,
the probability of the majority being correct is about the same (around two-thirds)
if everyone follows a single opinion leader π = 0.1 of the time or if people follow
one or another of twenty opinion leaders π = 0.5 of the time.
Table 11.1 Probability of correct majority decision among 1,000 voters with individual
competence pc* = 0.55 split evenly among multiple opinion leaders (pOL = 0.50), each voter
following his respective opinion leader with probability π.
Opinion leaders Group partition π (probability of followers adopting position of their
opinion leader)
0.1 0.2 0.3 0.5 0.75

1 1,000 0.69 0.50 0.50 0.50 0.50
3 333, 333, 334 0.91 0.75 0.56 0.50 0.50
5 5 × 200 0.95 0.80 0.69 0.51 0.50
10 10 × 100 0.98 0.87 0.75 0.62 0.60
20 20 × 50 0.99 0.93 0.83 0.66 0.57
11.3.2 Competent Opinion Leaders
Next suppose that there are multiple opinion leaders, but this time let us assume
that each of them is somewhat more likely to be right than wrong.
For ease of comparison, let us retain the same basic parameters as for the
Table 11.1 scenario. Specifically, assume there are 1,000 voters, equally split
among varying numbers of opinion leaders. Every voter follows some opinion
leader, and does so with probability π. The only difference between this and the
Table 11.1 scenario is that we shall now be assuming that each opinion leader is
not only independent of every other18 but also is of better than random compe-
tence, which we shall assume to be pOL = 0.55 for each of them.
Comparing the corresponding cells in Tables 11.1 and 11.2, we can surmise
the magnitude of epistemic gain that might be expected to come from making
opinion leaders themselves marginally competent. The short answer is ‘some,
but not much’. Take the case of five opinion leaders and probability of following
of π = 0.3. If opinion leaders are random (as in Table 11.1) the probability that
the majority among voters will be correct is Pn = 0.69. If opinion leaders are
pOL = 0.55 competent (as in Table 11.2), that goes up Pn = 0.77.
These results, however, depend heavily upon a presumption of symmetry in
the size of each leader’s potential followership. This is shown in Table 11.3. We
assume, once again, that each opinion leader is pOL = 0.55 competent and that
all opinion leaders are independent of one another in the relevant respects. In
the top row of Table 11.3, we assume that there are equal numbers of voters
following each of three opinion leaders with the same probability. In the b ottom
row, we assume that the number of followers is grossly unequal; specifically,
that 800 voters follow one opinion leader and 100 follow each of the other two
opinion leaders.
Table 11.2 Probability of correct majority decision among 1,000 voters with individual
competence pc* = 0.55 split evenly among multiple opinion leaders (pOL = 0.55), each
voter following his respective opinion leader with probability π.
opinion leader)
0.1 0.2 0.3 0.5 0.75

1 1,000 0.72 0.55 0.55 0.55 0.55
3 333, 333, 334 0.93 0.80 0.63 0.58 0.57
5 5 × 200 0.96 0.85 0.77 0.60 0.59
10 10 × 100 0.99 0.92 0.84 0.74 0.71
20 20 × 50 ≈1.00 0.97 0.91 0.81 0.73
18 Conditional on the state of the world.

Table 11.3 Probability that a majority of voters will be correct if they follow, to varying
degrees, three opinion leaders each with competence pOL = 0.55, uninfluenced voters
being competent with pc* = 0.55.
opinion leader)
0.1 0.2 0.3 0.5 0.75

3 333, 333, 334 0.93 0.80 0.63 0.58 0.57
3 800, 100, 100 0.84 0.56 0.55 0.55 0.55
As we see in Table 11.3, the probability of the majority being correct declines

in both cases with increases in the probability of voters following their opinion
leader (π). And as we see comparing the top and bottom rows of Table 11.3,
asymmetry in the size of the followerships does not make too much difference
if the probability of voters following their respective opinion leader is very low
or very high. But if voters are likely to follow their leader with intermediate
probabilities, then the epistemic competence of the majority can be seriously
compromised where the leaders’ respective potential followerships are of
appreciably different sizes.19 Consider the case of the probability of following of
π = 0.2, for example. The probability that the majority of voters will be correct
is Pn = 0.80 if voters are split equally among all three opinion leaders, but it is
only Pn = 0.56 if they are split unequally as in the bottom row of Table 11.3.
11.4 MANY MULTIPLY MEDIATED OPINION LEADERS
We tend to think of ‘opinion leaders’ as being relatively few and far between. In
ancient times, there were rarely more than a handful of demagogues in any
given assembly. In modern times, there are rarely more than a handful of mag-
nates controlling any given media market.
But there is another more common or garden variety of opinion leadership
that is much more widely dispersed throughout the community. That was the
variety that Paul Lazarsfeld and his colleagues had in mind when introducing
the concept of ‘opinion leaders’ into modern political science in connection
with their Columbia election studies dating back to the 1940s.
Lazarsfeld, Berelson, and Gaudet defined ‘opinion leaders’ as people who
answered ‘yes’ to both of the following two questions: ‘Have you tried to con-
vince anyone of your political ideas recently?’, and ‘Has anyone asked your
advice on a political question recently?’. On that definition, fully 21 per cent of
19 A similar result would obviously obtain if the probabilities of following differed substan-
tially between different leaders, as well.
State
Evidence
BOL 1 BOL 2 BOL 3
LOL 1 LOL 2 LOL 3 ... ... ...
Vote 1 Vote 2 Vote 3 ...
Figure 11.6 Many multiply mediated Opinion Leaders, direct links between Evidence
and LOLs and Evidence and Votes are omitted.
their sample counted as ‘opinion leaders’.20 Instead of Fox News, think of those
sorts of low-level opinion leaders as Uncle Fred.
Common or garden opinion leaders of that sort are not only more numerous
than the other ‘bigger’ sort. They also stand as intermediaries between the ‘Big
Opinion Leaders’ and the voters who look to their own ‘Local Opinion Leader’
for advice. What might be the epistemic impact of their intermediation?
To get a sense of that, let the polity be structured as follows. Suppose there
are 990 voters, thirty Local Opinion Leaders, and three Big Opinion Leaders.
Suppose each Local Opinion Leader has thirty-three voters who follow him
with probability π (the same for all voters and all Local Opinion Leaders).
Suppose that each Big Opinion Leader has ten Little Opinion Leaders who fol-
low him with that same probability π (again, the same for all Local Opinion
Leaders and all Big Opinion Leaders). This constellation is represented, sche-
matically in truncated form, in Figure 11.6.
As seen from the following Table 11.4, this new model, inserting lots of
Local Opinion Leaders (‘Uncle Freds’) between voters and the Big Opinion
Leaders (‘Fox News’), improves the epistemic performance of the electorate
markedly, at least for low to middling rates of followership. In that mediated
model of opinion leadership, the majority can expect to get it correct fully
72 per cent of the time even if voters follow their Local Opinion Leader half
the time (and Local Opinion Leaders their Big Opinion Leader half the time).
In the unmediated model where there are only Big Opinion Leaders who voters
follow half the time, the majority can only expect to be correct 58 per cent of
20 Lazarsfeld, Berelson, and Gaudet 1944, pp. 49–51. See further: Berelson, Lazarsfeld, and
McPhee 1954, pp. 109–15; and Katz and Lazarsfeld 1955.
Table 11.4 Probability of correct majority decision among 990 voters with individual
competence pc* = 0.55 split evenly among multiple opinion leaders (pOL = 0.55 when not
following another opinion leader), each voter following his respective Local Opinion
Leader (and each Local Opinion Leader his respective Big Opinion Leader) with
probability π.
Opinion leaders Group partition π (probability of followers adopting
position of their opinion leader)
0.1 0.2 0.3 0.5 0.75

3 Big Opinion Leaders, 10 LOLs per BOL; 0.99 0.97 0.89 0.72 0.59
30 Local Opinion Leaders 33 voters per LOL
3 Big Opinion Leaders only 330, 330, 330 0.93 0.80 0.63 0.57 0.57
the time. The driving force behind this result is this: the more independent,
competent common causes there are shaping people’s votes, the better (if there
have to be common causes at all). And if the number of common causes grows
at the same rate as the number of voters (effectively setting up more and more
subgroups influenced by new common causes of the same competence pOL),
then we conjecture that group competence converges to 1.
This result might go some distance towards explaining changes in the quality
of electoral decision in recent decades. Suppose that Lazarsfeld’s 1940s world
with lots of Uncle Freds talking politics over the back fence and in the local pub
has, nowadays, largely given way to a world in which we largely mainline our
political news and views by direct feed from television outlets like Fox News
and its ilk. Judging from Table 11.4, the loss of the mediating structure of lots
of Local Opinion Leaders would carry serious epistemic costs, even (indeed,
particularly so) among voters who follow whatever opinion leaders they do
only to a modest extent. If there is a sense that the quality of electoral decisions
has declined over the past decades, the loss of all those Uncle Freds may be
largely to blame.
12
Taking Cues
Most ordinary voters are pretty poorly informed about politics. How can they
make competent political choices given so little political knowledge? A now
widely accepted answer is: through the use of ‘informational shortcuts’. Voters
‘take cues’ from all sorts of things, using them as pointers as to how they
should vote.
We begin by rehearsing that theory of ‘low-information rationality’. Then
we look more closely at the particular sorts of cues that guide people politically,
to see just how much of a risk is posed to the epistemic performance of a group
when its members act in that way.
12.1 CUE-TAKING AND LOW-INFORMATION

RATIONALIT Y
The theory of ‘low information rationality’ within politics was prompted by

Downs’s insight that, given the low probability of a given voter’s vote changing
the election, it is irrational for him to pay any substantial costs to acquire political
information purely to inform his vote.1 Given that voters are acting based on
very limited information when voting, how can they vote correctly? That is the
problem to which the theory of low-information rationality offers a solution.
One way in which people who themselves are low in political information
might proceed is to consult others with more political information than them-
selves.2 This is the classic ‘political influence’ story from mid-twentieth-century
political science.3 We have already discussed that strategy, and how it fits with
1 Downs (1957, chs 11–13), as already discussed in Section 6.1.1. Even if voters themselves
appreciate only half of that cost–benefit equation, the cost side alone is enough to dissuade them
from becoming politically well informed just for the sake of casting their ballot (Hardin 2006).
2 Or they may defer to, rather than merely ‘consult’, such others—but note well that it matters
greatly for CJT purposes which.
3 Katz and Lazarsfeld 1955.
Taking Cues 179
the CJT, in the previous chapter (and will return to say a little more about it in
Chapter 14).4
Here we shall concentrate instead on other sorts of ‘informational shortcuts’
that uninformed voters might employ in their attempt to vote rationally. The logic
behind these informational shortcuts is continuous with that underlying the
many other habits and rules of thumb by which we negotiate our everyday lives.
They are ‘strategies of simplification that reduce the complexity of judgment
tasks, to make them tractable for the kind of mind that people happen to have’.5
Here are some of the cues most commonly mentioned by advocates of ‘low-
information rationality’ in the political realm:6
• Voters might rely on ideologies or abstract policy principles7 to simplify
their decision task.
• Voters might rely on party labels, just as relatively uninformed consumers
might rely on brand labels.8 Party identification is generally a pretty reliable
indicator of what policies a Congressional candidate will vote for in office.9
• Voters might rely on endorsements by newspapers or lobby groups with
whom they generally agree (or rely negatively on those with whom they
generally disagree).10 For example, when assessing a proposal to reform
4 Feldman et al. (2012) and Hochschild (2012) provide an overview of evidence suggesting
that communication from partisan elites is not the only cue at work among the mass electorate
(cf. Zaller 1992; 2012).
5 Kahneman et al. 1982, p. xii. See further: Ferejohn and Kuklinski 1990; Popkin 1991, ch. 3;
Lupia and McCubbins 1998; Gigerenzer et al. 1999; Iyengar 1990; Kelman 2011; Lupia 2015, ch. 3.
As Lupia and McCubbins (1998, p. 18, quoting Churchland 1995, p. 15) put it:
[P]eople use limited information to draw complex inferences. [This] connectionist activity
underlies ‘all of the distinctive cognitive properties displayed in living organisms, such as:
• the capacity for recognizing features or patterns through a veil of noise and distortion, or
given only partial information;
• the capacity for seeing complex analogies;
• the capacity for recalling relevant information, instantly, as it bears on novel circumstances;
• the capacity for focusing attention on different features of one’s sensory input . . . .’
Indeed, there are corresponding mechanisms at work in biological systems: Lupia (2015, p. 31) quotes
Andy Clark (1997, p. 25) as follows:
Biological cognition is highly selective, and it can sensitize an organism to whatever (often
simple) parameters reliably specify states of affairs that matter to the specific life form. . . .
A wide range [of organisms] rely on simple cues that are specific to their needs, and both
profit by not bothering to represent other types in detail.
6 Achen and Bartels (2016, ch. 11) argue that what drives votes is ‘group identity’, which could
be regarded as yet another cue. Whether acting on that cue is epistemically advantageous just
depends on whether the cue is truth-conducive.
7 Which Goren (2013) shows to be importantly different. 8 Zaller 1992.
9 Poole and Rosenthal 2000.
10 Kuklinski and Hurley (1994) for example report that people cue on the messenger in inter-
preting the content of the message. Far from being a ‘cautionary’ finding, it makes perfectly good
sense—as they end up saying in their concluding observations—for blacks to suppose ‘black
leaders will always look out for my interests better than white leaders’ (p. 749). Muirhead (2014,
pp. 124–6) couches his epistemic defence of partisanship in just those terms.
insurance laws in their state, Californian voters cued on whether or not

insurance companies supported the proposal; and that led them to vote in
the referendum almost exactly the same way as they would have done had
they full detailed knowledge of the proposal’s complicated provisions.11
• Voters might use polls to gauge what other voters think (which might be
epistemically useful, as long as the others are on average of better than
random competence). Polls are also useful to learn about the viability of
candidates: a candidate polling far below a threshold of success is often
best abandoned for the next best choice with higher viability.12
• Voters might rely upon more episodic cues. For example, during the 1976
US election Gerald Ford chomped into a tamale without first shucking off
the surrounding maize husk; from that fact, Mexican-American voters could
infer the extent of his awareness of their concerns as well as their culture.13
• Voters might also rely on subjective cues like name recognition, a candidate’s
appearance, or even just how the candidate’s name ‘sounds’ (‘foreign’, ‘black’,
etc.).
12.2 HOW EFFECTIVE IS CUE-TAKING?
There are various ways of assessing how effective voters acting on cues actually
are at voting how they would have voted had they been fully informed.
12.2.1 Experimental Evidence
Much of the evidence comes from laboratory experiments. Those generally

show that cue-taking enables uninformed voters to make decisions that are
almost as good as those of well-informed voters—at least when the issue is one
on which there is enough activity (discussion, advertising, party position-taking)
to provide voters with any real cues at all.14
Here is an example of one such experiment. In a large computer-aided
experiment to investigate information acquisition in an election campaign, the
subjects are shown a large number of different headline items of information
scrolling down the screen.15 Clicking on an item opens a box for reading, while
other pieces of information continue to scroll (and therefore disappear).
11 Lupia 1994b. See further Lupia (2006; 2015). 12 Lau and Redlawsk 2001.
13 Popkin 1991, pp. 1–3.
14 See e.g.: Lupia 1994a, b; Lupia and McCubbins 1998, pt 2; Gerber and Lupia 1999; Boudreau
2009.
15 Lau and Redlawsk 2006.
Taking Cues 181
In such an environment the subjects are challenged to pick some pieces of

information out of a plentiful but quickly changing menu, simulating the
situation a voter experiences during an election campaign. Two results stand
out. First, subsequent to the simulated election campaign, 70 per cent of
voters decided not to change their vote after carefully reading through all
available items of information. This means that most voters succeeded in
making the ‘right’ decision (measured by their own standards) with the fraction
of information they obtained during the campaign.16 Second, voters tend to
benefit from items that can be used as cues (endorsements, party affiliation,
ideology) especially if they have to choose from more than two alternatives.
Those findings suggest that cues are particularly useful when managing more
complicated choices.17
12.2.2 Evidence from Large-scale Surveys
Some evidence on the effectiveness of cues comes from sample surveys and other
such larger-scale studies.
Let us begin on a more negative note. The classic essay on the theme ‘voters
are fools’ is, of course, Philip Converse’s 1964 ‘The Nature of Belief Systems in
Mass Publics’, based on early American National Election Study surveys. Its
central finding was that most voters are ‘unsophisticated’ politically. By that,
Converse meant simply that they did not engage in the ‘active use of ideological
dimensions of judgment’ in thinking about politics. Only 15.5 per cent of voters
counted as ‘ideologues’ or even ‘near-ideologues’ on his measures.18
Now, ‘unsophisticated’ sounds like a pretty pejorative characterization. But
far from being bad news, his findings might actually be seen as very good news
from an epistemic point of view. It is, once we recall—with the aid of another
rightly famous chapter in the very same book—that ideological thinking is sys-
tematically distorted. There we read, for example, that ‘deviations from scien-
tific objectivity [are the] essential criteria of an ideology’; ‘the problem of
ideology arises where there is a discrepancy between what is believed and what
can be [established as] scientifically correct’; and so on.19
16 Lau and Redlawsk 2006, ch. 4. However, one problematic aspect of this study is the possibility
of confirmation bias.
17 Lau and Redlawsk 2006, ch. 10.
18 Converse 1964, pp. 214, 218. Converse’s findings suggest that more people cue on ‘group
interests’ (45% of voters) or the ‘nature of the times’ (22%). The latter refers to cases where ‘parties
or candidates were praised or blamed primarily because of their . . . association in the past with
broad societal states of war or peace, prosperity or depression’. Included there would presumably
be the classic pocketbook voting question that Reagan posed to the audience in closing the 1980
Presidential Debate: ‘Are you better off now than you were four years ago?’
19 Talcott Parsons, quoted by Geertz (1964, p. 50).
Converse’s finding that the vast majority of voters are not subject to one and
the same systematically distorted ideological bias when they vote is, therefore,
splendid news from a CJT perspective. It shows, in particular, that ideology
(although it could be one useful cue among many) is not the kind of pervasively
biasing cue that undermines the independence of all votes too severely—or at
least it is not, for the vast majority of the electorate.
Let us now sound a more positive note. For an example of how effective
low-information rationality strategies might be, politically, let us return to
the classic Bartels study of US presidential elections already introduced in
Section 6.1.1. His research design, recall, was to break up survey respondents
(1) first according to a range of demographic characteristics, and then within
each demographic (2) according to how well or poorly informed interviewers
thought the respondents were.
In Section 6.1.1 we were deeply critical of Bartels’s overall assessment of what
difference it made to the overall election outcome that some voters were rela-
tively uninformed. En route to those overall conclusions, Bartels also assessed
how different the voting intentions of uninformed voters were in comparison
to those of informed voters within the same demographic.20 It is that latter
assessment that will be the focus of our discussion here.
Averaging across all six presidential elections under study, Bartels found
that uninformed voters’ intentions deviated from those of informed voters
within the same demographic by 9.65 per cent. As Bartels concedes, that aver-
age conceals a fair bit of variation across elections; and only in three of the six
elections under study was that year’s variation between uninformed and
informed voters’ intentions statistically significant.21 Still, let us concentrate
on that 9.65 per cent headline figure, averaging across all those six elections.
Bartels himself asks precisely the question we would: ‘Are average deviations
of this magnitude surprisingly large or surprisingly small?’ He proceeds to pro-
vide this, for us telling, answer:
Obviously, the answer depends in large part upon one’s prior expectations. Some
useful perspective may be provided by noting that if every voter simply behaved
randomly, voting for each candidate half the time, the resulting average deviation
from ‘fully informed’ voting probabilities would be on the order of 20 percentage
points (ranging from 18 to 24 percentage points across the six elections). Thus, it
appears that the information voters bring to bear in presidential elections, albeit
20 It is worth emphasizing that he found that being uninformed made much less difference in
aggregate to the overall election outcome than it did in individuals’ voting intentions, just as a
proponent of CJT-like aggregation would hope. The former difference was on average 3.04%
across all six elections under study, the latter was 9.65% (Bartels 1996).
21 The difference was statistically significant only in the American National Election sample for
the presidential elections of 1972, 1984, and 1992 but not for those of 1976, 1980, and 1988 (Bartels
1996, pp. 209, 216).
Taking Cues 183
limited, reduces the average magnitude of their deviations from a hypothetical

baseline of ‘fully informed’ voting by about 50%.22
Whatever informational shortcuts voters are using, be they cues or some others,
seem to work—not perfectly, but at least tolerably well.
12.3 CAUTIONARY TALES
Cues are chosen precisely to be reliable indicators, and presumably they would
not persist in use were they systematically not so. In the modelling done in the
sections that follow, we will be relying heavily on that sort of assumption. But
we should acknowledge from the start that cues are certainly not perfect in that
regard. As those subsequent sections shall show, however, in the context of
the CJT we can afford for cues to be somewhat unreliable, as they still have a
powerfully positive epistemic effect.
12.3.1 Some Cues Might Be Unreliable
Cues are informational shortcuts, and (as any hiker knows) what looks like a
shortcut does not always take us to our desired destination. And even cues that
are ordinarily pretty reliable will almost inevitably sometimes mislead. Party
labels, for example, are ordinarily good indicators of the policies that a politician
is likely to pursue. But they are not invariably so. There are some rogue Democrats
who are to the right of some rogue Republicans.23
Here is one example of the effects that such misleading cues might have.24
A large survey asked US citizens to retrospectively identify the votes of their
respective two senators in important recent roll-call votes. Conveniently, there
were quite a few instances in which one of the two senators voted against the
party line, rendering the party affiliation cue useless; and these were the instances
on which the study focused. It found that voters who are most interested in
politics were in those instances least likely to correctly identify what the deviating
22 Bartels 1996, pp. 216–17.

23 Furthermore, the ‘possibility set’ open to political leaders may be inversely related to their
general pattern of policy preferences: only Cold Warrior Richard Nixon, who began his political
career looking for communists in pumpkin patches, could have opened US relations with Red
China, and only Zionist terrorist Menachem Begin, who began his career blowing up the King
David Hotel headquarters of the British army in Palestine, could have signed a peace accord with
Egypt. See Goodin 1983; Somin 2006, pp. 263–4.
24 Dancey and Sheagley 2013.
senator did. In other words: if the cue of party affiliation is misleading, it is

particularly misleading for the politically most interested.25
The authors of that study present this finding as a major worry.26 That is quite
probably an exaggeration.27 It is indeed true that even those who are most
interested in politics nonetheless rely on cues about politics and can get things
wrong when those cues are misleading. Just how often it happens that cues
mislead is, of course, a separate question.
12.3.2 Some Cues Might Not Be Chosen for Their Reliability
A second, related risk is that some cues might not have been selected for their
truth-tracking attributes in the first place.
Some cues might not have been selected at all. Such cues might act upon us
autonomously, as do (arguably, at least) psychological heuristics.28 We might
have some reason to think that cues that have been intentionally chosen for
their being generally good guides are indeed generally reliable. But if cues—or
still more psychological heuristics—have not been intentionally chosen we
would have no such reassurance. (Notice, however, there may still be some
evolutionary story to be told to vindicate our faith in the general reliability of
psychological heuristics, at least across the period over which they evolved.29
That may or may not make them well suited to the current environment in
which they are guiding our political choices, of course.)
Even cues that have been intentionally selected might have been selected for
other reasons than veracity. Take the case of party labels, once again. People
might affiliate with a political party, and follow it wherever it leads, in the same
way they follow a football team—for the sheer thrill of expressing tribal loyalties,
rather than out of any concern for the best public policies.30
25 Crucially, Dancey and Sheagley don’t classify the answer ‘Don’t know’ as a sign of misinfor-
mation, which makes the performance of the politically uninterested look much better than it
actually is.
26 ‘[T]he class of citizens many scholars presume will be the best citizens – the politically
interested – are actually the most misinformed segment of the population when heuristics go bad’
(Dancey and Sheagley 2013, p. 323).
27 It is true that sometimes more knowledge can also be more misleading, especially when
focusing on very specific questions. But it is an exaggeration to take the ‘retro-diction’ of a roll-call
vote as a representative measure of political information, let alone competence. In more realistic
political choices the voters will also be able to draw on more than one cue.
28 Lupia 2015, pp. 36 ff. Popkin (1991, chs 3, 4) treats cues and informational shortcuts as a
matter of ‘going without data’ and heuristics as a matter of ‘going beyond the data’, but as being
otherwise indistinguishable; he does not note that the one may be volitional and the other not.
Note that all the examples in Sunstein’s (2006b, pp. 34–6) discussion of ‘biases’ are actually psy-
chological heuristics (like ‘anchoring’).
29 Gigerenzer et al 1999. Gigerenzer 2008, esp. ch. 3.
30 Somin 2006, p. 261. Cf. Rosenblum 2008; 2014; White and Ypi 2011; 2016.
Taking Cues 185
We should be careful, however, about what we infer from the fact that some
particular cue was not intentionally selected on grounds that it is good at track-
ing the truth. We might then have no good reason for supposing that that cue
will systematically lead us to true judgements. But we have no good reason for
supposing that it will necessarily lead systematically to false ones, either.31
12.4 THE EPISTEMIC EFFECTS OF CUE-TAKING:

T WO MODELS
In discussing the effectiveness of cues, the empirical literature we have been

examining so far employs an internal criterion of correctness: to what extent do
ill-informed voters vote the same way, using cues, as they would have voted had
they been fully informed? For the purposes of the discussions that follow, we
now shift to the CJT’s external standard of correctness. Voters being correct in
this sense means being correct in their assessment of the true state of the world.
When assessing whether cue-taking enhances the epistemic performance of
a group of voters, there are various factors that we need to consider:
(1) How accurately informative are the cues?
(a) How likely is each cue to indicate the correct state of the world?
(b) How many different cues do the voters use?
(c) Are different cues independent of one another?
(2) Do the voters obtain all their evidence about the state of the world
through cues or do they also have independent access to other evidence?
And if the voters also have independent access to other evidence, are the
cues or that other independent evidence the dominant factor in their
voting?
(3) How likely are the voters to track the vote of a best responder of the
shared evidence provided by the cues?
Put briefly: Factor (1) concerns the epistemic quality of the cues, taken all
together. Factor (2) concerns whether the voters are completely reliant on the
cues or whether they also have other more direct access to the evidence.
Factor (3) concerns the competence of the voters to interpret the cues in the
best possible way.
To see why these considerations matter to the epistemic performance of the
group, consider two different scenarios. First, as a worst-case scenario, suppose
there is just one Cue. Assume that that Cue is not always pointing in the right
31 This may well be the case with Achen and Bartels’s (2016, ch. 11) ‘group identity’, for example.
State
Evidence
Cue
Figure 12.1 One Cue as the only access to the Evidence for all voters.
direction.32 Also assume that the voters do not have any other access to the
Evidence. That is the situation depicted in Figure 12.1.33
In the Figure 12.1 scenario, the more independent34 voters there are compe-
tently35 following what is suggested by the Cue (with a sufficiently high prob-
ability of following, π), the more likely the majority vote is to be correct with
the same probability that the cue itself is correct, which we denote pK. But the
collective competence of a group of such voters (i.e. voters whose only access
to the Evidence or the State of the world is through the information provided
by the Cue) cannot exceed the probability that the cue itself is pointing towards
the correct answer. That is implied by the Best Responder Corollary from
Section 5.3.
Epistemically, the Figure 12.1 scenario suffers badly from its monolithic Cue
structure. One and the same Cue informs everyone’s Vote; and if that Cue is
misleading then everyone will be misled at one and the same time by it. That is
the problem commonly discussed under the heading of ‘bias’. Race, class, and
gender are among the biases that are often supposed to operate in this way,
monolithically driving a great many voters in the same—all too often wrong—
direction, at one and the same time.36
32 We also assume that ‘Cue’ is a binary variable that points towards one of the two possible
answers.
33 Such a setting is functionally equivalent to having a single opinion leader, as discussed in
Sections 5.2.2 and 11.1.
34 Independent, here, conditional on the cue; see Section 5.2.
35 Competence here is defined in terms of voting in the way that the cue truly says to vote, not
(as in the standard CJT setting) in terms of voting in accordance with the true State of the world
or even in accordance with the Evidence about that—neither of which, in the Figure 12.1 scenario,
voters have any direct access to.
36 On race in particular, see Mendelberg (2001; 2008), Mendelberg and Oleske (2000), and
Mendelberg and Berinsky (2005).
Taking Cues 187
A bias might do that directly, by predisposing everyone who shares it to

answer some question in a certain way, without any reference to the evidence
whatsoever. Alternatively, a bias might shape everyone’s interpretation of the
evidence in the same way (in similar fashion partisan party-political biases
have been shown to give rise to selective perceptions, for example37).
The monolithic cue structure in Figure 12.1 upsets the classic CJT result in at
least one and potentially two ways. First, it violates the CJT’s Independence
Assumption. The votes in Figure 12.1 are not independent of one another in the
way that the classic CJT requires. Hence, no matter how many voters there are,
the probability of them being correct will not converge to certainty, as in the
classic CJT. Instead, as we have said, that probability will (at best) merely con-
verge to the probability that the Cue is itself not misleading.
Furthermore, if the Cue is more likely to be misleading than truth-conducive
then the vote of each voter who is more likely than not to follow it is more likely
than not to be wrong rather than right. That is the second way in which the
Figure 12.1 scenario might upset the classic CJT result—by violating the CJT’s
Competence Assumption.
At the opposite extreme, consider the Figure 12.2 scenario. There we find
each Vote being informed by multiple, independent38 Cues each of which
tracks the Evidence with more-or-less reliability. In that scenario, the probabil-
ity that the majority verdict will be correct can, given suitable parameters, be a
quickly increasing function of the number and reliability of Cues.39
Note that the same thing we say about Cues here could equally well be said
about psychological heuristics. They too are multiple. As Eskridge and Ferejohn
State
Evidence
Cue 1 Cue 2 Cue 3
Figure 12.2 Several Cues as well as direct access to the Evidence.
37 Layman et al. 2006. Marcus 2008. Shapiro and Bloch-Elkon 2008. Jacobson 2010. Cf. Fiorina
and Abrams. 2007.
38 Independent, here, conditional on the Evidence; see Section 5.2.
39 As in our discussion of multiple common causes in Section 5.4.3.
wryly observe, ‘demonstrated cognitive biases have grown like weeds in a vacant
lot. As documented biases have multiplied, it has become harder to reach con-
clusions from them. In any given institutional situation, there will be several
potentially applicable and potentially cross-cutting biases.’40
More than sixty psychological heuristics have now been identified. As long
as many of them operate on decision-makers simultaneously and independently
of one another, and as long as their evolutionary history makes them tolerably
good guides to present-day decisions, the happy epistemic consequences asso-
ciated with the Figure 12.2 scenario will still occur.41
Notice, finally, that in the Figure 12.2 scenario voters also have independent
access to the Evidence as well as the Cues. Assuming voters have better than
random competence at following the Evidence, that will improve the collective
epistemic performance of the group further still.
12.5 CALCULATING THE POTENTIAL EPISTEMIC

EFFECTS OF CUE-TAKING
In what follows, we shall attempt to explore a version of broadly the latter sort
of model, assuming (in line with the standard CJT setup) that the voters track
some independent truth that is the same for all of them.42 The particular values
we plug into the model are, here as always, purely speculative and not grounded
in any hard empirical evidence. Still, we think them defensible for exploratory
purposes—and we think that the results of the modelling based on them are
pretty compelling, as we shall show.
The setup we envisage here is one in which there are 990 voters, each of
whom is potentially influenced by (up to) nine cues. We suppose that each
voter is guided by each cue with probability π, a probability that we assume to
be the same for every voter and every cue. Hence, each voter’s cues are decided
by picking each of the nine cues with probability π. Furthermore, we assume
that each voter will consider simultaneously all selected cues.43 With high values
40 Eskridge and Ferejohn 2001, p. 633. Quoted approvingly by Elster (2013, pp. 86–7).
41 Wittman (1995, ch. 5) similarly notes the multiplicity of psychological heuristics and argues
that the effects of all heuristics taken together might not be systematic and might hence balance
out. For evidence on how people respond to multiple competing ‘frames’ see: Druckman and
Nelson 2003; Druckman 2004; Chong and Druckman 2007a, b.
42 Most of the empirical literature on correct voting assumes, implicitly or explicitly, a voter-
dependent standard of correctness: voters are assumed to be correct when they vote as their fully
informed counterpart self would. Note that that approach does not logically rule out that all voters
apply the same standard of correctness (though given that even fully informed voters almost
always disagree, it does so in practice).
43 Rather than considering them in some lexical sequence of cues, each of which is taken to
be one-way decisive. The latter sort of structure has been found to be operative among English
magistrates in deciding whether to grant prisoners’ bail: if the prosecution opposed bail or asked
Taking Cues 189
of π, most voters will then end up being guided by multiple cues, which may
conflict; and when they do, we assume the voter will vote in whichever way is
indicated by a majority of the cues she is guided by (with ties being broken
randomly). When guided by no cue, we assume that each voter casts a purely
random vote, so the probability of her voting correctly is then pc* = 0.5.
Our model of cue-taking crucially differs from our models of multiple opin-
ion leaders in Section 11.2 in the following three respects. First, people might
follow opinion leaders for all sorts of reasons unrelated to any expectations
about their epistemic reliability. People typically choose which cues to use, in
contrast, precisely on the grounds they are generally good guides. Hence, we
assume that pK—the probability of each cue being correct in its guidance—is
much higher than in cases of opinion leaders or other common causes.44
Specifically, we assume pK = 0.70.
Second, people often end up being influenced by opinion leaders or other
common causes unintentionally, and they may be influenced by them only epi-
sodically and intermittently in consequence. Cues, in contrast, have been
intentionally chosen by people in order to guide them. Hence, we assume the
probability of any given voter being guided by any given cue on any given
occasion is relatively high. Specifically we assume π = 0.70.
Third, in the models to follow we assume that voters can be guided by more
than one cue, while in Chapter 11’s initial model (but not later ones) each voter
was influenced by only one opinion leader.
12.5.1 Baseline Calculation
Let us now calculate, on the basis of those assumptions, the probability that a
majority of voters guided by cues in that way will reach the correct answer in a
dichotomous choice.45
Basically, the news is very good indeed. Here is the baseline calculation.
Suppose we have nine cues, each of which is pK = 0.70 likely to be correct and
which each of 990 of voters uses with probability π = 0.70; and suppose that
each voter is of purely random individual competence when not following any
cue (pc* = 0.5). Then the probability that the majority of such voters will be
correct is, Pn = 0.90. In the absence of any cues, of course, purely random voters
for it to be made conditional, that is the end of the matter and unconditional bail is denied; but
if that is not the case, the magistrate turns to consider whether a previous court imposed con-
ditions or remanded the prisoner in custody on some previous occasion; and so on (Dhami 2003;
Gigerenzer 2008, pp. 48–9).
44 And, in line with our normal assumptions, we assume that pK is the same for all possible
states of the world. We also assume that the cues are independent conditional on the evidence.
45 For convenience: similar results would obtain for many-option choice situations.
with individual competence pc = 0.5 would be collectively correct only Pn = 0.5

of the time.
Just as our discussion of multiple common cause influence in Section 5.4.3
would lead us to expect, what drives this result is the number and reliability of
independent cues. Suppose that in the above scenario we had not 990 voters,
but rather merely twenty-one voters, with all other factors remaining the same
as before. The probability that the majority of that much smaller electorate will
be correct is still Pn = 0.90.
Neither do these results depend on better-than-random individual competence
of voters, as in the standard CJT. For instance, the probability that the majority
of those voters will be correct remains Pn = 0.90, even if the individual competence
drops to pc* = 0.40. We will say more about how competent cues can epistemically
compensate for voters having worse-than-random individual competence and
about the limited circumstances under which higher voter competence can
improve group competence in Section 12.5.3.
12.5.2 Sensitivity to the Number

and Reliability of Independent Cues
What drives the above results in the cue-taking case is, as we have said, the
number and reliability of the cues. Now let us deploy some sample calculations
to see just how sensitive the results are to each of those factors.
First let us explore the effects of reducing the number of cues. In the previous
calculations we assumed there were nine cues, each of which each of 990 voters
uses with probability of π = 0.70. The probability that the majority of such voters
will be correct in that previous calculation was Pn = 0.90. Now let us assume
there are only three cues, with all else as before. The probability that the majority
of voters will be correct drops—but only to Pn = 0.78. That is lower, but not all
that much lower.
Next let us explore the effects of lowering the competence of the cues. For
that, let us revert to the case of 990 voters each of whom is guided by each of
nine cues with probability of π = 0.70. But instead of each cue being pK = 0.70
likely to be correct, as in our earlier calculation, let us now assume each cue
is only pK = 0.60 likely to be right. Then the probability that the majority of
voters will be correct is Pn = 0.73. Again, that is lower, but still a reasonably
high value.
In short, the results reported above are only somewhat sensitive to the num-
ber of cues and the reliability of the cues. Both can be reduced a fair amount
from our baseline model, and the majority of voters will still be very likely to be
correct when using cues on those bases.
Taking Cues 191
12.5.3 Relative Insensitivity to the Rate of Cue Use
Yet another factor at work in this model is the probability with which (or rate at
which) people use the cues. In all of our previous models we have held that
constant at π = 0.70. That is a pretty high rate, however. Might our results be
being driven by that particular parameter specification?
To see, let us now vary the probability with which people use cues. For the
purposes of this calculation, we will revert to our baseline model in all other
respects. We assume there are 990 voters, each of whom is no better than a coin
toss when not guided by any cue. We assume there are nine cues, each of which
is pK = 0.70 likely to be correct. But in the calculations that follow we substitute
for much lower values of π than in our baseline calculation.
In our baseline calculation we assumed that each voter uses each cue with
probability π = 0.70. In that case, the probability of the majority of the 990 voters
being correct in our baseline case was Pn = 0.90. What happens if we drop the
probability of each voter using each cue to π = 0.20? Nothing. The probability
of the majority of the 990 voters being correct remains Pn = 0.90. It does likewise
if we drop the probability of each voter following each cue to π = 0.10. And if we
drop π further to π = 0.05 the probability of the majority being correct falls ever
so slightly, to Pn = 0.89.
12.5.4 Cue-taking with Varying Individual Voter Competence
In our previous calculations, we have been assuming that voters are no better
than random in their individual competence levels (pc* = 0.5). Next let us explore
the effects of cue-taking under different assumptions about levels of individual
competence. Figure 12.3 shows the probability of the majority of 990 voters
being correct, for varying values of cue use (π) and for varying assumptions
about the competence of individual voters (pc*), assuming there are nine cues
each of which is pK = 0.70 likely to be correct.
In our baseline case, where voters were not better than random in their indi-
vidual competence levels, the probability that a majority among them would
have voted correctly in the absence of cues would itself be random, Pn = 0.5.
When such voters make even modest use of cues that are themselves even
moderately accurate, the probability of the majority of voters being correct can
be lifted substantially above that, as we have already seen.
Where voters are above pc* > 0.50 competent, as on the right side of
Figure 12.3, the probability of the majority of 990 voters being correct is higher
than random, increasingly so the higher individual voter competence is. That is
true without voters using cues at all, of course (that is the classic CJT result,
1.0
π = 0.7
0.8 π = 0.2
π=0
0.6
Pn
π = 0.1
0.4
0.2
π = 0.05
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Individual competence pc*
Figure 12.3 Probability of majority voting for correct alternative, with varying levels
of individual competence pc* and probability of being guided by any given cue π, for
990 voters, nine cues, and probability of any given cue being correct of pK = 0.70.
represented in Figure 12.3 by the line for π = 0). When better-than-competent

voters use cues, their collective competence is typically less than if they had not
used cues at all. (In Figure 12.3, the lines representing all rates of cue use above
zero are a little lower than the line representing the case of no use of cues at all.)
That is inevitable, because when many voters use the same cues the independ-
ence of their votes is compromised, and collective competence is reduced in
consequence. The important thing to notice on the right side of Figure 12.3 is
that cue use does not compromise the collective competence of individually
competent voters very much. Even if voters make heavy use of cues their col-
lective competence is still very high. Take the case of π = 0.7 for example: it
remains the case that Pn = 0.90. That is a high value, even though of course it
would have been higher still if they had not used cues at all.
Next let us consider the case in which voters have individual competence
(without cues) worse than random, pc* < 0.50. That case is represented by the
left half of Figure 12.3. In the classic CJT setup without any use of cues, a large
number of independent voters who have competence worse than random is
epistemically a disaster. Then the CJT goes into reverse: the majority of such
voters is less likely to be right than any single voter; and the probability that the
majority of such voters will be right approaches zero as the number of voters
approaches infinity.46
46 See our discussion in Section 4.4.

Taking Cues 193
If worse-than-random voters are guided by reasonably reliable cues, their

collective competence is always better than it would have been if they had not
used any cues at all. In Figure 12.3, all the lines for values of π > 0 are above the
line for π = 0, which represents the case where voters make no use of cues at all.
Not only do individually incompetent voters do better, epistemically, when
using cues. At least in some cases, they can do dramatically better. Take for
example the case of an electorate in which each voter is individually only
pc* = 0.2 likely to be right if voting purely on the basis of her own judgement.
If such voters are guided by the nine cues as described with probability of
only π = 0.20, the probability that the majority of the voters will be correct
remains Pn = 0.89, as shown in Figure 12.3. Indeed, even if voters use those
cues with a probability as low as π = 0.10, the majority remains more likely to
be right than not (Pn = 0.54, to be precise).47
Of course if the probability of voters using the cues drops too low, the
epistemic disaster ordinarily associated with incompetent voters in the classic
CJT remains. If voters who are individually only pc* = 0.2 likely to be correct use
each cue with a probability of only π = 0.05, then (as shown in Figure 12.3) the
probability that a majority of the voters will be correct is a minuscule Pn = 0.02.
Still, across a wide range of cases, using reasonably reliable cues can help to
prevent the epistemic disaster usually associated in the standard CJT with
incompetent voters.
12.5.5 Cueing Incompetents Only
In our previous calculations we have been assuming that all voters are alike in
their probability of being guided by cues. But in a way that is unrealistic. Voters
who are relatively well-informed will more typically decide how to vote on the
basis of their own knowledge. Cues are crutches used by poorly informed voters
with no other basis for voting one way rather than another.
Let us now adapt our modelling to reflect that fact about the real world of
politics. Suppose once again that we have 990 voters, but now suppose that
they are not all equally competent. Suppose that 330 of them are relatively
‘incompetent’, with a probability of being individually correct of pc* = 0.45; and
suppose that the other 660 of them are relatively ‘competent’, with a probability
of being individually correct of pc* = 0.55. Averaging across all 990 members
of the electorate, individual competence is around pc * = 0.517.
If we took a majority vote among all 990 voters, the probability of that majority
being correct would be Pn = 0.86. That is pretty good—although, of course, had
47 The slightly S-shaped sections in the figure, most pronounced for π = 0.10, are a numerical
artefact due to the ‘lumpiness’ caused by the small number of cues.
we restricted the electorate to the 660 competent voters alone the probability of
that majority being correct would have been Pn = 0.99.
But now look what happens when we introduce some cues for the incompe-
tent voters (and them alone) to use. Suppose as before there are nine cues, each
of which is pK = 0.70 likely to be correct. And suppose each of the 330 incompetent
voters are guided by each of those cues only just a little bit, with probability
π = 0.20, say. Now take a vote among all 990 voters once again, this time with
the incompetent guided by cues as just described. The probability that the
majority among those 990 voters will be correct jumps from Pn = 0.86 without
cue-taking to Pn = 0.99 with it. In short, in this case, a majority vote is as likely
to be correct if incompetent voters use such cues as it would be if we confined
the electorate to competent voters alone.
13
Pluralism
Differing Values and Priorities
The classical CJT is based on the assumption that there is one correct answer,
and that the correct answer is the same for everyone. Over a great range of
decisions, that is undoubtedly true. But sometimes we have to take account of
what Rawls calls ‘the fact of reasonable pluralism’. Different people have differ-
ent values or priorities or interests, and that leads to different options being
‘correct for them’ given their own particular values or priorities or interests.
Maybe that is merely a contingent (but exceedingly likely) consequence of
living in a large and heterogeneous society. Or maybe it is linked in some
deeper way to how free and equal people must relate to one another. Rawls
himself famously argues the latter.1 Be that as it may, it is undeniable that purely
contingent connections of the former sort will generate differences among
people in values, priorities, and interests with which any modern democratic
society must cope.2
1 In Political Liberalism Rawls (1993, p. 36) writes:

[T]he diversity of reasonable comprehensive religious, philosophical, and moral doctrines
found in modern democratic societies is not a mere historical condition that may soon pass
away; it is a permanent feature of the public culture of democracy. Under the political and
social conditions secured by the basic rights and liberties of free institutions, a diversity of
conflicting and irreconcilable—and what’s more, reasonable—comprehensive doctrines will
come about and persist if such diversity does not already obtain.
From that perspective he would focus on differences in people’s values rather than in their interests.
Rawls (1993, pp. 36–7), again, writes:
This fact of reasonable pluralism must be distinguished from the fact of pluralism as
such. . . . These . . . reasonable comprehensive doctrines . . . that reasonable citizens affirm . . . are
not simply the upshot of self- and class interests, or of peoples’ understandable tendency to
view the political world form a limited standpoint. Instead, they are in part the work of free
practical reason within the framework of free institutions.
2 Hirst 1989.
When elsewhere in the book we speak of the ‘correct option’, we mean by

that correct independently of the voter’s own values or priorities or interests. In
other words, the standard of ‘correctness’ in the classical framework is external.
In this chapter and the next, however, we will be focusing instead on a voter’s
choice among the options being correct as a function of and from the point of
view of that voter’s own values or priorities or (in Chapter 14) interests.
If the correct answer to the same question is different for different people,
what makes a group decision the correct one? One plausible proposal is to endorse
the answer supported by most voters.3 We call that outcome ‘democratically-
epistemically correct’, for short.
What the correct answer is for the group, understood in that way, thus
depends on the values, priorities, or interests of the voters in the group. Even if
there is no one option that is correct from the point of view of literally every-
one, it is democratically better (assuming some substantive safeguards against
majority tyranny are in place) to settle upon the option that is correct from the
point of view of the largest number of people. That is the ‘moral majoritarian’
position that we introduced in Section 4.1.4.
Here, as throughout the book, we assume the group makes decisions
through majority or plurality rule.4 Two important points motivate our setup.
The first is that it is not necessarily the case—it is not just analytically true—
that the majority/plurality chooses the option that is epistemically correct in
terms of the majority’s/plurality’s values, priorities, or interests. It is perfectly
possible for people to choose an option that is objectively less well suited than
some other for promoting their own values, priorities, or interests. Individually,
people make mistakes like that all the time. As we shall show over the course
of this chapter and the next, CJT-style logic provides some substantial reassur-
ance that (assuming its assumptions are satisfied) people are much less likely
to do so collectively, with democratic majorities correcting idiosyncratic
individual errors.
Second, even if it is unsurprising (albeit far from being guaranteed) that
individuals ordinarily choose the correct option given their own values, pri-
orities, or interests, it remains surprising—as CJT style results typically are—
just how quickly the probability that the majority/plurality of people will
collectively choose correctly can increase with increasing numbers of com-
petent voters.
The classic CJT reassures us that, with a large number of competent, inde-
pendent, and sincere voters, the majority/plurality is highly likely to be correct
from an ‘objective’ point of view. That reassurance is not available in the very
different settings involved in this chapter and the next. Still, it remains reassur-
ing that the majority/plurality is virtually certain to choose the correct outcome
3 As proposed by Miller (1986), Goldman (1999, ch. 10), List and Spiekermann (2016).
4 Majority rule in binary choice situations; plurality rule in choices among k > 2 options.
Pluralism: Differing Values and Priorities 197
from its own point of view, at least. While that is something less than the classic
CJT offers, it is nonetheless valuable from what we will call a ‘democratic-
epistemic’ perspective.
In this chapter, we develop that argument with respect to differing values
and priorities, reserving the case of differing interests for the next chapter.
Mathematically, the structure of all three cases is identical. But the substantive
interpretation of the results is sufficiently different with respect to differing
interests that it behoves us to treat that case separately.
13.1 DIFFERING VALUES
Many different things might be meant by ‘value pluralism’. Here we will explore
only one aspect of it. Specifically, we will assume that different people have dif-
ferent values, and that the best option for serving those different values differs.
Thus, it is the electoral consequences of ‘conflicts of values’ that will concern
us in this discussion. Of course, adjudicating conflicting values via a vote is not
the only way of dealing with such conflicts, and it may not always be the best.
We will return to say a little more about that in Section 19.3. But for purposes of
this chapter, we will focus upon that way of dealing with them.
13.1.1 Baseline Scenario
For our baseline model, let us suppose there are n voters, with an equal number
harbouring each one of i values V1, V2, . . . , Vi.5 Suppose that for each value Va
there is a corresponding outcome Xa that is the correct one for the voters sub-
scribing to Va to prefer.6 We also assume for the baseline case that all such out-
comes X1 . . . Xi are different (so that no two voters with different values agree on
the same best outcome). For purposes of the baseline scenario suppose that each
voter is pcV likely to vote for the outcome that is correct, given the value V that he
harbours; and suppose (for now) that the level of pcV is the same for all voters.
Suppose furthermore that the probability of each voter’s voting for the outcome
that is correct from the point of view of his value is higher than the probability
of his voting for the outcome recommended by any other value. Finally, we sup-
pose that when a voter makes a mistake and votes for some option other than
that which is correct from his own value’s point of view, he has equal probability
of voting for each of those other options that are incorrect from his point of view.
5 This means that n/i is a natural number > 0 in this baseline scenario.
6 The outcome variables X1, X2, . . . are akin to x1, x2 . . . in Section 3.2. We capitalize them here
to emphasize that the outcomes here under discussion are associated with a particular value.

V1 voters for X1 V2 voters for X1 V2 voters for X2 V1 voters for X2
0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.1 Baseline scenario with two groups of equal size and all pcV = 0.55.
Now we increase the numbers of such voters in a specific way: we hold the
proportions of voters subscribing to different values constant while increasing
the population. If we do so, it will be increasingly certain that the proportion of
voters backing each option will be nearly the same as the proportion of voters
who harbour each value. That is the law-of-large-numbers effect that drives the
CJT—and we will observe it here likewise, even in the very different context of
the models here under discussion.
Let us start by considering what would happen in the baseline scenario.
Supposing the number of voters is large, we would expect voters to split virtu-
ally equally among each of the options. The outcome will thus be, in our base-
line case, a virtual tie among all the options favoured by any of the values.
Figure 13.1 provides an illustration of that, for the case of a great many voters,
half of whom harbour value V1 and half of whom harbour value V2. Suppose
that each of those voters is pcV = 0.55 likely to vote for the correct outcome from
the point of view of his own value, which is X1 for those who harbour value V1
and X2 for those who harbour value V2. Figure 13.1 shows that the distribution
of votes among those options, under those circumstances, is likely to be a tie.
(The thick, dark line in the centre, in that figure and equivalents in this chapter,
represents the winning threshold.)
13.1.2 Six Variations
Now let us vary that baseline scenario in several ways.

Scenario 1: All else is as in the baseline case, except let the number of voters
harbouring value V1 be greater than the number of voters harbouring any of
the other values.
In Scenario 1, the option that is correct from the point of view of value V1 (i.e.
option X1) is highly likely to win, and increasingly so the larger the electorate.
Unsurprisingly, more voters competently supporting value V1 results in more

0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.2 60% of voters with value V1, 40% with V2 and with all pcV = 0.55.
votes for the option X1 that is correct from that value’s point of view, more than
for any other option.
Thus, where the electorate is large, the option in accordance with the value
supported by the larger group of voters will in all likelihood be the winner. That
is to say, the democratic outcome will accurately reflect the values (or, as we will
go on to say in Section 13.2, the priorities) that are most common across the com-
munity. That is the ‘democratic-epistemic’ good news that we advertised earlier.
Figure 13.2 provides an illustration of that. Here again, we assume there are
many voters, but now we assume that 60 per cent of them harbour value V1 and
only 40 per cent of them harbour value V2. Once again, we suppose that each of
those voters is pcV = 0.55 likely to vote for the outcome which is correct from the
point of view of his own value (X1 for those who harbour value V1; X2 for those
who harbour value V2). Figure 13.2 shows the likely distribution of votes among
those options, under those circumstances. As we see from Figure 13.2, the out-
come that is correct from the point of view of the majority value V1 is likely to
get safely over the dark bar indicating the winning threshold.
Scenario 2: All else is the same as in the baseline case, except let the compe-
tence of voters harbouring value V1 be greater than that of voters harbouring
each of the other values.
Given a large electorate once again, the proportion of the group voting for the
outcome that is correct from the point of view of that group’s value is virtually
the same as the probability of each member of that group voting for it. Ex
hypothesi, in this case there are equal numbers of voters in each group. So if
voters in one group are more competent, and hence more likely to vote for the
outcome that they should given their values, then there will be more votes for
that outcome than any other.
Figure 13.3 provides an illustration of that. Here again, we assume there
are many voters, half of whom harbour value V1 and half of whom harbour
value V2. But in this scenario, we suppose that voters who harbour value V1 are

0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.3 Two groups of equal size, but voters with V1 have pcV1 = 0.8 while voters
with V2 have pcV2 = 0.55.
each pcV1 = 0.8 likely to vote for the outcome which is correct that point of view
(X1), while voters who harbour value V2 are each pcV2 = 0.55 likely to vote for
the outcome which is correct that point of view (X2). Figure 13.3 shows the
likely distribution of votes among those options, under those circumstances.
As we see in Figure 13.3, the outcome X1 that is correct from the point of view
of the value V1 supported by more competent voters is likely to win.
In the next pair of variations on our baseline model, we allow variation in
both the numbers of voters supporting each value and individual competence
levels within each group.7 In Scenario 3 both of those factors vary in the same
direction, whereas in Scenario 4 they will vary in opposite directions.
Scenario 3: All else is the same as in the baseline case, except supporters of
value V1 are both more numerous than supporters of any other value and
they are also individually each more likely to vote for the correct outcome
from the point of view of their value (i.e., X1) than are supporters of each of
the other values to vote for the outcome that is correct from the point of view
of their respective value.
In Scenario 3, the option that is correct from the point of view of the value that
is supported by the largest number of people is obviously even more likely to be
the winner than in either of the first two scenarios.
Figure 13.4 provides an illustration of that. Here we assume there are many
voters, 60 per cent of whom harbour value V1 and 40 per cent of whom harbour
value V2. We further assume that voters who harbour value V1 are each pcV1=0.8
likely to vote for the outcome which is correct that point of view (X1), while
voters who harbour value V2 are each pcV2=0.55 likely to vote for the outcome
which is correct that point of view (X2). Figure 13.4 shows the likely distribu-
tion of votes among those options, under those circumstances.
7 On the idea of ‘competence asymmetries’ in general, see Bendor and Bullock (2008 pp. 11–12).
We return to these issues, with particular reference to different groups of people. being differen-
tially competent in assessing where their true interests lie, in Section 14.1.2 below.
0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.4 60% of voters subscribing to value V1 with pcV1 = 0.8 and 40% of voters
subscribing to value V2 with pcV2 = 0.55.
As we see in Figure 13.4, the outcome X1 that is correct from the point of view
of the majority value V1 is likely to be even further over the winning threshold
than in the Figure 13.3 case, representing Scenario 2.
But what if the number of voters and their competence at choosing the right
option given their own value pull in opposite directions? For the fourth vari-
ation on our basic scenario, let us consider the case in which the group that
contains the smallest number of voters is also the one whose voters are indi-
vidually most competent at choosing the right option from the point of view of
the value that they support.
Scenario 4: All else is the same as in the baseline case, except supporters of
value V1 are less numerous than supporters of any other value, but they are
individually each more likely to vote for the correct outcome from the point
of view of their value V1 (i.e. X1) than are supporters of each of the other
values to vote for the outcome that is correct from the point of view of their
respective value.
In Scenario 4, it is an open question whether or not the option that is correct
from the point of view of the value supported by the largest number of voters
will actually prevail in the vote. Everything depends on just how much less
competent voters in that group are at choosing the correct option given their
values, and on just how many more of those voters there are in that group
compared to the others.
It is perfectly possible, however, for smaller but more competent groups to
prevail over larger but less competent ones. Here is an illustration of that.
Assume there is a large number of voters, 60 per cent of whom harbour value
V1 and 40 per cent of whom harbour value V2. Assume that voters who harbour
value V1 are each pcV1 = 0.55 likely to vote for the outcome which is correct that
point of view (X1), while voters who harbour value V2 are each pcV2 = 0.8 likely
to vote for the outcome which is correct that point of view (X2). Figure 13.5
shows the likely distribution of votes among those options, under those

0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.5 60% of voters subscribing to value V1 with pcV1 = 0.55 and 40% of voters
subscribing to value V2 with pcV2 = 0.8.
c ircumstances. In the case depicted in Figure 13.5, the option X1 that is correct

from the point of view of the majority value V1 is defeated by X2, which is the
option that is correct from the point of view of a smaller number of more
competent voters.
A larger group of less competent voters will prevail over a smaller group of
more competent voters only under certain well-defined circumstances that can
be characterized as follows. Let pcS represent the probability of each individual in
the smaller group voting for the option that is correct from the point of view of
her value, and pcL represent the probability of each individual in the larger group
voting for the option that is correct from the point of view of his. Let nS and nL
represent the number of voters in each group. Then, at the limit,8 the correct
outcome from the point of view of the value supported by the larger group is
likely to garner more votes than that supported by the smaller if and only if:
pc L > (nS / nL )( pc S − 1 2 ) + 1 2 . Eq. 13.1
The derivation of that formula can be found elsewhere.9 But here is an example
of its implications. Suppose that the smaller group is 10 per cent smaller than
the larger (i.e. nS = 0.9nL ), and suppose that the individual competence of
each member of the smaller group is once again pcS = 0.8 individually compe-
tent. Then the outcome that is correct from the point of view of the larger
group’s value is likely to prevail (increasingly so, the larger the electorate) so
long as pcL > 0.77.
For the fifth variation on our baseline scenario, let us drop the assumption of
a one-to-one mapping of values onto options.
8 I.e. as the population grows, keeping the ratio nS : nL fixed.

9 Miller 1986, pp. 178–9; Boland 1989, p. 184. The formula in the text rearranges terms to make
our point stand out more clearly.

for X1 for X3 for X4 for X5
0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.6 Five equally large groups, but two groups support X1.
Scenario 5: Take the baseline scenario (with more than two values), but
now let values V1 and V2 both converge on the same option X1 as being the
correct option from each of their points of view; as in the baseline scenario,
let each of the other values point to its own unique option (different from X1)
as being the correct option from its point of view.
Ex hypothesi, in Scenario 5 each value has an equal number of voters support-
ing it, and the voters supporting each are equally probable to vote for the option
that is correct from their own value’s point of view. So, assuming a large elect-
orate, option X1 is likely to get roughly twice as many of the correct votes as any
other option, and hence win comfortably under the plurality rule.
Figure 13.6 provides an illustration of that. Suppose there are many voters,
20 per cent of whom harbour each of five different values V1 through V5.
Suppose each voter has the same probability pcV = 0.55 of voting for the out-
come that is correct from the point of view of her own value (and votes for
other outcomes with equal probability). But suppose that outcome X1 is correct
from the point of view of both value value V1 and V2, while each of the other
values has a differing correct outcome associated with it (X3 being correct from
the point of view of value V3, X4 from the point of view of value V4, and X5
from the point of view of value V5; X2 is not on the agenda). Figure 13.6 shows the
likely distribution of votes among those options, under those assumptions.
Democratically, that is presumably precisely as we would wish. The option
that wins the plurality vote is the option that is democratically-epistemically
correct, given the values of the greatest number of voters. In this case, the
winning option collected its support from supporters of two different values,
V1 and V2—but democratically, no matter.10
For a sixth and final variation on our baseline scenario, let us drop its final
assumption—the assumption that when people vote erroneously from the
10 Think of this as akin to an ‘overlapping consensus’ (Rawls 2001, section 11).


for X1 for X2 for X3 for X4 for X5
0.0 0.2 0.4 0.6 0.8 1.0

vote distribution
Figure 13.7 Five equally large groups, four of which systematically err in the same
direction.
point of view of their own value they are equally likely to support any of the
other options.
Scenario 6: All else is as in the baseline scenario (with more than two
options), except that when voters vote ‘incorrectly’ from the point of view of
their own value, they all vote for option X1.
Among a large electorate, option X1 would on Scenario 6 be expected to be
the hands-down plurality winner, unless voters are extremely competent. In
addition to the votes X1 garners from all those who actually support value V1
and vote correctly in supporting X1, X1 would also collect all the votes of the
voters from all the other groups who vote incorrectly. Even if some other value
is supported by substantially more voters than V1, the outcome X1 that is
favoured by V1 might still prevail thanks to the boost given it by all those ‘sys-
tematic errors’ in its favour. And that can obviously happen, even if the pattern-
ing of the errors is substantially less dramatic than assumed in this scenario.
Figure 13.7 provides an illustration of that. Suppose there are many voters,
20 per cent of whom harbour each of five different values V1 through V5. Suppose
each voter has the same probability pcV = 0.55 of voting for the outcome which
is correct from the point of view of her own value. But suppose that when sup-
porters of values V2 through V5 vote incorrectly from their own value’s point of
view (X2 through X5, respectively), they always vote for option X1 (the option
that is correct from the point of view of value V1). And suppose when supporters
of value V1 vote incorrectly they distribute their votes randomly among options
X2 through X5. Figure 13.7 shows the likely distribution of votes among those
options, under those assumptions.
In Scenario 6 it is stipulated that all groups are of equal size, so there is no
democratic-epistemic reason for preferring any outcome over any other. There
is nothing ‘wrong’ with X1 winning as a result of supporters of other values
systematically erring in its direction. Of course, if V1 had more supporters than
any other value, X1 would be the correct outcome from a democratic-epistemic
point of view for the group as a whole. Then it would be fortuitous, from a
democratic-epistemic point of view, that supporters of other values systemat-
ically support X1 when they vote erroneously from their own point of view. But
suppose that V1 has fewer supporters than some other value, V2 for example.
Then X2 would be the correct outcome from a democratic-epistemic point of
view for the group as a whole—yet it might be defeated by X1 as a result of
supporters of other values systematically erring in favour of the latter. That
would be a worry, from a democratic-epistemic point of view, somewhat akin
to the worry associated with Scenario 4.
13.1.3 The Democratic Upshot
Among all the scenarios just discussed, most are broadly reassuring from a
democratic-epistemic point of view. If there is one value that has more sup-
porters than any other, the option that is correct from that value’s point of view
will, on most of the scenarios discussed, win in a democratic vote. That result
might be upset if differential competence (Scenario 4) or systematic error
(Scenario 6) favours some other outcome. But absent those untoward influ-
ences, the democratically-epistemically correct outcome will almost certainly
prevail within a large electorate.
13.2 DIFFERING PRIORITIES
Notice now that everything that has been said in Section 13.1 can be reapplied,
with simple relabelling, to the case in which what differs is not people’s values
but merely their priorities.
People might actually take the very same position on every issue, but none-
theless disagree over the priority to be attached to each issue. Some voters
think ‘the most important issue’ is national security, while others think it is
unemployment. Those differences in priorities will naturally lead one group of
voters to think that more of the tax revenue should be spent in one way, and
the other group of voters to think that it should be spent in another.
We can model that case in exactly the same way as before. The analogous
baseline scenario would have n voters, with an equal number harbouring
each of i priorities (or ‘urgencies’) which can be represented by U1, U2, . . . , Ui.
Suppose once again that there is one outcome Xi which is uniquely correct,
from the point of view of each priority Ui . Suppose once again that, in the base-
line scenario, each voter is pcU likely to vote for the outcome that is epistemically
correct, given his own priority; and suppose that is the same for all voters.
Suppose once again that the probability of each voter’s voting for the outcome
that is correct from the point of view of his priority is higher than the probabil-
ity of his voting for any outcome that is correct from the point of view of any
other priority. Finally, suppose once again that when a voter makes a mistake
and votes for some option other than that which is correct from the point of
view of his own priority, he has equal probability of voting for each of the
options indicated by any other voter’s priorities.
Then we can run all of the analyses in Section 13.1 exactly as before, substi-
tuting the term ‘priority’ for ‘value’. When we do, the same broadly reassuring
outcomes will emerge: we can be reasonably confident (or better) that the
plurality winner will be that outcome indicated by the priority of the largest
number of voters. We can be highly confident of that on most of the scenarios
under discussion. Once again, there is democratic-epistemic cause for con-
cern only with respect to Scenarios 4 and 6—although even there it is at least
possible that the democratically-epistemically correct outcome might none-
theless prevail.
13.3 DEMO CRATIC COMPETITION OVER

VALUES AND PRIORITIES
In the preceding discussion, we have assumed that the political process simply
aggregates the votes for outcomes that people cast more-or-less accurately in
accordance with their own fixed and exogenously given values and priorities.
But we should note, in closing this discussion, that political competition in
real-world democracies is often an exercise in persuasion designed precisely to
change people’s values and priorities. We can think of this as a form of ‘opinion
leadership’ of the sort discussed in Chapter 11.
Take the case of priorities, where this phenomenon is most well docu-
mented.11 On one highly plausible account, competition between parties is less
over ‘which is the right position to take on issue X, ours or theirs?’ and more
over the question of ‘which is “the issue of this election”, issue X or issue Y?’. The
thought is that one party Px ‘owns’ issue X, while the other party Py ‘owns’ issue Y.
If issue X is seen by voters as ‘the issue of this election’, party Px will win; if Y is
seen by voters as ‘the issue of this election’, party Py will win.
Traditionally, for example, parties of the right ‘own’ the issue of ‘national
security’ whereas parties of the left ‘own’ the issue of unemployment. Of course
11 By the Comparative Manifesto Project: Budge et al. 1987; Klingemann et al. 1994. See also:
Hammond and Humes 1993; Seeberg 2017. Voters’ basic values are probably more stable and
harder for political parties to alter, although there may be more scope for ‘preference-shaping’
understood as policy interventions designed to alter voters’ interests and hence their votes
(Dunleavy and Ward 1981; Dunleavy 1986; 1991; McAllister and Studlar 1989; Saunders 1995;
cf. Stubager 2003).
the manifestos of all parties will have someting to say on all topics. But the
competition is less to persuade voters that your party is right in its approach to
all the issues, and more to persuade voters that the issue that you ‘own’ is the
principal issue on which this election should turn. Succeed in that, and you will
win the election.
Of course, this amounts to an attempt by political parties to make voters’
priorities endogenous to the process of democratic competition. Instead of
treating voters’ priorities as an exogenous fixed given, and trying to satisfy
them, democratic competition organized on this basis tries to change the pri-
orities of voters to fit those of the parties. From a democratic point of view, that
is the ‘wrong direction of fit’.12
Of course, parties have to persuade voters that the issue that they own or the
values that they champion are priorities or values that the voters should share.
No doubt there is a certain amount of behind-the-back subliminal suggestion
at work, as well as a fair bit of hidden-text ‘dog-whistling’.13 Still, the democratic
critique of that practice is substantially blunted if at the end of the day voters
change their priorities or preferences of their own volition, influenced but not
controlled by the parties trying to shape them.
12 See Humberstone (1992) and sources discussed therein for analogous critiques in other
areas of philosophy.
13 Goodin and Saward 2005.
14
Factionalism
Differing Interests
Rousseau taught us to distinguish between particular interests, which are

narrow and sectoral, and the public interest that truly represents the common
good of all.1 Rousseau, and Condorcet following him, strove to suppress fac-
tions on the grounds that they would pursue the former at the expense of the
latter.2 The American Founders offered a counterproposal. They crafted insti-
tutional structures that would hopefully enable the public interest to emerge
out of the interplay of private interests.3 But Marx saw that whole project as
fatally flawed, denying as he did that there is any good common to all classes.4
For him, the distinct interests of the different classes are diametrically opposed
to one another’s.
Marx’s is the most Manichean version of the story and (if only for dramatic
effect) we will follow him in talking in terms of Elites and Masses throughout
this chapter. But notice that we can model other much weaker forms of distinct
factional interests and priorities in precisely the same way.5 Whenever limited
social resources are such that it is impossible simultaneously to pursue each
faction’s differing interests or priorities—which is to say, whenever we are
merely in a situation of pluralism combined with scarce resources6—similar
issues arise in determining whose interests or priorities to promote and whose
to set aside.
The Condorcet Jury Theorem is, as we said in the last chapter, most com-
monly applied to situations where there is some truth that is common to all,
and the ‘correct’ way to vote is the same for all voters. But where there are
distinct factional interests and priorities, with no overarching ‘common good’
1 Rousseau 1762/1997, bk 2, ch. 3.

2 Condorcet 1785/1976, p. 61; 1789/1994, p.170; 1793/1994, pp. 192–4.
3 Madison Federalist nos. 10 and 51. Goodin 1996. 4 Marx and Engels 1848.
5 For a survey of differing bases for group identities see Huddy (2013).
6 Which liberal democrats have long supposed we are. See e.g. Dahl 1967; 1982.
Factionalism: Differing Interests 209
standing above them,7 that will not be the case. Instead, the ‘correct’ way for
members of each faction to vote will vary, depending on the faction they are in.8
In Section 14.1 we extend the previous chapter’s results to show how some-
thing akin to the CJT’s Asymptotic Result still follows where voters are voting on
the basis of factional interests rather than any common good. In the previous
chapter we saw that, among a large number of voters, the majority or plurality
vote almost certainly corresponds to the values or priorities of the largest group
in the population. Here we will similarly show that, when a large group of people
of better-than-random competence vote on the basis of their own factional
interests, the majority vote is likely to correspond to the true interests of the
majority faction.9 As we will show, that probability quickly converges to 1 as
the number of voters increases.
In Section 14.2 we discuss another way in which factions themselves can take
advantage of CJT-like mechanisms, practising epistemic solidarity within each
faction to identify the true interests that members of the faction share, and then
voting as a block in pursuit of the interests thus revealed. Whereas Section 14.2
takes factions as given, Section 14.3 relaxes that assumption.
14.1 A FACTIONAL INTERPRETATION OF THE CJT
14.1.1 Uniform Voter Competence
Suppose, with Marxists and a great many others alike, that people’s interests
diverge. The ‘correct’ way for members of each faction to vote is whatever way
is in that faction’s true interests, which will differ across factions. Suppose, further,
that voters in each faction are minimally competent—better than random—at
determining where their true interests lie and voting in that way. For the pur-
poses of this initial discussion, further suppose that the competence of voters
in that sense is identical across all voters.
Then it follows that whichever outcome is in the true interests of the largest
faction is likely to win a plurality vote among the electorate as a whole. Nicholas
Miller has proven one version of this theorem.10 List and Spiekermann develop
7 Or below them, in the ‘least common denominator’ way of identifying the public interest
(Goodin 1996).
8 The idea of group-specific truths has been discussed previously by: Goldman 1999, ch. 10;
Goodin and Spiekermann 2014; List and Spiekermann 2016.
9 We restrict our discussion here, for convenience, to the case of a society with only two dis-
tinct classes. But everything we say about the workings of majority rule in that two-option case
can easily be extended to the workings of plurality rule in many-option cases.
10 Miller (1986, p. 178) characterizes his result as showing that ‘the victorious position . . . is the
one that would win in the event that all voters were completely informed’ (or as fully informed as
the best available evidence allows them to be, following the logic of Section 5.4.2). He goes on to
the point in a slightly different way.11 Their setup takes a core population
consisting of two factions with two different correct answers (for example:
three members of faction A and two of faction B) and then scales up that popu-
lation by multiplication with natural numbers (for example: six of A and four
of B, nine of A and six of B, and so on). If the voters are all equally competent to
vote for their true interest and independent conditional on their interest, then
List and Spiekermann conjecture that the larger faction is increasingly likely to
win a majority if the population size increases in the way described. This prob-
ability tends to 1 as the group sizes go to infinity.
This result is illustrated in Figure 14.1. That figure compares different initial
population mixes (2:1, 3:2, and 5:4), scaling them up to higher sizes.12 All voters
have competence 0.55 to vote for the alternative in their faction’s interest. The
vertical axis displays the probability of the majority faction winning the vote.13
For comparison, the result of the standard CJT (without factions) is also shown.
The comparison reveals what the theorem claims: the probability of the major-
ity faction winning increases in population size and will eventually converge
1.0
Probability that majority interest will prevail
CJT
0.9
0.8
2:1
0.7
3:2
0.6
5:4
0.5
0 100 200 300 400 500
Population size
Figure 14.1 Probability of victory for the majority faction from majority voting, as
population size increases.
say, ‘once the electorate achieves some minimal size, this probability is greater than the average
competence of all voters, increases further as the size of the electorate . . . increases, and in due
course (though not as rapidly as in the case of the original Jury Theorem) approaches perfection’.
11 List and Spiekermann 2016.
12 Only population sizes that are multiples of the initial population are shown. This is why the
data points are less tightly packed for larger initial populations.
13 Ties are broken by a coin toss.
to 1. However, the closer the ratio of the two factions is to 1:1, the longer the
convergence process takes.
All this goes to show that what Rousseau says, complaining of factions, is
correct. ‘When one of these [factions] is so large that it prevails over all the
rest, the result you have is no longer . . . a general will’; and instead ‘the opinion
that prevails is nothing but a private opinion’ of the majority faction.14 The
point remains, however, that, with large numbers of voters, the position that
prevails is almost certain to be in the true interests of the majority within the
community, at least.15
That is a different understanding of what the ‘correct’ outcome is. Rousseau
may be disappointed that we have substituted ‘what is truly in the interest of
the majority’ for ‘what is truly in the common good of the whole society’. But
liberal democrats of a different stripe would be well pleased. Where there are
two irreconcilable positions and no common ground for compromising
between them, liberal democrats must surely say (subject to substantive con-
straints to prevent the majority from becoming tyrannical) that it is better to
serve what is correctly judged to be the position of the majority than the minor-
ity. That is our ‘moral majoritarianism’ from Section 4.1.4.
14.1.2 Unequal Factional Competence
Notice that that result will often hold, even if members of the larger faction are
individually less competent on average than members of the smaller faction.16
The reason is simple: being larger, the majority faction can afford for a larger
proportion of its members to vote for the wrong position, from their own point
of view. That is straightforward. But what is perhaps more surprising is just how
much less competent members of the larger faction can afford to be, and still
prevail over the minority in this way. Here we use some numerical examples to
demonstrate.
Let us call the larger faction ‘the Masses’ and suppose they have M members.
The option that is in the true interests of the Masses will be written M. Let us
call the smaller faction ‘the Elites’ and suppose they have E members. The
option that is in the true interests of the Elites will be written E. Let pcM repre-
sent the probability of each member of the Masses perceiving his or her own
true interests correctly, and pcE the probability of each member of the Elites
perceiving his or her own (different) true interests correctly. Then we can apply
the formula introduced in the previous chapter. At the limit,17 the position M
14 Rousseau 1762/1997, bk 2, ch. 3, p. 60. For a discussion of this remark in connection with the
CJT, see Grofman and Feld (1988, p. 571) and Estlund (1989, p. 1381).
15 Or ‘largest group’ more generally, assuming the community makes decisions by plurality
rule.
16 These are the sorts of cases evoked by Goldman (1999, pp. 327–8).
17 I.e., as the population size grows, keeping the E:M ratio fixed.
that is in the true interests of the Masses will win a majority vote across the
whole electorate just so long as:
pc M > (E / M )( pc E − 1 2 ) + 1 2 . Eq. 14.1
That is just Equation 13.1, with the notation changed to reflect the current
application.
Here are some sample calculations to show in just how wide a range of cases
that inequality will hold. Based on Equation 14.1, Table 14.1 reports how high
individual competence among Mass voters pcM must be, in order for the Masses
to likely prevail electorally over individually more competent but less numer-
ous Elites.
From Table 14.1 we see that, just so long as they are reasonably numerous
relative to the Elites, the Masses can afford for their members’ individual compe-
tence at judging their own interests to be substantially lower than that of the Elites
at judging theirs, and the position M truly in the interests of the Masses will still
most likely prevail electorally. Suppose for example the Elites are one-fifth as
numerous as the Masses, and members of the Elite are individually pcE = 0.70
likely to judge their true interests correctly. How good do members of the Masses
have to be at judging their own (differing) interests correctly for the position in
the true interests of the Masses to win a majority vote? Just pcM > 0.54.
Higher individual Mass competence is required where Elites are relatively
more numerous or individually more competent. But even in the most extreme
setting in Table 14.1—where there are a third as many members of the Elites
than of the Masses and where individual Elite competence is pcE = 1.0—the
position M that is in the true interest of the Masses will win an electoral major-
ity just so long as the individual Mass competence is pcM > 0.667. In absolute
terms, that may seem awfully high. But seen in relative terms it is not: in the
scenario there in view, it would mean that the individual competence of mem-
bers of the Masses could be thirty-three percentage points lower than that of
members of the Elites, and the position in the true interests of the Masses would
still prevail. In short, this is a pretty strong result.
Table 14.1 Competence threshold Mass voters have to

exceed to make true Mass interest more likely to win
than Elite interest, for various values of pcE and E/M.
E/M = pcE =
0.55 0.60 0.70 0.80 1.00
1/3 0.517 0.533 0.567 0.600 0.667

1/5 0.511 0.520 0.540 0.560 0.600
1/10 0.505 0.510 0.520 0.530 0.550
1/100 0.5005 0.501 0.502 0.503 0.505
14.2 EPISTEMIC SOLIDARIT Y AND BLO CK VOTING
The next question that naturally arises is whether there is anything that
larger but less well-informed Masses can do to avoid defeat by smaller but
better-informed Elites, in cases where the Masses cannot (as in Section 14.1)
count on prevailing by sheer force of numbers alone.18
There is: they can organize. That is the traditional way for the Masses to over-
come the power of Elites, after all. And a version of that strategy can be used by
the Masses to overcome at least a certain amount of false consciousness as to
where their true interests lie, as well. The form of ‘organizing’ we have in mind,
here, is for the Masses to collectivize epistemically, pooling information about
what is in their interests and how best to pursue them.
The thought is simply that the CJT can work within factions, as well as (as in
the previous section) across them. Ex hypothesi, all members of each faction
have identical interests, different from the interests of members of other fac-
tions. If members of the faction are individually better than random at discern-
ing what those shared interests within the faction are, then whatever a majority
of members of that faction say those interests are is quite probably correct—that
probability increasing as the number of members of the faction increases.
That fact forms the basis of the strategy we dub ‘epistemic solidarity’. The
solidarity in question is practised strictly within factions. Our suggestion is
that, if in doubt as to what the true interests of your faction are, a good way to
find out might be to ask the other members of your faction what they think
they are, and then to vote in whatever way a majority of them say. The practice
of epistemic solidarity is thus a two-step process. It involves, first, epistemic
pooling of information with others sharing the same true interests, and second,
block voting by everyone in that same situation in line with the result of that
information-pooling.19
18 How big a window is this, in which the strategy here discussed is both needed and will work?
Maybe not very. Take the case discussed above, where E/M = 0.2 and pcE = 0.7. The position in the
interests of the Masses will win, according to Equation 14.1, whenever pcM > 0.54. And the strategy of
epistemic solidarity discussed in the present section will work only if pcM > 0.5. So the strategy is
useful (both needed and successful) only if 0.5 < pcM < 0.54. While that seems like a small window,
it may well be that lots of real-world cases lie within it. That is to say, it may well often happen that
members of the Masses are individually better than random at judging their interest but only by a
small margin.
19 This approach was foreshadowed by Estlund (1989, p. 1318) when contemplating what would
happen ‘if . . . the “party line” within groups is determined by majority rule within the faction’;
then, he suggests, ‘the Condorcet jury theorem can perhaps be brought to bear to show that group
competence can exceed the average individual competence’. Looking further back into history,
something like it was arguably foreshadowed by Rousseau’s (1762/1997, bk 2, ch. 3, p. 60) initial
discussion of factions, when he wrote that ‘when factions arise, small associations at the expense
of the large association, the will of each one of those associations becomes general in relation to
its members and particular in relation to the State’. (A similar rule was employed for very different
purposes within the Problem Solvers Caucus in the US Congress in 2017; see Washington Post
Editorial Board 2017b.) Rousseau, and Estlund following him, think that the principal problem
with that is ‘there can then no longer be said to be as many voters as there are men, but only as
This strategy will work only within limits. First, people have to be relatively
confident with whom they share the same interests, even if they are unsure
exactly what those interests are. Second, the people in the group thus identified
have to be more likely to be right than random regarding the content of those
interests.20 Third, the less-competent Masses must be more numerous than the
more-competent Elites. How much is required in each dimension in order for
the strategy to work is a function of how much is present in both of the other
dimensions.
We will discuss just how sensitive the strategy is to relaxation of the first two
of those assumptions shortly. But first let us consider what happens in the base-
line case where all those assumptions hold.
First suppose that the Masses, and they alone, practise ‘epistemic solidarity’.
That is to say, they take a pre-election ballot among themselves to decide
which option in the upcoming election is in their true interests. In that
pre-election ballot, members of the Masses vote independently of one another.
But then, in the subsequent election, each member of the Masses votes for
whatever option was the majority winner of that pre-election ballot among
the Masses themselves.
What will the result of that be? Well, assuming members of the Masses are
individually better than random at correctly identifying what is truly in the
Mass interest (and the other CJT assumptions are met), the majority among a
large number of such voters is almost certain to be correct. And since, ex
hypothesi, there are more members of the Masses than the Elites, if all members
of the Masses vote in line with the true interests of the Masses, that option is
almost certain to win.
The same is true if Elites as well as Masses practise epistemic solidarity within
their own faction. Assuming a large number of members in each group (so the
law of large numbers applies for both), by the same logic the vote share for the
option truly in the interests of each faction will be roughly in proportion to
their fraction of the population. And since the Masses are more numerous, the
position M that is truly in the interests of the Masses will win a majority across
the electorate as a whole.
Figure 14.2 provides an illustration. For these purposes, we take a concrete
numerical example: a population of 1,200,000 voters, of whom 200,000 are
Elite and 1,000,000 are Mass voters. For the purposes of that example, the indi-
vidual competence of members of the Elites is taken to be pcE = 0.70, and that of
members of the Masses pcM = 0.51.
many as there are associations’. Insofar as the position of each association is (as Estlund puts it)
‘determined by majority rule within the faction’, and as long as voters within each faction vote
independently of one another per the CJT Independence Assumption, you have as many inde-
pendent assessments of what is correct for that faction as you have voters within that faction.
20 Both of which are to say: false consciousness must not run too deep.
1,200,000 Masses for м

Elites for м
1,000,000 Masses for ɛ
Elites for ɛ
Vote distribution
800,000
600,000
400,000
200,000
0
None only Masses Both only Elites
Pooling
Figure 14.2 Approximate expected vote distribution, E = 200,000, M = 1,000,000;

pcE = 0.7, pcM = 0.51.
With those population sizes and voter competences, we know from Equation
14.1 that the position M that is in the true interests of the Masses will be defeated
without epistemic solidarity. That happens, basically, because too many mem-
bers of the Masses will ‘mistakenly’ vote for option E which is in truth the wrong
option from their point of view. We can see this in the first column of Figure 14.2,
where the white-hatched part of the bar represents members of the Masses
‘mistakenly’ voting for E. That is what is primarily responsible for E getting over
half of the votes, when neither group practises epistemic solidarity.
But as long as the Masses practise epistemic solidarity, the position M that is
truly in their interests is likely to win. It is likely to win by a somewhat larger
margin if the Elites do not also practise epistemic solidarity (the second column
in Figure 14.2), because then M attracts not only all the votes of all members
of the Masses but also a few from members of the Elite who vote that way
‘mistakenly’ (shown as the unhatched grey portion of the column). But even if
both groups practise epistemic solidarity, the position M that is truly in inter-
ests of the Masses is still likely to win by a wide margin.
What happens, however, if the Elites practise epistemic solidarity while the
Masses do not? There the outcomes can vary. In certain cases, the Elites might
in that way manage to triumph over the Masses and install the option E that is
in the Elites’ true interests as the election winner. The scenario depicted in
Figure 14.2 is one such case, as the final column there shows. If the Elites prac-
tise epistemic solidarity and the other assumptions hold, all of the members of
the Elite vote for the option E that is the correct one from their point of view.
But if the Masses do not practise epistemic solidarity, a substantial number of
them are likely to vote ‘mistakenly’ for E as well—enough, in the final column
of Figure 14.2, to make it the majority winner.
Thus it can happen that Elites defeat Masses if Elites, and Elites alone, p ractise
epistemic solidarity. The example in Figure 14.2 is proof of that. The big ques-
tion, however, is whether that is likely to be a typical case or a very special one.
In general, it can be shown that, at the limit, the Elites will win in this way if and
only if pcM < E / (2 M ) + 0.5 which is just a special case of Equation 14.1 where
pcE = 1.0.
For some numerical examples to illustrate just how serious that risk is, just
look at the last column of Table 14.1. That describes how high the individual
competence pcM of Mass voters must be for the option M that is in their true
interests to prevail, despite the fact that the Elites practise epistemic solidarity
and the Masses do not. As we see there, if the Elites are relatively small compared
to the Masses, the Masses do not need to be very individually competent to
defeat them even if the Elites practise epistemic solidarity and the Masses do
not. If there are one hundred members of the Masses for every member of the
Elites, for example, the position truly in the interests of the Masses prevails just
so long as the individual competence of members of the Masses is pcM > 0.505.
And even if the Elites are ten times as numerous as that, the individual com-
petence of members of the Masses needs only be pcM > 0.55 in order for the
position truly in the interests of the Masses to prevail.
In short, epistemic solidarity is a much better way for the Masses to beat
the Elites than vice versa. Even if the Elites practise epistemic solidarity while
the Masses do not, the circumstances in which the Elites will win in that way are
strictly limited.
14.3 WHO’S WITH US?
The model of epistemic solidarity just sketched assumes that people are
completely certain as to who shares the same interests with them, but that they
are uncertain what exactly the content of those interests are. That is pretty
unrealistic, however. More commonly, people will be more-or-less uncertain
about both.
Furthermore, those uncertainties might not be uniformly distributed.
Identifying with whom they share the same interests might well be harder for
the Masses than the Elites. Not only are members of the Elite (ex hypothesi)
individually more competent in judging what is in their interest. They might
also have a better idea who shares the same interests: they are socially more
mobile and better networked; they ‘know people who know’; they and their kind
dominate the public discourse. In all these ways members of the Elite might be
better able to identify one another and vote for their interests as a block. And as
we have seen, if the Elites coordinate and vote as a block while the Masses do
not, they can sometimes in that way prevail over the disorganized Masses.
In this section we will introduce uncertainty as to who shares the same interests,
and let the extent of that uncertainty and responses to it vary across the two
groups, to see to what extent that upsets the main results of the previous sec-
tion. The analyses will necessarily be more complicated than the previous ones.
But, to foreshadow our principal conclusion: the result remains basically the
same. As long as the Masses engage in epistemic solidarity, the position truly in
their interests will prevail in all but the most extreme settings.
14.3.1 Differential Abstention from Epistemic Solidarity
If people are uncertain with which group their own interests align, they can
respond in at least two ways. One is to abstain from practising epistemic soli-
darity with either group. Another is to practise epistemic solidarity with which-
ever group they think their own interests are most likely to align. These two
responses will be considered in turn.
The basic results reported above are not, in general, terribly sensitive to some
members of the Masses abstaining from epistemic solidarity. Of course if all of
them abstained, that could be a problem. But the Masses can afford for quite a
few of them to abstain, as can be seen from size of the grey-shaded area in
Figure 14.2 for those particular parameter settings.
In our running example of pcM = 0.51 and pcE = 0.70 and E / M = 0.20 , the
position truly in the interests of the Masses will prevail so long as just over
6 per cent of the Masses practise epistemic solidarity and none of the Elite do. If
half of the members of the Elite practise epistemic solidarity and all else is the
same, then the position truly in the interests of the Masses will prevail so long
as more than 12.2 per cent of the Masses practise epistemic solidarity. Even if
80 per cent of the Elite practise epistemic solidarity, the position in the true
interests of the Masses will prevail so long as more than 15.9 per cent of the
Masses practise epistemic solidarity.
14.3.2 Differential Group Selection Competence
Thus, the results are not terribly sensitive to abstentions from epistemic soli-
darity on the part of members of the Masses. But might they be more sensitive
to members of the Masses being more likely than the Elite to mistake their true
type and practise epistemic solidarity with the wrong group?
Suppose that all Elite-type individuals have the same probability pgE > 0.5 of
correctly identifying with which group to align, and that all Mass-type individuals
have probability pgM > 0.5. Call this the ‘group selection competence’ of the Mass
type and Elite type, respectively. Further suppose the population is exhaust-
ively partitioned into two groups, one composed of self-assessed members of
the Masses and the other self-assessed members of the Elites. Finally, suppose
just as before that all Mass-type individuals have probability pcM > 0.5 and all
Elite-type individuals have probability pcE > 0.5 of being correct in their per-
sonal assessment of their own interests in the case at hand.
Based on 1,000 simulations for each data point, Figure 14.3 plots the propor-
tions of Mass majorities as a function of group selection competence, which for
purposes of this initial exercise we assume to be equal for both types, pgM = pgE.
The number of Elite types is twenty-one and of Mass types one hundred. The
former have competence pcE = 0.7 and the latter pcM = 0.55. The circle markers
show the probability of a Mass majority when only the self-assessed Elite group
pools, the diamonds when only the self-assessed Mass group pools, and the
stars when both groups pool.
If the Elites alone practise epistemic solidarity, the position in the true inter-
ests of the Masses is increasingly unlikely to prevail as group selection compe-
tence increases. In the limiting case of group selection competence pgM = pgE = 1.0,
the probability of a majority for the position that is in the true interests of
the Masses is only around 16 per cent (as we see from the rightmost circle in
Figure 14.3). But whenever the Masses practise epistemic solidarity, the larger
size of the self-assessed Mass group leads increasingly reliably to a victory for
the position truly in the interests of the Masses as group selection competence
increases. And that is equally true whether the Masses are alone in practising
E = 21, M = 100, pcE = 0.7, pcM = 0.55

1.0
0.8
Probability of Mass majority
0.6
0.4
0.2
0.0
0.5 0.6 0.7 0.8 0.9 1.0
Group selection competence pgM = pgE
both pooled only Elite only Mass
Figure 14.3 Probability of Mass majorities as a function of group selection competence.

epistemic solidarity or whether both Masses and Elites engage in it (the diamond
and star markers trace out essentially the same line in Figure 14.3).
But as we have said, it might be easier for members of the Elites to identify
one another than it is for members of the Masses to do so. What would happen
if group selection competences were not equal for both groups, but rather
favoured the Elites?
To explore this question, let us hold group selection competence constant for
the Masses at pgM = 0.55, while letting the group selection of the Elites pgE vary.
The results are displayed in Figure 14.4, for the case once again of M = 100, E = 21,
pcE = 0.7, and pcM = 0.55.
In that scenario, once again, the position truly in the interests of the Masses
is very likely to win if the Masses alone practise epistemic solidarity, and the
position truly in the interests of the Elites is very likely to win if the Elites alone
practise epistemic solidarity (and are good enough at selecting the right
pooling group). In that respect, Figure 14.4 is similar to Figure 14.3. But notice
what happens when both Elites and Masses practise epistemic solidarity (indi-
cated by the row of star markers in Figure 14.4). The more competent Elites
are at recognizing their true type, the more the Elites benefit from practising
epistemic solidarity and the less likely a victory for the position truly in the
interest of the Masses becomes.
E = 21, M = 100, pcE = 0.7, pcM = 0.55, fixed pgM = 0.55

1.0
0.8
Probability of Mass majority
0.6
0.4
0.2
0.0
0.5 0.6 0.7 0.8 0.9 1.0
Group selection competence Elites pgE
both pooled only Elite only Mass
Figure 14.4 Probability of Mass majorities as a function of the Elite group selection
competence, Mass group selection competence fixed.
Notice that, for the example given in Figure 14.4, the row of star markers
crosses the 0.5 threshold. This shows that if the small Elites are very good at
selecting their pooling group they become more likely to win than the much
larger Masses. But this only happens if the Elites have a substantially better
group selection competence than the Masses. Of course, those results would
change if the parameters were set differently—whether the Elites or the Masses
tend to prevail depends on the details. But for a wide range of the most realistic
scenarios, it seems unlikely that the position truly in the interests of the Masses
will be defeated, so long the Masses practise epistemic solidarity at all.
In our earlier discussion, we showed that that was true despite members of
the Masses abstaining from the practise of epistemic solidarity at a higher
rate than members of the Elites. With this last set of figures we have shown
that it is also true if the Masses practise epistemic solidarity less competently
than the Elites.
14.3.3 Strategic Leadership and Coordination
Success for the Masses depends, as we have seen, on finding their epistemic
peers. They might be able to find their peers reliably enough based on their
individual competence to do so. The situation gets trickier, however, if there
are many possible policies and the Masses need to coordinate on one—a problem
we already touched upon in Section 4.2.3. Suppose several policies are equally
good to endorse for the Masses, but none sticks out as special or salient. If the
Masses fail to coordinate on one, or if the Elites manage to split the Masses
using a divide et impera strategy, then the Masses are likely to fail in their pursuit
of their interests.
The challenge the Masses face is akin to Keynes’s ‘beauty contest’.21 Beauty con-
tests are games in which the winner is chosen among those who vote with the
plurality (largest group). For example, the task could be to choose the most beau-
tiful artwork among ten.22 The winners are all those who vote for the most popular,
most often-chosen artwork. To win this beauty contest, the first-order strategy is
to work out the likely aesthetic judgements of the other participants, anticipate
the most popular artwork, and vote for it. The second-order strategy is to work
out what most people think most people think, and so on.
Because the outcome of beauty contests depends on expectations, leadership
to create or influence expectations comes into play. A leader can provide a focal
point that allows individuals to coordinate on one option. This can help the
21 Keynes (2007 [1936]), chapter 12, section V.

22 Keynes took his cue from a newspaper contest in which participants had to vote for photos
of women, trying to anticipate which woman is seen as the most beautiful by most. We avoid this
sexist setting for obvious reasons.
Masses to achieve their goals. However, there are various strategic incentives
for leaders.23 For example, non-coercive leaders can promote their own prefer-
ences by providing biased information, benefiting from the voters’ needs to
coordinate.24 These forms of leadership-as-coordination are obviously very
different from the opinion leadership we discussed in Chapter 11. Nevertheless,
the epistemic analysis of factionalism shows that leadership-as-coordination is
an important piece of the epistemic puzzle.
23 As Dewan and Myatt (2008), Landa and Tyson (2017), and others have shown.
24 Coercive leaders can of course also effectively enforce options that are not in the public’s best
interests because the threat of non-coordination is effective.
Part IV
Structures of Government
In this part of the book, we adduce some lessons from the Condorcet Jury
Theorem for the structure of government. We begin with the threshold question
of whether a system of government should be a democracy or an ‘epistocracy’
(‘rule by the experts’). Insofar as good epistemic performance is our goal—and
surely it should be one of our goals—then it seems natural to suppose that
epistocracy is to be preferred. But as Chapter 15 shows, the CJT does not drive
us particularly strongly in that direction. Indeed, if you believe in ‘learning
from experience’ (as advocates of expansions of the democratic franchise have
historically done), then CJT reasoning would actually argue for a wide demo-
cratic franchise.
Having thus seen how the CJT points us towards democracy, the question
then becomes what form that should take. From the CJT perspective a
natural inference might seem to be that direct democracy is to be preferred
to representative democracy, since larger numbers of voters are always epi-
stemically more reliable than smaller ones, other things being equal. But as
Chapter 16 shows, the epistemic difference between a mass electorate num-
bering millions and a representative assembly numbering several hundreds
is not great; and there might be other epistemic considerations, such as the
advantage of deliberation in smaller groups, that make up for the modest gap
between them.
The next two chapters discuss, respectively, institutional hindrances and
institutional aids to epistemic success. In Chapter 17’s discussion of the former,
one recurring theme is that of ‘epistemic bottlenecks’.1 Since on CJT assump-
tions a larger group of voters is always epistemically better than a smaller group
(all else equal), any institutional arrangement that funnels the views of a larger
number of people through a body composed of a smaller number of independ-
ent decision makers will, on the face of it, constitute an ‘epistemic bottleneck’
1 In Vermeule’s (2009a, pp. 50–3) felicitous phrase.

that compromises the epistemic performance of that institution overall. There

may sometimes be countervailing considerations that make up for that, epi-
stemically. But that is the baseline presumption from which any CJT-style
assessment begins: epistemic bottlenecks should be presumed to be bad things,
in the absence of any countervailing considerations.
In Chapter 18 we discuss institutional aids to epistemic success. Some of
those measures work by making the decision situation more truth-conducive.
Others work by increasing the independence or sincerity or, less promisingly,
competence of voters. One suggestion that recurs across several of those dis-
cussions is for smaller and more deliberative groups to be used to craft and
winnow alternatives among which larger groups are then to decide.
15
Epistocracy or Democracy
Insofar as ‘getting it right’ is our goal, deferring to those who are more likely to
get it right is the most intuitively appealing strategy. As Montesquieu says,
‘decisions go by majority vote, but . . . it would be better to follow the minority
opinion, for there are very few good minds, and everyone agrees that there is an
infinite number of bad ones’.1
Assuming that some people are more likely to get it right than others—and
assuming that they are independently identifiable (a big assumption, which we
will here query2)—letting those people make decisions on our behalf would
seemingly be the best way for us to reach the right decisions. We follow Estlund
in calling that scheme of Rule by the Wise ‘epistocracy’, and in contrasting that
with democracy understood as Rule by the Multitude.3
Contrary to the seeming implications of his own jury theorem, Condorcet
was himself inclined towards a version of the Rule by the Wise. In the Essai
he explicitly cautions against hyper-democratic interpretations of his jury
theorem’s results:
A very numerous assembly cannot be composed of very enlightened men. It is
even probable that those comprising such an assembly will on many matters
combine great ignorance with many prejudices. Thus there will be a great number
of questions upon which the probability of the truth of the vote of each voter will
be below 1/2. It follows that the more numerous the assembly, the more it will be
exposed to the risk of making false decisions.
Now since these prejudices and this ignorance can exist in relation to very import-
ant matters, it is clear that it can be dangerous to give a democratic constitution to
an unenlightened people. A pure democracy . . . would only be appropriate to a
people much more enlightened, much freer from prejudices than any of those
known to history. . . .
1 Montesquieu 1721/1973, Letter 86, p. 167. Cf. Dahl 1989, ch. 5.

2 See Section 15.2. ‘Who will know the knowers?’ is of course Estlund’s (1993, pp. 71–2,
84–92; 2008, pp. 30–1) principal objection to epistocracy. See similarly Viehoff (2016).
3 Estlund (1997, p. 183; 2008, pp. 277–8 n 16) apologizes for the fractured Greek, but we follow
him in it.
[W]here [a smaller] assembly can be formed in such a way that there is a very great
probability of its decisions being true, then there is just ground for men less
enlightened than its members to submit their will to the decisions of this
assembly.4
The American Founders acted on that same thought, with James Madison’s
Federalist no. 58 arguing for limiting the size of the proposed House of
Representatives on the grounds that, ‘the larger the number, the greater will be
the proportion of members of limited information and weak capacities’.5
It is important to note that logically this constitutes a case for epistocracy,
not aristocracy or its contemporary equivalents. It is a case for Rule by the
Few, to be sure; but the qualification is one of intellectual enlightenment, not
social status.6 Most of all, Condorcet says, we ought to select the Few who will
rule over us purely on the basis that they are substantially immune to ‘preju-
dices and ignorance’ (as he puts it).7 The rich as well as the poor, the noble as
well as the peasant, all have prejudices peculiar to their social location. So too
do the well-educated, who were Condorcet’s preferred rulers.8 If anything,
Condorcet’s criterion for selecting rulers might point most strongly towards
the socially unconnected and free-floating.9
Let us set all those mere staffing issues to one side, however. The more fun-
damental issue is whether Rule by the Few—even if properly chosen—is actu-
ally to be epistemically preferred. The CJT says that more voters are typically
better, from an epistemic point of view—subject of course to the CJT’s assump-
tions being satisfied.10 Furthermore, extensions of the CJT tell us that, when
adding more voters, those additional voters do not have to be individually as
4 Condorcet 1785/1976, pp. 49–50.

5 Madison, Federalist no. 58 (1788/2003, p. 285). The theme echoes down the centuries.
Warning against the increasing power of the lower classes, Bagehot (1867/2003, p. 278) writes:
‘their supremacy, in the state they are now, means the supremacy of ignorance over instruction
and of numbers over knowledge’. Hayek (1960, p. 110) insists that majority decisions ‘are bound . . . to
be inferior to the decisions that the most intelligent members of the group make after listening
to all opinions’.
6 ‘Enlightened men, free from prejudice’, in Condorcet’s (1785/1976, p. 62) formulation.
Although himself of aristocratic origins, Condorcet was no friend of aristocracy: he denounced
that, at the same time as arguing for extending voting rights to women (Condorcet 1790/1994,
p. 337).
7 And there might be a trade-off between those two desiderata, insofar as education brings
with it prejudices of its own; see sources in the next footnote.
8 Estlund 2008, ch. 11; Bovens and Wille 2010; 2017. Cf. Condorcet (1785/1976, pp. 62–3).
9 See Frazer (2014) for an elaboration of such an argument. Note the ‘well connected’ have a class
consciousness all their own (Calhoun 2002), and presumably have always done so. Condorcet’s
logic would lead to the very opposite of the ‘rule by the socially well-connected’ that Anti-
federalists complained was embodied in the US Constitution. See e.g. Letter III of Brutus
(1787/2003, pp. 456–8); cf. the reply in Federalist no. 57 (Madison 1788/2003, pp. 277–82).
10 Maybe not much better, once we already have a large number of voters already—maybe the
epistemic gains to the community are not worth the costs those additional voters incur in turning
out to vote (Brennan 2011a). But that is a separate issue. From a purely epistemic perspective, more
is better.
Epistocracy or Democracy 227
competent as the already existing voters in order for their addition to increase
the probability that the majority of the new larger group will vote for the c orrect
alternative.
This chapter is thus devoted in a way to adjudicating a dispute between
Condorcet and Condorcet. Which is epistemically preferable, when,11 and by
how much: Rule by the Many (democracy) or Rule by the Few epistemically
chosen (epistocracy)?12
15.1 BEATING THE SMARTEST GUY IN TOWN
Let us start with a simple demonstration of the superior epistemic power of

large groups of individuals, who are themselves only slightly better than ran-
dom at choosing the correct alternative, over single or small groups of more
competent individuals.
15.1.1 Beating the Smartest Single Guy
Suppose there is one person, instantly recognizable to everyone, who is the

‘smartest guy in town’. Call him ‘individual 1’ and let his individual voter com-
petence be represented by pc1. Suppose there are n − 1 other individuals in his
town, each with identical individual competence pcREST > 0.5. Ex hypothesi, the
smartest guy is smartest, so pc1 > pcREST for each of the others in town, taken
individually. But what is the probability that a majority vote among all of the
n − 1 others might be more likely to be correct than the smartest guy in town
would be, had he been allowed to decide all on his own?
Here is sample calculation bearing on that question. Suppose n = 1,000,
which is actually a pretty small electorate. Suppose that the smartest guy’s
competence is pc1 = 0.9, which would be extraordinarily high. In that case the
11 I.e. ‘in what decision situations?’. Of course, some people are substantially more competent
on any given topic than others (Section 3.1.2); but as long as mean competence among the elect-
orate as a whole is better than random, and the other conditions of the CJT are met, it will ordin-
arily be epistemically preferable to let the decision be made by the electorate as a whole (cf. Peter
2016). But if the decision situation is systematically misleading for some people but not others
(Section 4.6.1), it might be epistemically preferable for the decision to be made by those in the
epistemically more favourable decision situation.
12 That is the way the options are usually posed, and we will in this chapter stick with that
framing. Notice however that Arneson (2016, p. 156) poses an interesting hybrid model: ‘In pass-
ing, note that it is not clear why a knowledgeable elite concerned to base decisions on all relevant
information could not gather the dispersed bits of knowledge possessed by the less knowledgeable
by opinion polls or other social science techniques of information collection.’ Or as Weinstock
(1999, p. 12, emphasis in original) had put it in an earlier unpublished paper, ‘this still hasn’t given
us a reason to engage in democratic decision-making, as opposed to democratic discovery’.
probability of a majority being correct is higher than that of the smartest guy’s
being correct, just so long as the individual competence of each of the other
townspeople is pcREST > 0.520. Make the electorate larger—10,000, say (still only
the size of a small town)—and the required level of individual competence
required for the majority to be more likely to be right than the smartest guy
drops to pcREST > 0.5064.
15.1.2 Beating the Smartest Clique of Guys
So far we have been talking about beating ‘the smartest guy’, in the singular.
Do the results just reported turn heavily on that feature of the example, or
would something similar be true even if we were talking about a ‘clique of
smartest guys’?
We can address that question with the aid of the Grofman–Dummkopf–
Witkopf theorem mentioned in Section 2.4. Using that formula, we can con-
struct a graph showing how individually competent members of a larger group
(for varying values of group size) would need to be, in order for that group to
be as collectively competent as a smaller group of epistocrats of fixed compe-
tence.13 For the purposes of this example, let us suppose that the clique of
‘smartest guys’ numbers nSMART = 100 and that each of them is individually
pcSMART = 0.7 likely to be correct.
0.60
0.58
0.56
pcREST
0.54 Isocompetence with nSMART = 100

and pcSMART = 0.7
0.52
0.50
0 10,000 20,000 30,000 40,000 50,000
nREST
Figure 15.1 Isocompetence curve showing points at which a group with nREST and pcREST
has the same epistemic performance as a group with nSMART = 100 and pcSMART = 0.7.
13 Grofman (1975) suggests using it in precisely this way.

Figure 15.1 shows how individually competent pcREST the rest would have to
be, in order for a majority vote among them to be more likely to be correct than
a majority vote among the ‘clique of smartest guys’ alone. As Figure 15.1 shows,
that depends on the number of voters there are among ‘the rest’. But if ‘the rest’
numbers nREST = 10,000, for example, each of them would need to be only
pcREST > 0.522 individually competent in order for a majority vote among them
to be more likely to be correct than a majority vote among the hundred ‘smart-
est guys’ who are pcSMART = 0.7 likely individually to be correct.14
15.2 MODELLING THE EPISTEMIC EFFECTS

OF EXPANDING THE ELECTORATE
In assessing the epistemic effects of increasing the number of voters, we can

distinguish the following different scenarios.
15.2.1 Each Individual’s Competence Level Is Known
First, suppose—unrealistically—that the competence of every potential voter is

known in advance. A social engineer could then start by first enfranchising the
most competent voter, and then continuing to add voters in decreasing order of
individual competence.15
Two questions arise, in the process. First, under which conditions does
the group competence increase monotonically? Second, if the group compe-
tence does not increase monotonically, which group size maximizes group
competence?16
In the classic CJT setup with a homogeneous individual competence pc > 0.5
across all voters, increasing the group’s size is always epistemically beneficial,
and the optimal group size is always the largest group size possible. However,
if competence is heterogeneous, then, when adding voters in order of decreas-
ing competence, it is possible that not all additional voters would increase
group competence.17 And if at some point individual competence falls below
14 Thus bearing out Lippert-Rasmussen’s (2012, p. 246) observation that ‘there is nothing, logically
speaking, that prevents it from being the case that the group of people that collectively knows best
what the right decisions are has none of those individuals who, individually speaking, knows
best, i.e., no epistocrats, as its members’.
15 Strictly speaking, it is useful to think in terms of adding pairs of voters (rather than single
voters)—that avoids changing from odd to even group sizes and the corresponding issues with
ties that can then arise.
16 Estlund 2012.
17 Karotkin and Paroush (2003) state a sufficient (but not necessary) condition for a strict
monotonic increase of group competence when adding pairs of voters.
(or even just quickly approaches) 0.5, it can be better to stop adding voters and
not include the whole population.18
Note well, however, that these outcomes rest on an entirely unrealistic
assumption—namely, that we have perfect knowledge of all individual com-
petence values and can add voters in that specific order from high to low
competence. If we really did know that much about all voters, and if we did
know that some voters are barely competent or even incompetent, then we
might have a case for restricting franchise.
15.2.2 Only Average Individual Competence Is Known
In reality, the most we can ordinarily know is the average competence of the
whole group of potential voters. That gives rise to the second (and more realis-
tic) scenario to be considered.
Assume that we draw voters randomly from that population to form our
electorate. It can then be shown that, as long as the average population compe-
tence from which voters are drawn is a value above 0.5 and each individual
voter is also at least somewhat better than random, then larger groups of voters
can always be expected to be more competent than smaller groups of voters.19
15.2.3 Enfranchising Batches of Voters with Heterogeneous

Individual Competence
More realistic, yet again, is a third sort of scenario. Political realities are
typically such that new voters are added to the electorate only in ‘batches’—all
property holders, all adult males, all adult females, all persons over age sixteen,
and so on. Furthermore, individual competence is inevitably heterogeneous
within those batches. Some individuals within the batch being added have rela-
tively high individual competence, while others have low (perhaps even worse-
than-random) individual competence. Yet the most we can ordinarily know,
with any confidence at all, is the average competence across the batch of poten-
tial new voters to be added to the existing electorate.
If there are a great many people in the batch being added, and mean individ-
ual competence within that batch of people is safely over pc > 0.5 , then adding
that batch to the existing electorate can often improve collective competence.
That is to say, the collective competence of the new, expanded electorate can
exceed that of the old, smaller electorate. And that can be true despite the fact
that some people in the heterogeneous newly added bunch have individual
competence pc < 0.5.
18 Karotkin and Paroush (2003) provide an algorithm to determine the optimal group size.
19 Berend and Sapir 2005.
15.2.4 Enfranchising Voters with Heterogeneous

Knowledge Bases
One of the great benefits long claimed for extending the franchise is that doing
so will lead to the inclusion in the electorate of people with different views,
perspectives, experiences, and knowledge bases than found among the exist-
ing, more restricted electorate. In part that is urged as a matter of fairness to
those whose interests and perspectives might otherwise have been overlooked.
And that is obviously an important consideration, as well. But for the purposes
of this book it is the epistemic consequences that are of principal concern.
Insofar as increasing the size of the electorate also increases the diversity of
the electorate in relevant respects, that tends to improve the epistemic per-
formance of the electorate.20 The outcome is obvious when the newly enfran-
chised voters bring new information to the table. It can also be the case, as we
have shown in Section 7.2.3, when what the newly enfranchised voters bring to
the table are different cognitive models of the world. And it is also typically the
case even when what they bring to the table are different cues and heuristics,
insofar as those common causes are sufficiently independent of one another.21
15.3 THE EPISTEMIC LO GIC OF ENFRANCHISING

THE LESS COMPETENT
Next let us rehearse why including some additional voters, even if they are
individually less competent than the existing voters, might actually improve
the epistemic performance of the group as a whole.
According to the formula at the heart of the CJT, the probability Pn that the
majority of a group will vote for the correct alternative is a function of two
things.22 One is the number of voters, n. The other is (in the form of the CJT
extended in Section 3.1.1 to heterogeneous voters) the mean competence pc of
voters in the group individually voting for the correct alternative themselves.
What we are doing, when adding additional voters who are individually less
competent than the already existing ones, is increasing n at the same time as
decreasing pc . The former has the effect of increasing the probability of the
majority vote in the new, expanded group choosing the correct outcome, Pn;
the latter has the effect of decreasing that probability. But since competence
20 Landemore (2013b, p. 104; see similarly 2013a, p. 217) asserts this claim, albeit cautiously: ‘Of
course, this assumption that cognitive diversity positively correlated with numbers will not always
be verified but it is generally more plausible than the reverse assumption that cognitive diversity
increases as the number of people go down.’
21 As per the multiple-common-cause implications of the Best Responder Corollary to the CJT
discussed in Section 5.4.3.
22 The formula is given in Section 2.4; Section 3.1.1 discusses its extension to mean individual
competence for heterogeneous voters.

1.00 1.00
pc = 0.6
0.95 pc = 0.52 0.95
0.90 0.90
Pn
0.85 0.85
0.80 0.80
0 100 200 300 400 500 1,000 5,000 10,000
n
Figure 15.2 Group competence as a function of n, given that the first fifty voters have
pc = 0.6, while all others have pc = 0.52.
grows quickly in n, the former effect can often dominate the latter, just so long
as pc remains above 0.5.
Remember, our epistemic goal is to maximize Pn. That, emphatically, is not
necessarily the same as maximizing pc .23 Adding individually less competent
voters will always drag down average individual competence pc , but that does
not always inevitably drag down collective group competence Pn. Sometimes it
can improve it.
That has already been shown, in one way, in Figure 15.1. But here is a small
numerical example to drive home the point. Suppose the original group con-
tained fifty voters with homogeneous individual competence pc = 0.6. Now
suppose we begin adding more and more less competent voters with pc = 0.52
to the electoral rolls. The curve in Figure 15.2 shows the effect of doing so on
group competence.
To see the effects for both small and large groups we have separated the hori-
zontal axis, using a more compact scale on the right. Looking at the curve we
see that, when adding each of the first fifty voters with pc = 0.6, the initial ascent
in group competence is steep. At n = 51 we begin adding voters who are less
individually competent, with pc = 0.52. When adding the first such voters, group
competence initially declines sharply. But as we add more and more of those
less individually competent voters, the epistemic performance of the expanded
23 As Condorcet (1789/1994, p. 169) himself points out, ‘It does not always have to be the wor-
thiest men who are elected. . . . We need simply ensure that the plurality of the votes will always be
obtained by men who have a perhaps mediocre, but sufficient, amount of the qualities necessary
to fulfill the functions entrusted to them.’
group eventually recovers. At around n = 1,000 voters, the competence of the

larger group with all those less individually competent voters included exceeds
the competence of the smaller group with just the first fifty individually more
competent voters. And when we add yet more of the individually less compe-
tent voters beyond that, the competence of the still-larger group actually
exceeds that of the group consisting purely of the first fifty more individually
competent voters.
15.4 COMPETENCE-WEIGHTED VOTING RULES
Including less competent people in the electorate is one thing; including them
on a par with all other members of the electorate is another. The best formula-
tion of the ‘Guardianship Argument’ includes an ‘Epistemarchy Principle’ of
the following sort: ‘Political wisdom entitles the politically wise person to a
share of political power directly proportionate to his wisdom.’24
Voting rules that weight each person’s vote according to that person’s compe-
tence have been formally explored. It has been proven as a corollary to Bayes’s
theorem that, in a group of individuals with heterogeneous competence, the
group decision rule that maximizes the probability that the majority decision
among the group will be correct is one that assigns weights to each voter in
line with the competence of that voter. The optimal weight is proportional
to log( pc / (1− pc )) .25 This of course is just a formal refinement of the practice
commended by Mill of giving graduates of Oxford and Cambridge extra votes
in Parliament.26
Politically, of course, we might be reluctant to discount some people’s votes
for the same reason we think it is wrong to deny them a vote altogether. Even
though they are marginally less competent epistemically, we might think it is
24 Talisse 2005, p. 79, emphasis added.

25 Shapley and Grofman 1984; Grofman et al. 1983, pp. 274–5. The Bradley and Thompson
(2012) setup uses self-assessed competence, where those self-assessments are perfectly reliable;
but the range of possible weights is limited, owing to the fact that voters in their setup must cast
whole votes and voters have only ten votes each to cast across all the matters that are up for
decision.
26 Mill 1861/1977, ch. 8. As he writes (p. 473):
When two persons who have a joint interest in any business, differ in opinion, does justice
require that both opinions should be held of exactly equal value? If with equal virtue, one is
superior to the other in knowledge and intelligence—or if with equal intelligence, one excels
the other in virtue—the opinion, the judgment, of the higher moral or intellectual being, is
worth more than that of the inferior: and if the institutions of the country virtually assert that
they are of the same value, they assert a thing which is not. One of the two, as the wiser or
better man, has a claim to superior weight.
important they should be not only included but included as full equals in the
electorate. The same symbolic issues might arise in both cases.27
The question is just how much of an epistemic price we might have to pay to
obtain those symbolic benefits. Of course the Shapley–Grofman mathematics
are beyond reproach: weighted voting of the sort they prescribe is indeed the
way literally to maximize the probability that the majority vote will be correct.
For a small electorate of voters of heterogeneous individual competence, that
can make a difference. But among a large electorate of heterogeneous voters
with mean individual competence better than random by any appreciable
amount, there simply is not much to be gained by competence-weighting votes.
That can be seen clearly in Figure 15.3, which provides a numerical example
based on the Figure 15.2 scenario. In Figure 15.3, the line representing the
‘unweighted’ case is the same as in Figure 15.2. As we have seen there, without
weighting, group competence drops when adding the first hundred or so less
competent voters (i.e. voters fifty-one through 150). Now compare that result to
what happens with competence-weighted voting, as represented by the
‘weighted’ line in Figure 15.3. As we can there see, weighting avoids that drop in
group competence as we add new and less competent voters. The competence-
weighted voting rule achieves that result by according the less-competent
voters much less electoral importance—a vote of a more-competent voter has a
weight more than five times greater than that of the less-competent voter, in
this example. (So the epistemic success of the competence-weighted voting
rule comes at a cost of substantially unequal treatment.)
Yet as we see in Figure 15.3, the epistemic difference between the competence-
weighted and unweighted election rule reduces as the group size increases.
Indeed, it virtually disappears as the group grows large (to 5,000 in Figure 15.3,
for example); and as group size approaches infinity, group competence con-
verges to 1 in both the competence-weighted and unweighted cases. The upshot
is that competence-weighted voting rules make a real epistemic difference only
in relatively small groups.28
For any large group of even somewhat competent voters, then, the epistemic
gains from competence-weighting their votes are virtually zero. Among a million
voters each of whom is only just pc = 0.505 likely to be right individually, the
probability that the majority among them will be correct is already practically
indistinguishable from Pn ≈ 1.0 even operating on the basis of unweighted
majority rule. If there are some more competent voters in the population, then
27 That is particularly the case, since the Shapley–Grofman weight that should be given to each
voter is approximately ri - 0.5, where ri is the ‘percentage of times [that voter i is] in agreement with
the majority choices’ (Grofman et al. 1983, p. 275). There is of course a ‘competence’ backstory
driving that result. Nonetheless, appearances matter politically—and applying that rule cannot
help giving the impression of creating a ‘discrete and insular minority’ among the voting public
(Stone 1938, pp. 152–3 n. 4; Ackerman 1985).
28 As Bradley and Thompson (2012, p. 63) have also noted.

1.00 1.00
weighted
0.95 0.95
unweighted
Pn
0.90 0.90
0.85 0.85
pc = 0.52
pc = 0.6
0.80 0.80
0 100 200 300 400 500 1,000 5,000 10,000
n
Figure 15.3 Group competence as a function of n, given that the first fifty voters have
pc = 0.6, while all others have pc = 0.52., shown with unweighted (equal) votes and with
Grofman and Shapley’s weighted voting rule.
implementing instead a scheme of weighted majority rule, in Shapley–Grofman

fashion, would undoubtedly increase that in an infinitesimal way—but that
hardly matters among electorates of the size that politically most concern us.29
15.5 EPISTEMIC CONSIDERATIONS BEYOND

COMPETENCE
15.5.1 Other Ways Smaller Groups Might

Outperform Larger Ones
So far we have been focusing on just one way in which smaller groups might be
able to outperform larger ones, epistemically—viz., because individual compe-
tence of members of the smaller group is greater than that of members of the
larger group. That is the way the case for epistocracy has classically been cast.
Yet from our analyses in previous chapters, we know that there might be vari-
ous other ways in which the smaller group might perform better epistemically
than the larger group.
29 Even in an electorate of 10,000, Pn would be 0.8414 on the basis of unweighted majority rule.
That leaves more room for improvement by competence-weighting of votes. But the probability is
so high already that further improving it would not be worth any great costs in terms of symbolic
offence, etc.
Here are some of them:

• It might be possible to find smaller groups for whom the decision situation
is more truth-conducive and hence who are better at avoiding being mis-
led by the evidence, as per Sections 5.4 and 5.5.
• It might be the case that within small groups people’s votes are more inde-
pendent of one another, in any of the various ways described in Chapter 5.
• Smaller groups, being more conversable, might benefit from Section 9.3’s
Deliberation Effect.
• By devolving decisions to a set of smaller groups, we might benefit from
Chapter 8’s epistemic division of labour.
• It is possible (although it seems unlikely) that within small groups
people’s votes may be more diverse30 or even negatively correlated with
one another in ways that would produce the epistemic benefits described
in Section 7.2.
Of course, within the CJT classical framework, as long as (i) mean individual
competence in the larger group is pc > 0.5 by any appreciable margin, (ii) the
group is indeed large, and (iii) the other conditions of the CJT are met, the
probability that the majority of that large group will reach the correct conclu-
sion is already very near Pn ≈ 1.0 . In such cases, there is simply not much room
for improvement on the larger group’s epistemic performance. The epistemic
gain from relying on the smaller group rather than the larger one would be so
vanishingly small as to be of no practical consequence.
Yet we should acknowledge that there can be other cases in which smaller
groups can epistemically perform better than large ones by a wide enough mar-
gin to matter. In non-classical cases such as discussed in in Chapter 5, common
causes undermine the Independence Assumption—and perhaps in different
ways for small and large groups.
For instance, it might be that, owing to the influence of common causes or
other attributes of the decision situation, the probability of the majority among
a large group reaching the correct decision is upper-bounded at something less
than the upper bound of the probability that the majority among a smaller
group might do so. If the best responder facing the decision situation of the
larger group can only get the decision right in 51 per cent of cases, while the
best responder facing the decision situation of the smaller group is right in
75 per cent of cases, then it might be epistemically better to let the decision
be made by the smaller group.
Notice, however, that it is a failure of the Independence Assumption, much
more than of the Competence Assumption, that drives that result. And for
30 Contrary to the speculation in Section 15.2.4 (which we regard as more plausible) that larger
groups are likely to be more diverse.
that result to obtain, it must be the case that the Independence Assumption is

adequately satisfied more frequently in the smaller group than in the larger
one—which empirically may or may not be very likely.
When Condorcet refers to the ‘many prejudices’ that are more likely to influ-
ence large electorates than a select group of the well-educated, he was appar-
ently thinking about such effects in terms of his Competence Assumption. But
that has the awkward effect that the majority of a large group of voters would
nearly always vote for the wrong answer, which is not a very plausible conclu-
sion to reach. The Best Responder Corollary offers a more plausible and more
nuanced way to model Condorcet’s ‘many prejudices’. It explains how preju-
dices lower the convergence threshold, but it does not force us to deny that the
epistemic performance of an electorate subject to prejudices (or other biases)
can still benefit from the aggregation of votes. It suffices to assume that the
electorate, because it is more prone to be led astray by prejudices and biases,
faces misleading decision situations with higher frequency.
15.5.2 Other Ways Larger Groups Might Outperform

Smaller Ones
Translated into those terms, Condorcet’s ‘many prejudices’ claim amounts to a

claim that the best responder of the electorate gets things wrong more often
than the best responder of the small groups of experts. That is what proponents
of epistocracy should most plausibly try to claim. But is that so?
The problems with the epistocratic proposition are both theoretical and
empirical. First the theoretical problem. Proponents of epistocracy simply
assume that the ‘experts’ are less often misled by common causes. Justifying this
assumption is not easy, as there is a very obvious reason why the experts are
influenced by a set of common causes that is specific to experts. On the face of
it, they are more likely to be influenced by common causes because they were
likely selected as experts by the same procedure.
It is not just the (empirically) contested phenomenon of ‘groupthink’ that
could hamper the epistemic performance of the experts. That would be the case
if the experts undermined their independence by influencing each other dir-
ectly. The probably more important problem is that the experts are too correl-
ated in their judgements because they are all influenced by the same factors.
Among these factors could be a similar education, class, social background,
gender, a commitment to similar theories and ideologies, a similar network of
acquaintances, similar sources of evidence, and so on. And if the experts are
influenced by any such common causes, then this means they are likely to all
vote the same wrong way when they face a misleading decision situation.
It is true, of course, that the electorate as a whole is also influenced by many
common causes. Therefore, it is a difficult empirical question how much
c orrelation the common causes bring about, and how often different groups
face misleading decision situations. However, the debate between democrats
and epistocrats is now much more in balance.
The classical literature only ever stressed the competence parameter, insist-
ing that experts are more competent, if not by definition, then at least by over-
whelming empirical plausibility. The new jury theorems (represented here by
the Best Responder Corollary) show that the real challenge is not only compe-
tence but also independence. Because this is so, it is very much an open ques-
tion whether a small group of experts or the whole electorate is more likely to
perform better, epistemically.
15.5.3 The Political Upshot
What is the political upshot of all this? Basically, so long as the newly added
members are not too much less competent than the original members (or as
long as the genuinely incompetent among the group of new members are not
too large a proportion of the batch being added), adding new members to the
electorate will ordinarily improve the group’s epistemic performance. That is
the classical CJT-based argument for an extensive franchise—and for democ-
racy over epistocracy, more generally.
Where the group is very large already, however, adding extra voters (even if
they are reasonably competent) might not be worth all that much, epistemically.
Take the case of the Second Reform Act of 1867, for example. That increased the
British electorate from roughly 1 million voters to roughly 2 million. Now, the
majority among 2 million voters is more likely to be correct than that among
1 million, assuming the conditions of the CJT remain satisfied in each case. But
if you already have 1 million voters, and the conditions of the CJT are satisfied,
the probability of the majority being correct is already very close to 1.0. There
simply is not a lot of room for improvement by adding another million.31
Given an already large electorate, therefore, no great case for expansions of
the franchise can be made in epistemic terms alone within the classical CJT.
Instead, we do that for reasons of symbolism, or legitimacy. We do that because
we think it morally important that people have a say in the laws governing
them, or because we think it socially important for people to feel welcomed and
included in the national community.32 Those are the reasons for enfranchising
previously excluded subjects. Maximizing the epistemic power of the demo-
cratic majority simply does not play any important positive role in the story.33
In a non-classical framework, however, we can also identify potential
epistemic reasons for extending the franchise, in line with our discussion in
Section 15.2.4. For instance, if the extension of franchise improves the diversity
31 Lagerspetz 2010, p. 41. 32 Shklar 1991. 33 Hill 2016.

of the population and thereby reduces the influences of misleading biases or

other problematic common causes, then real epistemic benefits might accrue.
Take the Second Reform Act again: it dramatically extended the franchise to
the working classes and possibly reduced epistemically harmful bias in terms of
class interests, removing some blinkers that prevented good decision-making.34
The larger lesson we take away from the CJT is simply that epistemic con-
siderations certainly should not typically play any important negative role in
those discussions. We can afford to give votes to substantial numbers of mar-
ginally less competent (but still on average minimally competent) citizens—by
dropping the voting age to sixteen, for example—without seriously jeopardiz-
ing the probability that the majority among the new, enlarged electorate will
be epistemically correct in its decisions. It may well improve it. Especially if
we increase diversity by extending the franchise, that may well have a positive
effect on the group’s collective epistemic performance. Maybe dropping the
voting age to sixteen would not contribute much in that direction (how
different can sixteen-year-olds be from eighteen-year-olds?). But extending
the franchise to women or to foreigners or to other marginalized groups might
make a difference to diversity. It might reduce the overall influence of bad
common causes (prejudices) and therefore might raise the upper threshold of
convergence.
15.6 THE DIFFERENTIAL BENEFITS OF LEARNING

FROM EXPERIENCE
We have saved for last what is perhaps the most classical, and most reassuring,
argument for enfranchising less competent (yea, even positively incompetent)
members of the community. That points to the benefits of learning from
experience—and, we will here add, the differential impact that such learning
from experience will have on the probability that mass electorates compared to
small groups of experts will reach correct decisions.35
15.6.1 The Classical Argument of Participatory Democrats
There are various ways by which people acquire competence. One is through
book learning or, more generally, through systematic programmes of instruc-
tion and study. Another is through learning from experience.36 Experts excel in
the former sort of knowledge, but experts and laypersons alike benefit from the
34 Or it might not: whether the extension of the franchise reduces or increases harmful
common-cause-induced dependence is an empirical question.
35 This line of analysis arises from a suggestion by Ana Tanasoca. 36 Arrow 1962.
latter. Experts may ‘know’ what to expect from various economic strategies,
based on book learning; but laymen and experts alike come to know all the
better what it is like for their economy to be subjected to IMF ‘shock therapy’,
once they have been through the experience.
Democratic theorists have long supposed that one of the more important
experiences from which citizens can learn is that of voting and feeling the
effects of elections in which one has played a part. That ‘the major effect of
[political] participation is an educative one’ is a familiar lesson.37 John Stuart
Mill waxes lyrical on the ‘the exercise of political franchises by manual labour-
ers’ as ‘a potent instrument of mental improvement. . . . [. . .] Among the fore-
most benefits of free government’, Mill writes, ‘is that education of the
intelligence and of the sentiments, which is carried down to the very lowest
ranks of the people when they are called to take part in acts which directly
affect the great interests of their country.’38
15.6.2 Improving Already Competent Voters
Here we will offer a simple, stylized model to illustrate the power of learning
from experience. Suppose that each person’s individual competence increases
by 1 per cent with each round of experience—a tiny amount.39 Suppose, thanks
to their book learning, experts have uniform initial individual competence of
pcEXPERT = 0.60 on some subject. Suppose that laypersons are barely competent
on that subject at all—their uniform initial individual competence on that
37 Pateman (1970, p. 27) sources it to Rousseau. See similarly Thompson (1970, pp. 19–22) and
Macpherson (1973, esp. chs 1, 3). For evidence in support of this proposition from Switzerland,
where direct democracy is rampant, see Benz and Stutzer (2004); see similarly evidence from
American states with direct democracy in Smith and Tolbert (2004) and Donovan et al. (2009).
38 Mill 1861/1977 , ch. 8, pp. 467–8. Mill (p. 469) continues:
It is by political discussion that the manual labourer, whose employment is a routine, and
whose way of life brings him in contact with no variety of impressions, circumstances, or
ideas, is taught that remote causes, and events which take place far off, have a most sensible
effect even on his personal interests; and it is from political discussion, and collective polit-
ical action, that one whose daily occupations concentrate his interests in a small circle round
himself, learns to feel for and with his fellow-citizens, and becomes consciously a member of
a great community. But political discussions fly over the heads of those who have no votes,
and are not endeavouring to acquire them. Their position, in comparison with the electors,
is that of the audience in a court of justice, compared with the twelve men in the jury-box. It
is not their suffrages that are asked, it is not their opinion that is sought to be influenced; the
appeals are made, the arguments addressed, to others than them; nothing depends on the
decision [they] may arrive at, and there is no necessity and very little inducement to them to
come to any.
39 So any given individual’s competence in the next round is 1.01 times that individual’s com-
petence in the previous round.

1.0
0.9
initial pcEXPERT = 0.6

0.8
Pn
0.7
initial pcLAY = 0.501

0.6
0.5
0 1 2 3 4 5 6 7 8 9
Learning rounds
Figure 15.4 Learning by experience among twenty voters with an initial pcEXPERT = 0.6
and 1,000 voters with an initial pcLAY = 0.501, with the competence of each increasing by
1% in each round.
subject is pcLAY = 0.501. And suppose that experts and laypersons alike are
exposed to exactly the same subsequent experiences.
Now suppose that there are twenty experts and 1,000 laypersons in the com-
munity in question. In the initial setting, before any further experience, the
probability of the majority of experts being correct would be PnEXPERT = 0.814,
and the probability of the majority of laypersons being correct would be
PnLAY = 0.525 (assuming the CJT conditions are met). But now subject both
experts and laypersons to experiences, from which they learn in such a way as to
increase their individual competence by 1 per cent with each round of experi-
ence.40 Figure 15.4 shows what happens to the probabilities of those two groups
being collectively correct with increasing numbers of rounds of experience-based
learning.
As we see from Figure 15.4, the more rounds of experience-based learning
there are, the more the competence of laypersons as a group catches up with
the competence of experts as a group. Indeed, in the example described in
Figure 15.4, the collective competence of laypersons actually surpasses that of
experts after just three rounds of experience.
Two things are driving those results. One is that, in the CJT framework,
increases in individual competence make a greater contribution to collective
40 Nothing turns on the assumption that experts and laypersons learn at the same (1%) rate
from experience. We could let experts learn more from experience than laypersons, and qualita-
tively the same results would obtain.
competence when those increases are in individual competence levels that

were nearer random to begin with. Thus, a 1 per cent increase in competence
makes more of a difference to the collective competence of laypersons than
experts. The second thing driving the Figure 15.4 results is of course simply the
greater size of the group of laypersons than of experts.
15.6.3 Rendering Initially Incompetent Voters Competent
Notice, importantly, that this ‘learning from experience’ model also has the
capacity literally to transform incompetents into competent voters. Applying
the CJT framework all by itself, having voters who are on average worse
than random leads to catastrophic results: group competence rapidly drops
toward zero the larger such a group that is voting. But in our learning model,
voters gain competence with each round, so people who started out with
below-random competence can after a few rounds achieve above-random
competence.
Let us adapt our earlier example to illustrate that. Suppose now that the
uniform competence of laypersons on the subject under discussion is pcLAY = 0.49,
and everything else remains as before. Figure 15.5 displays what would
then happen.
In the Figure 15.5 scenario, the collective epistemic performance of the
layperson group is catastrophic in the first couple of rounds, during which time
each layperson was more likely to be wrong than right in his vote. But after
three rounds of experience, each layperson has learned enough to be more
likely to be right than wrong in his vote. Once that crucial threshold has been
crossed, the CJT effect from aggregating the more numerous votes of that
group flips into its familiar mode, magnifying individual competence (rather
than, as in earlier rounds, magnifying individual incompetence). After just a
few further rounds, the larger group of individually less competent laypersons
has once again overtaken the smaller group of individually more competent
experts as the better indicator of the correct position on the issue at hand.
15.6.4 The Political Upshot of Learning from Experience
This learning from experience model itself contains various presumptions, of

course. One is that recognizably the ‘same’ issue arises from time to time, fre-
quently enough for any given voter to have the benefit from several rounds of
experience with it (but not so rapidly that she does not have an opportunity to
digest the lessons of the previous round). That is true of many—but certainly
not all—important political issues. Major wars and great depressions are, hap-
pily, relatively infrequent phenomena; and arguably each is peculiar unto itself,

1.0
0.9 initial pcEXPERT = 0.6
0.8
0.7
0.6
Pn
0.5
initial pcLAY = 0.49

0.4
0.3
0.2
0 1 2 3 4 5 6 7 8 9
Learning rounds
Figure 15.5 Learning by experience with twenty voters with an initial pcEXPERT = 0.6
and 1,000 voters with an initial pcLAY = 0.49, with the competence of each increasing by
1% in each round.
rather than exactly the same as before. Still, there are many regularly recurring
political issues (economic policy, for example) that are essentially the same
from one time to the next—or anyway sufficiently analogous to the previous
ones—for us to be able to learn from past experience.41
The point of the simple model developed in this section is that the benefits of
such learning from experience are much greater for laypersons than experts.
Both learn from the same experiences, of course. But the benefits of the same
learning among the larger community of laypersons, when they vote, is vastly
greater than among the smaller group of experts. This is one important way in
which the collective competence of less individually competent lay voters
might exceed that of experts as a group.
41 Sunstein 1993.
16
Direct versus Representative Democracy
The previous chapter showed that a concern with the epistemic competence of
our government will, under certain plausible circumstances, lead us to prefer
democracy over epistocracy. But there are various alternative forms of democracy.
The business of this chapter is to reflect upon which broad form of democracy is
to be preferred from an epistemic perspective.
The first choice to be made, in Section 16.1, is between direct and representative
democracy. On the face of it, the CJT logic might seem automatically to favour
direct democracy, purely on the grounds that decisions reached in that way can
take more people’s independent assessments into account. But representative
assemblies are not only more select, and hence potentially competent, decision-
making bodies; they are also smaller, and hence potentially more conversable.
Those two factors, taken together, may well make representative democracy
epistemically superior to direct democracy.
The second choice, taken up in Section 16.2, concerns what attitude repre-
sentatives should take toward their task. When initially discussing the choice
between direct and representative democracy in Section 16.1, we do so on the
assumption that all members of the assembly are trustee-style representatives
exercising their own independent judgement in deciding how to vote. But in
Section 16.2 we go on to discuss the case in which some or all members of the
assembly take a delegate-style stance toward their role, acting strictly on the
instruction of their constituents.1 As we there show, that can make representative
government even better, from an epistemic point of view.
In subsequent chapters we will go on to discuss how other institutional design
features might help or hinder a political system’s overall epistemic perform-
ance. But those are details that can only be properly addressed once these more
fundamental choices have been resolved.
1 Burke 1774. Pitkin 1967.
Direct versus Representative Democracy 245
16.1 HOW CAN A SMALLER GROUP OF

REPRESENTATIVES BE BET TER THAN
A LARGER GROUP OF VOTERS?
Which is to be preferred, direct or representative democracy? There are many

considerations to weigh in any overall assessment of their comparative merits,
which have been much discussed over the years.2 Participatory democrats
point to many virtues of direct democracy, and political scientists show it to be
more practical a model than often supposed.3 Yet the conventional wisdom is
that, owing to problems of scale, in any large society democracy must be repre-
sentative in form.4 Still, there are many forms that representative government
might take, each with its own advantages and disadvantages.
Those are all important matters to ponder. But here we will engage with only
a very particular subset of them. Given the subject of this book, we shall in this
chapter focus narrowly on comparing direct and representative democracy
purely on the basis of their likely epistemic performance.
At first brush, which of them is to be epistemically preferred might seem to
follow straightforwardly from the CJT. After all, the CJT teaches us that (so long
as its conditions are satisfied, and other things being equal) more voters are
always better epistemically than fewer. And with direct democracy of course we
have far more voters than we do representatives, under representative democracy.
Q.E.D., or so it may seem.
But as we shall proceed to show, that would be too quick a conclusion.
There may well be good reasons for supposing representative democracy to be
epistemically superior to direct democracy. Not only might voters in a repre-
sentative democracy elect representatives who are substantially more compe-
tent than they themselves are. Furthermore, the possibility of deliberation and
discussion in smaller assemblies might give representatives an epistemic advan-
tage over the mass electorate. Taken together, those effects might make represen-
tative democracy epistemically superior to direct democracy.
2 For more historically based accounts see Manin (1997) and Urbinati (2006). Working more
from first principles are Mayo (1960) and Pennock (1979).
3 Among the former, see: Pateman 1970; 2012; and Barber 1984. Among the latter see: Cronin
1989; Budge 1996; Lupia and Matsusaka 2004.
4 Mill (1861/1977, ch. 3, p. 412) writes:
[T]he only government which can fully satisfy all the exigencies of the social state, is one in
which the whole people participate; that any participation, even in the smallest public func-
tion, is useful; that the participation should everywhere be as great as the general degree of
improvement of the community will allow; and that nothing less can be ultimately desirable,
than the admission of all to a share in the sovereign power of the state. But since all cannot, in
a community exceeding a single small town, participate personally in any but some very
minor portions of the public business, it follows that the ideal type of a perfect government
must be representative.
16.1.1 Incompetent Masses Choosing Competent

Representatives
In the passage quoted at the beginning of the last chapter, Condorcet himself
conjectured that the masses would ordinarily be worse than random at choosing
the correct alternative. If so, they could not be entrusted to vote on policy
directly. The flip side of the CJT tells us that, if voters are incompetent, then
the more of them there are the more certain it is that the majority among
them will be incorrect. Condorcet proposes, instead, representative democracy.
His idea is that the masses would choose representatives who are more likely than
random to choose correctly on their behalf.
That assumes, however, that voters who are bad at choosing policies will be
good at choosing people. Condorcet is explicit on that point:
[I]n the majority of matters submitted to the decision of an assembly, the same voters
whose opinions have such a small probability of being true can be enlightened
enough . . . to choose, as the most enlightened, one of those whose opinions will have
a large enough probability of being true. Thus a numerous assembly who are not very
enlightened could be usefully employed only to choose the members of a less numer-
ous assembly to whom the decision on other matters would then be entrusted.5
Madison, too, is cheerfully confident on that score.6
That assumption is obviously absolutely crucial to Condorcet’s analysis on
this point. If Condorcet is right that the masses are on average individually
worse than random at choosing the correct alternative, then that fact would give
ironclad epistemic grounds for not entrusting choices to them directly.7 But if
the masses were equally bad at choosing the correct representative as they are at
choosing outcomes directly, then they are virtually certain to choose the wrong
representatives, who are likely to choose the wrong alternatives in turn.8
What grounds might we have for supposing that people who are systematically
worse than random in their judgement concerning the substantively correct
alternative will be systematically better than random at assessing who will
5 Condorcet 1785/1976, p. 61. Condorcet (1789/1994, p. 170) offers a similar argument for ‘two-
stage elections’. Hume (1754/1760) before him offered a parallel argument for multi-stage elections,
with groups at each successive stage being more competent than the group that chose them but
choosing as their representative in the next round someone more competent than themselves.
6 As he put it in a speech before the Virginia Ratifying Convention, it is merely necessary that
the people possess the ‘virtue and intelligence to select men of virtue and wisdom’ (quoted in
Lupia and McCubbins 1998, p. 1). And as Madison had written in the Federalist no. 57, ‘[I]n so
great a number, a fit representative would be most likely to be found’ (Madison 1788/2003, p. 280).
7 The majority vote among any large number of independently worse-than-random voters is
virtually certain to be wrong in the classic CJT framework. In terms of the Best Responder
Corollary in Section 5.3, it almost certainly fails to track the best responder.
8 The epistemic cost being mitigated, in that case, only by the fact that there are fewer represen-
tatives than voters in the general electorate. If pc < 0.5 the more voters there are, the more certain
the outcome is to be incorrect.
r ecognize the substantively correct alternative?9 Here are some reasons that
might be given for thinking that that is the case:
1. Perhaps each voter is not systematically better than random at making
complex decisions, but each voter is nonetheless better than random at
choosing representatives on the basis of their evaluations of their general
character and competence.10 Voters might thereby end up electing repre-
sentatives who are much better than electors themselves at making com-
plex decisions.
2. Direct democracy asks people to decide on many issues in which their per-
sonal stakes are very low.11 Under such conditions biases are likely to guide
their thinking and attempts to influence voters are more likely to succeed.
3. Similarly, if voters think their competence is low, they might be tempted
to ‘epistemically free-ride’ by just siding with the majority.12 Such a lack of
independence may lead to cascades and hence bad outcomes.13
4. It is possible that many people do not make up their own minds but rather
follow opinion leaders. In the extreme case, everyone slavishly follows
the same opinion leader (or more than half the voters slavishly follow the
same opinion leader, whose side therefore always wins), and n is effectively
reduced to 1.
All those arguments are plausible, but note their limitations. The first argument
works only if people are better judges of other people’s characters than is probably
9 The issue has exercised lawyers, for example, in contemplating whether lay juries can adequately
assess expert scientific testimony: how can they know which experts to trust, if they do not
themselves have the expertise required to decide on the substance of the matter under discussion?
Cf. Hurd (1991), Brewer (1998), Anderson (2006), and Lane (2014).
10 Brennan and Lomasky (1993, p. 211) conjecture that ‘people are better judges of other people
than they are of rival policies’. ‘Converse’s studies support their proposal, at least insofar as people’s
attitudes toward candidates are more stable than their attitudes toward policies’, concludes Gaus
(1995, p. 261), referring to Converse and Markus (1979).
11 Downs 1957. Owen and Grofman 1984. As Schumpeter (1942, p. 262) complains,
the typical citizen drops down to a lower level of mental performance as soon as he enters the
political field. He argues and analyzes in a way which he would readily recognize as infantile
within the sphere of his real interests. He becomes a primitive again. His thinking is associa-
tive and affective. . . .
12 List and Pettit 2004. Vermeule 2009a, pp. 46–7. Landemore 2013b, pp. 193–5. Bentham
(1788/2002, p. 122) made the same point more obliquely: ‘[Claim:] with the number of members
increases the chance of wisdom. So many members, so many sources of light. Response: the
reduction which that same cause operates in the strength of the motive to bring out this light . . .
offsets this advantage’ (translated by and quoted in Elster 2013, pp. 152–3). For evidence of such
epistemic free-riding see Latané et al. (1979) and Diehl and Stroebe (1987). If people have to pay
some costs to ascertain the opinion of others who are likely to be better informed than they are
themselves, they might be expected to free-ride in another way, viz. letting others pay those costs
while not doing so themselves. There is some evidence of this in McCubbins and Rodriguez
(2006, p. 31).
13 On cascades, see our discussion in Chapter 10.
the case, given recent work on the ‘fundamental attribution error’.14 The last
three arguments point to factors that will presumably also be at work in voters’
choices of representatives, contaminating those as well.15 Furthermore, those
last three factors will also work to contaminate the decisions of representatives
themselves—who will after all be accountable to their electorates come the next
election, and who are likely to defer to the wishes of their electorates, at least to
some extent, in view of that fact.16
Finally, notice that Condorcet’s whole approach here presupposes that the
electorate is better at choosing people than policies in a very precise way. Voters
are presumed to be individually (a) less competent than 0.5 at choosing policies
but (b) more competent than 0.5 at choosing people. It is important that both
assumptions be satisfied for this argument to do the work that it is supposed to
do in support of representative rather than direct democracy. If (a) is not the
case, then the epistemic case against direct democracy fails; if (b) is not the case,
then the case in favour of representative democracy fails.
But on the face of it that is not a particularly plausible combination of assump-
tions. Perhaps sometimes the electorate’s differential competence at choosing
people and policies might take that very specific knife-edge form. Of course that
is possible—but is it at all probable? It seems unlikely to be the case often enough
to be of any general interest in designing overall social institutions.
16.1.2 Competent Voters Choosing Even-More-Competent

Representatives: The Selection Effect
A more natural starting point seems to be to assume that voters in general are
more likely to be right than wrong about both questions, both the question of
substance and the question of staffing—although they might be even more
likely to be right about the latter than the former. From that starting point, we
can then proceed to ask ‘how much better do they have to be on the latter than
the former question, for substituting representatives’ judgments for their own
to make epistemic sense?’.
To frame our thinking about that question, let us turn to another of the
American Founders. Writing in the Federalist no. 68, Alexander Hamilton
commends the Philadelphia Convention’s plan for the indirect election of the
president in these terms:
It was desirable that the sense of the people should operate in the choice of the
person to whom so important a trust was to be confided. This end will be answered
14 Nisbett and Ross 1977.

15 Although with respect to claim 2 it might be argued that a representative structure raises the
stakes for voters because they vote for a whole bundle of policies over many years.
16 We discuss issues to do with trustee versus delegate forms of representation in Section 16.2.
by committing the right of making it . . . to men chosen by the people for the special
purpose. . . .
It was equally desirable that the immediate election should be made by men most
capable of analyzing the qualities adapted to the station, and acting under circum-
stances favourable to deliberation. . . . A small number of persons, selected by their
fellows from the general mass, will be most likely to possess the information and
discernment requisite to so complicated an investigation.17
Here, Hamilton offers what are in effect two distinct reasons for thinking that
the smaller group (the Electoral College) will be epistemically superior to the
larger (the general electorate). The first reason has to do with the supposed fact
that members of the smaller group will be more ‘capable’: they will have been
selected because they ‘possess the information and discernment’ required for
the task. We call that the ‘Selection Effect’.18 Hamilton’s second reason has to do
with the supposed fact that, as a smaller group, members of the Electoral College
would be ‘acting under circumstances favorable to deliberation’. Call that (per
our discussion in Section 9.3) the ‘Deliberation Effect’.
We will discuss the Deliberation Effect separately, in Section 16.1.3 to follow.
That is going to turn out to be crucial. But for now let us focus on the Selection
Effect alone.
Let us begin by acknowledging that—at least in certain political communities—
there might indeed be a positive Selection Effect at work such that elected
representatives are more competent than the population at large. The best evidence
we have comes from Sweden, where researchers have gained access to the
results of IQ-style tests administered to some 90 per cent of the male population
as they were conscripted for military service. Members of Parliament on average
scored almost (0.85 of) one standard deviation higher on the quasi-IQ test
than the rest of the population, while broadly resembling them in most other
respects.19 That is impressive. But is it enough for the much smaller number of
MPs to overcome the epistemic advantage enjoyed by the much larger number
of voters across the community as a whole?
In addressing that question, let us start with a worked example of a relatively
small-scale society—a factory, for example, where workers have to choose
between making decisions on industrial action either in mass meetings of all
1,001 workers or by electing eleven shop stewards to make those decisions on
their behalf. Suppose that the 1,001 workers, following Hamilton’s advice,
decide to leave the choice to a small group of eleven elected representatives who
are more capable than themselves. The crucial question is this. How competent
does each of those eleven persons have to be, in order for that smaller group to
outperform the larger group epistemically?
17 Hamilton, Federalist no. 68, 1788/2003, p. 331. See also Mill’s (1861/1977, ch. 9) discussion of
indirect election of representatives.
18 Following Vermeule 2009b; 2011b, ch. 4. 19 Dal Bó et al. 2017.
Well, supposing each worker is individually pc = 0.55 likely to choose the cor-
rect alternative in a binary choice setting (and supposing they vote independently
of one another), the probability that a majority among them is correct is 0.999.
For a majority vote among merely eleven people to be as reliable as that, those
eleven people would have to be individually over pc > 0.80 likely to be correct.
Maybe on rare occasions we can find eleven people who have such high compe-
tence. But it certainly seems unlikely that, as a matter of course, 1,001 people can
find eleven people that much more competent than they are themselves. And if
they cannot, then they are epistemically better off entrusting the decision to a
mass meeting of all 1,001 workers rather than to just eleven shop stewards.
So far as the Selection Effect is concerned, the Federalist’s strategy becomes
even more implausible as the number of voters grows very large, as is the case
in real-world electorates.20 (When, here and below, we talk about applying our
models to ‘real-world settings’, it is—unless otherwise stated—the numbers of
people typically involved, rather than any other aspect of the ‘real world’, to which
we will be referring.)
Consider, for example, a case akin to the 2008 US election.21 There, the presi-
dent was elected indirectly by an Electoral College consisting of 538 members,
instead of being elected directly by the 131,000,000 people who voted in that
year’s presidential election. For the sake of this example, suppose—contrary to
the current fact, but in line with the Founders’ hopes and expectations—that
members of the Electoral College exercise their own judgement rather than
necessarily voting the same way as the citizens of the state they represent. And
let us make the modest assumption that voters among the electorate as a whole
are individually pc = 0.51 competent. For a majority among the 538 members of
the Electoral College to be epistemically equally reliable to a majority among
131,000,000 such voters, members of the Electoral College would have to be
individually pc = 0.976 competent.22 That seems inconceivable.
In short, the Selection Effect in and of itself is hardly likely to provide any
epistemic warrant for the Federalist’s turn away from direct and toward repre-
sentative democracy. Among groups the size of contemporary electorates, any-
way, it seems unlikely that a small group of individuals can be chosen that is
sufficiently more competent for its collective competence to equal or exceed
that of the electorate as a whole.
20 Most legislatures of moderately large countries have in the order of 500–600 members
(Inter-Parliamentary Union, 2010). To avoid computational overload, we calculate Pn for large
values of n by normal approximation.
21 ‘Akin to’, in the sense of ‘of the same size’—but with an Electoral College that operates on
different rules, as we go on to describe.
22 Calculated according to the Grofman–Dummkopf–Witkopf theorem (Grofman et al 1983,
p. 265). See Section 2.4 for the formula and Goodin and Spiekermann (2012, Appendix 2) for
details of the procedure.
Notice, however, that the Selection Effect might well yield a representative
assembly that is epistemically almost as good as the electorate as a whole.
Take the case of the 2008 US election, once again. On the assumptions just
stated, the probability of a majority of the electorate as a whole being correct
is Pn = 0.99999 . . . (with many more 9s following that). For the 538 members
of the Electoral College literally to equal that, each of them must (as we have
just said) be individually pc = 0.976 competent. But suppose we would be
content for the Electoral College’s decision to be just Pn = 0.99990 likely to be
correct. In order to achieve that, each member of the Electoral College would
only need to be about pc = 0.579 competent. It is not remotely unrealistic to
suppose that the masses might be able to choose representatives at least that
much more competent than themselves. That fact will become important in
the demonstration that comes next.
16.1.3 Epistemic Benefits of Smaller Groups:

The Deliberation Effect
The second prong of the Federalist’s conjecture is that a smaller group might be
more likely to be correct by virtue of its deliberative superiority. While this
possibility is less discussed among commentators on The Federalist Papers or on
the Electoral College, it is much discussed by a host of contemporary deliberative
democrats.23
In Chapter 9 we discussed various mechanisms by which this discussion and
deliberation might improve the epistemic performance of a group. We cannot
say just how big those effects are likely to be. Deliberative democrats offer many
reasons to think that they may be large. But—and this is our central claim here—
they do not need to be very big, in order to make the smaller and more deliberative
set of representatives superior epistemically to the mass electorate.
To demonstrate that, let us here (as in Section 9.3) take a deliberately low
estimate of the epistemic value of deliberation. Suppose that the superior delib-
erative circumstances of smaller representative assemblies will make them at
least one percentage point more likely to reach the correct decision than the
electorate as a whole. We can then calculate how individually competent each
representative would have to be, in order for the representatives’ collective
competence to be within one percentage point of that of the electorate as a
whole.24 In that way, we can calculate how big the Selection Effect needs to be
in order for the Deliberation and Selection Effects combined to vindicate
23 Bohman and Rehg 1997. Dryzek 2000. Luskin et al. 2002. Goodin 2008.

24 Again, calculated according to the Grofman–Dummkopf–Witkopf theorem (Grofman et al.
1983, p. 265). Again, see Section 2.4 for the formula and Goodin and Spiekermann (2012, Appendix 2)
for details of the procedure.
Hamilton’s conjecture, even if the Deliberation Effect is no stronger than this

very conservative estimate of what it might be.
Let us now examine in this way the collective epistemic competence of three
actual assemblies:
• First is the case most immediately on Hamilton’s mind: the Electoral
College for the first US presidential election in 1789. There were sixty-nine
members voting in the Electoral College that year.25 There were 38,818 votes
cast for electors that year.26
• Second is the Electoral College in 2008. That year 538 members of the
Electoral College were chosen on the basis of 131,237,603 popular votes.
• Third is the UK House of Commons in 2005. That year 27,110,727 popular
votes were cast for 646 MPs.
For each of those representative assemblies, we calculate two statistics. The
first, reported in the first column of Table 16.1, is how competent each individ-
ual in the group of representatives would have to be in order for their collective
competence literally to match that of the electorate as a whole, assuming each
voter is individually pc = 0.51 likely to be correct. That first column represents
the individual competence required of each representative for the Federalist’s
conjecture to be true by virtue of the Selection Effect alone (along the lines
discussed in Section 16.1.2). In the second column of Table 16.1, we report how
individually competent each of those representatives would individually have
to be in order for the assembly’s collective competence to come within one
percentage point of that of the electorate as a whole.27 That second column
Table 16.1 Estimated necessary individual competence of representatives to make

their collective decision epistemically equal to that of the electorate or no more than
1 percentage point worse (assuming voters are individually pc = 0.51 competent).
Equal to that of the Within 1 percentage point of
electorate as a whole that of the electorate as a whole
US Electoral College, 1789 0.714 0.637

US Electoral College, 2008 0.976 0.551
UK House of Commons, 2005 0.986 0.546
25 Each cast two votes, so there were 138 electoral votes cast but only sixty-nine actors exercising
independent judgement à la CJT.
26 That may seem a surprisingly small number of votes, but less than two-thirds of states that
year chose electors by any form of popular election, and even where they did the franchise was
severely limited.
27 Calculated, again, according to the Grofman–Dummkopf–Witkopf theorem (Grofman et al.
1983, p. 265) from Section 2.4. Goodin and Spiekermann (2012, Appendix 2) provide details of the
procedure underlying these calculations.
represents the individual competence required of each representative for the

Federalist’s conjecture to be true with the help of the Deliberation Effect.
The results differ for different assemblies, owing to the different numbers of
representatives and voters involved in each case.28 Still, the basic pattern across
all three cases is qualitatively similar. So for purposes of discussion, we will
focus once again on the case of the 2008 US presidential election, as depicted in
the middle row of Table 16.1.
Assuming the Deliberation Effect improves the collective competence of the
smaller Electoral College by just one percentage point, compared to that of
the mass electorate, then this is what we can see from Table 16.1. Voters who are
pc = 0.510 individually competent would only have to choose representatives who
are pc = 0.551 individually competent in order for the collective competence of the
538 members of the Electoral College to exceed the collective competence of
the electorate as a whole. That would seem to be an eminently realistic ambition.
Thus, the Federalist’s conjecture is plausibly correct. It might be very nearly
so by virtue of the Selection Effect alone. But it will be completely so only by
combining the Deliberation Effect with the Selection Effect. Still, if there is
indeed even a modest Deliberation Effect of the sort we here envisage, then
(contrary to what we might naturally assume from the CJT) representative
democracy might actually be epistemically at least as good as or even better
than direct democracy. A moderate Selection Effect suffices to get the smaller
group of representatives quite close to the epistemic performance of the popu-
lation, as the second column in Table 16.1 indicates. But to close the final gap,
the Deliberation Effect is ordinarily required. If the assemblies are large and the
Selection Effect quite strong then the contribution of the Deliberation Effect
needs to be only very small. However, less numerous assemblies or voters less
able to select the best representatives need to rely more strongly on a delibera-
tive boost (but, typically, it need not be large).
Indeed, if we make more optimistic assumptions about the Deliberation
Effect, then epistemic benefits of representation become even more pronounced.
This is particularly true if we leave the classic CJT framework and think in terms
of the Best Responder Corollary (from Section 5.3). If deliberation makes the
occurrence of systematically misleading situations less likely (by reducing
the influence of misleading common causes, as pointed out in Section 9.2.4),
then deliberation can systematically lift the convergence threshold and not
just match but outperform non-deliberative large groups. But to what extent
28 Column one in Table 16.1 increases from the top to bottom row because the increasing
number of voters bring the group competence of the electorate very close to 1 (from 0.99996 in the
first case to a number of the shape 0.9999999999 . . . with many more 9s to follow), so that the
representatives need to get very competent to literally match the population. In column two, by
contrast, the representatives only need to match a group competence of about 0.99, which, with
increasing group size, does not require very high individual competence of the representatives.
deliberation is really able to do that is an empirical issue that requires more

investigation of a different sort than we are conducting in this book.
16.2 DELEGATE VERSUS TRUSTEE REPRESENTATIVES
Depending on the conception that representatives have of their role, there

might or might not be all that much epistemic difference between direct and
representative democracy, however.
Up to this point, we have been assuming that representatives see themselves
as ‘trustees’, looking out for their constituents’ interests through the exercising
of their own judgement and without any regard to their constituents’ own
views on the matter.29 Then letting decisions be taken by a vote among the
representatives rather than by a direct vote of the people could make a huge
difference, not just from a democratic perspective but perhaps (as we have just
been discussing) from an epistemic perspective as well. Alternatively, how-
ever, representatives might conceive of themselves as ‘delegates’, acting only upon
the direct instruction of their constituents and faithfully following their con-
stituents’ views on the matter.30 Then substituting representatives for a direct
vote of the people may not make much difference epistemically (or perhaps
democratically either).31
Given the concerns of this book, our interest will be purely in the epistemic
aspects of the difference between trustee- and delegate-style representatives. We
begin by assessing just how much (or, as it turns out, how little) is epistemically
lost by entrusting decisions to an assembly of purely delegate-style representa-
tives as compared to making decisions by a vote of the electorate as a whole
(Section 16.2.1). We will go on, in Section 16.2.2, to discuss just how much scope
there is for the Deliberation Effect to make up the epistemic gap that would
otherwise exist between assemblies composed entirely of trustee-style repre-
sentatives and assemblies composed entirely of delegate-style ones. But there is
no reason to suppose that all representatives in the same legislature will neces-
sarily take the same attitude towards their role. Section 16.2.3 therefore assesses
29 Note that from the epistemic point of view of this chapter, what matters is that aspect of the
trustee/delegate distinction that Rehfeld (2009, p. 215) calls ‘Source of Judgment: whether the
representative lawmaker relies on his or her own judgment’ rather than the judgment of those
being represented in determining his or her vote.
30 Condorcet (1787/1994, Letter 1, p. 294) writes, when ‘limiting our examination to representative
constitutions alone, there can clearly be just two methods of decision-making: either by the plurality
of the representatives voting in accordance with their own reason, or by the plurality of represen-
tatives obliged to vote as their constituents dictate’. See further Mill (1861/1977, ch. 12).
31 Participatory democrats would of course insist that it would make a huge democratic dif-
ference (Pateman 1970; 2012).
the epistemic performance of ‘mixed chambers’ of representatives, some of

whom are acting as ‘trustees’ and others as ‘delegates’.
16.2.1 Delegate-style Representation: The Epistemic Costs

of Bunching Voters into Constituencies
First let us assess the epistemic consequences of making public decisions by a

legislative assembly composed purely of representatives with a delegate-style con-
ception of their role, rather than by a direct democratic vote of all the electorate.
Insofar as delegate-style representatives act purely on the instructions of
their electorates, and vote for something if and only if a majority of their con-
stituents would have done so, one might be tempted to assume that there would
not be much material difference in the outputs of both those systems.
But here is one difference. Direct democracy would take a vote of the entire
electorate. Representative democracy typically breaks the electorate down into
constituencies.32 Those constituencies are substantially smaller than the electorate
as a whole; and we know from the CJT that, in general, smaller groups are epi-
stemically disadvantaged compared to larger groups. So even if the representa-
tives elected by those constituencies take a purely delegate-style attitude
towards their role, and vote strictly in accordance with the preferences of the
majority of their constituents, there is bound to be some epistemic loss that
comes simply from bunching voters into constituencies.33 The question is ‘just
how much?’.
To calculate the probability that delegate-style representatives acting on
instruction in this way will reach the correct collective decision, we need to
apply the CJT formula given in Section 2.4 twice. First we use it to calculate the
probability that the majority of constituents is correct in the instructions that
they give to their representative.34 Since the representatives always vote as
instructed, we equate that with the probability that each representative’s own
vote will be correct. We then use the CJT formula for a second time to calculate
the probability that a majority of representatives with that level of individual
competence will reach the correct collective decision.
32 Sometimes the entire country can be a single constituency, with all voters electing all
representatives through a system of proportional representations. But even where proportional
representation is used, it is often done so within constituencies that, while large, are not coterminous
with the entire country.
33 Boland (1989, pp. 86–8) proves as much.
34 Typically, voters do not actually vote on the instructions they give, and delegate-representatives
have to rely upon other more imperfect mechanisms for surmising the views of their constituents;
and therein lies another source of epistemic slippage. But that is one we will not explore in our
modelling here. Operationally, for the purposes of our modelling, we simply equate the probability
of the majority choosing the correct delegate-representative with the individual competence of that
delegate-representative voting correctly in the representative assembly.
Calculating in this way allows us to see just how little epistemic cost arises
from breaking the electorate down into smaller constituencies, each of whom
is represented by a delegate-style representative. To show that, let us start with
a small-scale example: the factory discussed in Section 16.1.2.
Suppose the 1,001 workers in that factory elected their eleven shop stewards
by breaking themselves down into sections of ninety-one each. Suppose each
of the workers was individually pc = 0.55 likely to make the correct choice among
the alternative candidates. From the CJT formula, we know that the probability
that a majority of ninety-one of them in each section voting for the correct
alternative is Pn = 0.831. Since we are interested in the case of delegate-style
representatives here, we assume the elected shop steward votes for industrial
action if and only if a majority of the members of his section have voted for it as
their chosen alternative. Hence we assume that the individual competence of
that shop steward, when he votes on such matters, is the same as the collective
competence of his section. The probability of a majority of eleven shop stewards
each with individual competence of pc = 0.831 voting for the correct alternative
will be Pn = 0.995. Now compare that to the probability of all 1,001 workers with
pc = 0.55 coming to the correct conclusion through a majority vote in a mass
meeting. That is Pn = 0.999. That is higher than the probability that the shop
stewards representing their sections in delegate-style fashion would have been
correct in that matter—but only marginally so.
That conclusion gets powerfully reinforced as we shift from the small-scale
factory example to large-scale representative democracies. Think for example
of the 2005 British General Election. Each of the 646 Members of Parliament
was elected by around 40,000 voters. So long as the individual competence of
each voter is relevantly over 0.50, with numbers like 40,000 the probability that
the majority among them will be correct is expected to be very close to 1.35
Representatives who faithfully follow instructions given to them by these large
constituencies will reach correct decisions with near-certainty. Thus, there is
virtually no epistemic difference between delegate-style representative govern-
ment and direct democracy, where electorates are of the size we typically see in
real-world mass politics.
16.2.2 Trustee-style Representation: The Deliberation Effect Again
Is there any credible scenario by which the collective competence of a group of

trustee-style representatives might overtake that of groups of delegate-style
representatives?
35 Of course there is some pc so close to 0.5 that that will not be true. But, for instance, as long
as individual competence of 40,000 voters is above pc > 0.5108, collective competence is greater
than Pn > 0.99999.
Looking at it one way, that seems unlikely. After all, for the collective
c ompetence of an assembly composed purely of trustee-style representatives lit-
erally to equal that of the same number of otherwise identical representatives who
take a delegate-style conception of their role, each trustee would have to be
individually as competent as each delegate. But delegates act strictly on the
instruction of a large number of constituents; and if the individual competence of
each of those constituents is relevantly better than random, the instructions con-
tained in their majority vote for their delegate-representative are virtually certain
to be correct. Hence the individual competence of the delegate-representative
who acts strictly in accordance with those instructions is very close to 1. It is
hard to imagine that the individual competence of trustee-representatives could
ever match that.
But notice that trustee-style representatives enjoy one epistemic advantage
that delegate-style ones do not. Their decisions—and theirs alone—are capable
of benefiting from the Deliberation Effect described above (in Sections 9.3 and
16.1.3). However much delegate-style representatives discuss matters with one
another in the assembly ahead of a vote, at the end of the day delegates must by
definition vote as they have been instructed by their constituents.36 Trustee-
style delegates are not so bound, and their votes can indeed be changed by delib-
eration within the chamber. That affords trustee-style representative government
some real scope for ‘closing the epistemic gap’ between it and delegate-style
representative government in the same way as discussed in Section 16.1.3. That
might constitute a considerable advantage for trustee-style representation over
delegate-style representation. And if we are more optimistic about the effects
of deliberation to reduce or avoid misleading common causes, then this advantage
would be even greater.
16.2.3 Mixed Assemblies with Both Delegates and Trustees
Up to this point, we have been treating representative assemblies as if they were

comprised entirely either of trustee-style representatives or of delegate-style
representatives. A more realistic scenario is that some representatives will
behave as trustees and some as delegates.
Maybe some of the representatives will always act as delegates and others
will always act as trustees, on all occasions. More likely, different representatives
will assume each of those roles on different occasions, with each representative
36 With one possible exception: perhaps it would be possible for delegates to ask their constitu-
ents to instruct them again after the deliberation, and for the constituents to use the advantage of
better information or better options after the deliberation in issuing these new instructions. Then
the collective competence can improve even with a delegate-style approach. Of course, this depends
on constituents benefiting from deliberations to which they were not themselves directly party.
sometimes acting as a delegate and sometimes as a trustee. The model we shall

go on to develop applies equally well, whichever is the case.
In our modelling of a mixed representative assembly, with some trustee-style
representatives and some delegate-style representatives, we will take (for analytic
convenience) the small-scale case of an assembly with ninety-nine representa-
tives, each of whom is elected by 101 constituents. We assume the individual
competence of each representative, when acting as a trustee on the basis of his
or her own judgement, is pc = 0.55; and for purposes of this modelling exercise,
we will take no account of the Deliberation Effect. We assume the individual
competence of each constituent is also pc = 0.55, which (from the CJT formula
in Section 2.3) means that the collective competence of a majority among the
101 of them is Pn = 0.844. By definition, a representative who acts as a delegate
votes strictly in accordance with the majority of his or her constituents, so each
delegate-representative’s individual competence in this scenario is also, there-
fore, pc = 0.844.
Figure 16.1 displays the way in which the collective competence of the
assembly varies, depending upon the number of the representatives who
behave as delegates.37 Where there are no delegates and only trustees among
the ninety-nine members of the assembly, the collective competence of the
assembly is (taking no account of any Deliberation Effect, as we have said) 0.841.
But as we see from Figure 16.1, it rises rapidly with just a few delegate-style
1.0
0.9
99 representatives, mix of delegates and trustees.

0.8
Trustees with pc = 0.55; delegates instructed by
Pn
101 voters with pc = 0.55 each.

0.7
0.6
0.5
0 20 40 60 80 100
Delegates
Figure 16.1 The epistemic competence of an assembly mixing delegates and trustees,
with the number of delegates among ninety-nine representatives on the x-axis and
group competence on the y-axis.
37 Calculated as described in Goodin and Spiekermann (2012, Appendix 4).

representatives, closely approaching 1 by the time there are just twenty delegate-
style representatives in the ninety-nine-member assembly.
The upshot is thus that it does not take anything like half of the representatives
behaving as delegates for the collective competence of an assembly to come
very close to one, even setting the Deliberation Effect aside. At first brush this
might come as a surprise. But upon reflection, it should not be so surprising.
If the individual competence among representatives when acting on their own
is relatively low, most results will end up around an equal split of opinions
among trustee-style representatives. If there is a small group of delegate-style
representatives that votes more reliably, that increases substantially the assembly’s
collective competence overall.38
Given that the basic lesson of the CJT is that ‘more voters are always better’,
we might naturally suppose that direct democracy is to be epistemically preferred
to representative democracy, and delegate-style representatives to trustee-style
ones. But as this chapter has shown, that may well not be the case.
Even without taking any account of the Deliberation Effect, a moderately
numerous representative assembly is likely to be almost as epistemically powerful
as the mass electorate, certainly if it contains within it any appreciable number
of representatives who take a delegate-style view of their role. And the Deliberation
Effect may well further boost the epistemic power of groups of trustee-style
representatives within that assembly. Even in assemblies composed purely of
trustee-style representatives, the Selection and Deliberation Effects can easily
combine to make such decision-making bodies epistemically more reliable than
the electorate as a whole.
38 As List (2004) has shown, the probability of the group being correct, given a specific out-
come of the vote, depends on the absolute margin (and not the proportion). Even just a few highly
competent delegates make a certain margin between correct and incorrect votes very likely.
17
Institutional Hindrances
to Epistemic Success
The institutions of representative government can be structured in many different

ways, which can either help or hinder its epistemic performance. We will discuss
features of institutional design that can serve as an aid to collective competence
in the next chapter. This chapter will be concerned with features of institutional
design that hamper it.
All of them constitute, in one way or another, what Adrian Vermeule dubs
‘epistemic bottlenecks’. He elaborates that concept as follows:
The judgments of many minds may be the input to a decisionmaking process, but
if the structure of that process requires or allows few minds to accept or reject the
many-minded judgment . . . then the resulting decision may be little better than if
the one mind had simply decided for itself, right from the start. Gold in, garbage out.1
Assuming its conditions are satisfied, and all else being equal, the CJT tells us
that the majority in a larger group is more likely to be right than that in a smaller
group.2 Any element of institutional design that gives the smaller group a
decisive—rather than purely advisory3—role in the decision-making process can
thus be presumed to compromise the epistemic performance of that process.4
1 Vermeule 2009a, p. 50.

2 A case can also be made in terms of democratic legitimacy for laws to be made by larger
groups (legislatures) rather than smaller ones (cabinets, courts), of course (Waldron 2000). In this
book, however, we focus purely on epistemic issues.
3 ‘Statutes often require administrative agencies to consult with other agencies or officials or
with advisory committees, thus obtaining a second opinion before taking action. However, these
statutes do not usually give the party consulted a veto over the decision or require the decision-
making agency to follow the consulted party’s opinion’ (Vermeule 2011a, pp. 1147–8; see further
Vermeule 2009c).
4 Bentham (1823/1928, p. 154, quoted in Vermeule 2009b, p. 37) alludes to this argument when
writing:
As to ‘common reason,’ or the reason of the majority of the people who use their reason about
the matter, whose reason is it most to be apprehended should run counter to it? That of many
hundred [members of Parliament] chosen the greater part of them by the people . . . or that of
four [judges], appointed by the Crown.
Institutional Hindrances to Epistemic Success 261
We will be looking at alternative institutional structures of representative

government through the lens of the Condorcet Jury Theorem. Before doing so,
however, we must enter one blanket caveat. The CJT holds only insofar as its key
assumptions are satisfied5—and in cabinet, coalition, or legislative settings, some
of them may well not be. The CJT’s Independence and Sincerity Assumptions
seem to be particularly at risk in those settings, which are not only sites of opinion
leadership but also sites of much strategic manoeuvring.6 The analyses that
follow, therefore, should not be taken as representing what actually occurs in
those settings, necessarily. Instead, they are offered as limiting thought experi-
ments, showing what the epistemic consequences of alternative institutional
structures would be, were the CJT’s assumptions satisfied—which, in terms of
collective epistemic competence, is probably the best-case scenario.7
17.1 STRONG LEADERS
A strong leader can constitute a bottleneck right at the very beginning of the
political process. If there is a single strong leader who literally dictates what
views his followers hold, or are allowed to express, then there may be multiple
voices, but they will all be mouthing but one view. The ‘wisdom of crowds’
would then be reduced to ‘the wisdom of the chaperone’.8
The ‘strong leader’ in question might be an opinion leader of any of the sorts
discussed in Chapter 11. He might be a media mogul controlling all access to
politically relevant information. He might be a messianic national leader that
people follow blindly. Or the ‘strong leader’ might be a political party that people
vote for without fail, whatever the content of its policies.
Notice, however, that our specification of a ‘strong leader’ above is a very
demanding one. We have been talking about a ‘strong leader’, so strong that
the majority of people always follows the leader. And we have been speaking of
‘a strong leader’ in the singular, as if there were only one of them at work in any
given polity. If either of those conditions is relaxed to any large degree, then the
epistemic effects of leaders will be much less dire, as we shall show in Section
17.1.2. First, however, let us consider a case where a ‘strong leader’ of this special
sort might indeed exist.
5 At least tolerably well, in ways delimited by the extensions to the CJT discussed in Chapter 3.
6 In negotiations between the two houses of a bicameral legislature, for example, as discussed
in Tsebelis and Money (1995).
7 It is logically possible that an opinion leader might be pc = 1.0 competent or that strategic
voting might lead to better epistemic results. But presumably each of those eventualities is very
highly unlikely.
8 Vermeule 2009a, p. 53.
17.1.1 Party Leaders Dictating Party Policy
For that example, consider the case of a parliamentary party whose policy
positions are dictated purely by the party leader. It is an open question, perhaps,
whether that has ever literally been the case in any parliamentary democracy.
But it has surely been the ambition of a great many party leaders, which is in
itself sufficient to make this a case worth considering.9
To model a case of that, imagine a two-party democracy in which one party
is in the majority and each party leader is pOL = 0.55 likely to choose the correct
alternative. But suppose that all of any given party’s MPs vote strictly in accord-
ance with the dictates of their party’s leader. In that case, the probability that
the winning party will be correct is Pn = 0.55, corresponding just to the prob-
ability that that party’s leader will be correct (pOL= 0.55). And that remains the
same, no matter how many MPs took part in the voting.
Notice that things do not get appreciably better even if there are more than
two parties, if all those parties have strong leaders of this sort. Table 17.1
considers two such cases, involving three and five parties respectively. In the
two-party case, the probability that the majority of the legislature will vote
correctly is just the probability of the leader of the majority party, who dictates
that party’s policy, being correct himself. But in three-party case, a coalition
of two parties is required to form a majority. So the probability that the major-
ity outcome will be correct in that case is the probability of two or more of
the leaders being correct—and in the five-party case, of three or more of the
leaders being correct.
As we see from Table 17.1, the more different party leaders who have to
agree, the more likely the legislative majority is of being correct—but only by
a little. As we have observed above, the probability that the majority of the
legislature will be correct in the two-party case is Pn = 0.550, corresponding
to the probability that the strong leader of the winning party will be correct
(pOL = 0.550). The probability that the majority of the legislature will be cor-
rect rises only to Pn = 0.593 in the five-party case. The reason is simple. If each
party has a strong leader who utterly determines the votes of all of that party’s
MPs, then the move from a two-party to a five-party parliament brings only
three more distinctive, independent points of view to bear on the issues being
voted upon.
9 It is one of the examples of an epistemic bottleneck that Vermeule (2009a, pp. 51–2) offers:
Even if legislators of high average competence could pool their many minds in epistemically
impressive ways, the legislative leadership may form a kind of chokepoint that prevents
them from doing so; perhaps the wisdom of the legislative multitude must be approved by,
or at least refracted through, the mind of a Nancy Pelosi.
On the growth of the power of the Speaker of the US House of Representatives, a role which Pelosi
occupied, see Maass (1983, ch. 3).
Table 17.1 The collective competence of assemblies divided into

similarly sized parties with party leaders dictating party policy (where
the probability of the party leader being correct is pOL = 0.550).
Pn =
2 parties (1 required to form government) 0.550
3 parties (any 2 required to form government) 0.575
5 parties (any 3 required to form government) 0.593
Even this modest effect might be further reduced when the choice is among
more than two alternatives, and there are strong party leaders who are only a
little better than random at making that choice. Take for example the case of
five equal-sized parties choosing among three alternative policies. Supposing
each party leader is pOL = 0.34 likely to choose correctly (and votes with prob-
ability 0.33 for each of the incorrect alternatives), and all of the party’s MPs vote
strictly as their party’s leader dictates, then the probability that a majority of
MPs (i.e. a three-party coalition) will be correct is only 0.345.
The problem, in all those cases, is that there is a ‘strong leader’ who serves as
an epistemic bottleneck, preventing the wisdom of other MPs in his party from
helping to inform policy choices.
17.1.2 Mitigating Factors: Many Independent Leaders
The examples above illustrate just how deleterious it can be to have just one, or
even just a handful, of strong opinion leaders who are followed 100 per cent of
the time. But as we have said, those are very strong assumptions which are
not typical of the real world. More typically, there are several opinion leaders,
and they are followed often but not always. In cases like that, the deleterious
epistemic consequences are substantially mitigated.
We have already discussed the ways that can happen in Section 11.3.2. Here
let us simply adapt, in truncated form, some results presented there. In the
Table 17.2 scenario, there are 1,000 MPs, a (roughly) equal number of whom
follow each of a varying number of party leaders, and they do so only half of the
time, on average. Each party leader is assumed to be independent of each other
party leader; and each party leader is assumed to be pOL = 0.55 likely to be cor-
rect. When voting independently of the influence of his party leader, each MP
is also assumed to be individually pc* = 0.55 likely to be correct.
The results in the first three rows of Table 17.2 are much like those in Table 17.1,
showing that making MPs partially independent of their party leaders does not
make much difference where there are only a few parties (or hence party leaders).
But with partially independent MPs and larger numbers of independent leaders,
the collective competence of the assembly rises markedly—to fully Pn = 0.81 in
case of twenty party leaders whose MPs follow them only half the time.
Table 17.2 Probability of correct majority decision among 1,000

MPs with individual competence pc* = 0.55 split evenly among
multiple party leaders (pOL = 0.55), each voter following his
respective party leader with probability π = 0.5.
Party leaders Group partition Pn =
1 1,000 0.55
3 333, 333, 334 0.58
5 5 × 200 0.60
10 10 × 100 0.74
20 20 × 50 0.81
Having fully twenty equal-sized parties in any given legislature may be a

rarity. So even semi-strong leadership might be a persistent epistemic concern
in most actual parliamentary democracies. But if we shift now to think about
mass politics, it is not at all unrealistic to assume there might be twenty or more
independent opinion leaders at work there. If there are, and people do not
follow their lead slavishly, then the impact of those opinion leaders on the col-
lective competence of the electorate overall might be epistemically tolerable.
The epistemic bottleneck effect is there mitigated, first by having more ways
through the bottleneck (twenty opinion leaders), and second by having wider
bottlenecks (people follow their own opinion leader only half the time). And,
of course, if any given voter does not half-follow just one of those opinion lead-
ers but rather half-follows the pooled advice of all twenty of them, things are
epistemically even better, yet again.10
17.2 SMALL UPPER CHAMBERS OR COMMIT TEES
Another sort of epistemic bottleneck can be found in the institutional structure

of legislative assemblies, when smaller decision-making units are given a decisive
say in the outcomes. This might occur at the start of the legislative process, as
when a committee is assigned the task of pre-screening proposed legislation,
and only proposals that are passed out of the committee are put to a vote of the
legislature as a whole.11 Or it might occur late in the legislative process, as in a
bicameral legislature where a smaller upper house must also endorse the larger
10 Models like that are presented in Section 12.5 (just think of the ‘cues’ there as being ‘opinion
leaders’).
11 Vermeule (2009a, p. 52) includes them among his ‘epistemic bottlenecks’: ‘The numerous
“vetogates” in the federal legislative process ensure that the whole group of many minds often
cannot bring its group-level epistemic competence to bear as such. Rather decisions are made
serially by a series of separate groups of few minds (such as committees). . . . ’
lower house’s proposals in order for them to become law. Here we discuss each
of those cases in turn.
17.2.1 Legislative Committees as Epistemic Bottlenecks
Legislatures almost invariably divide themselves into smaller committees to do

much of the preliminary work on legislation. Furthermore, the larger assembly
is customarily quite deferential towards committee recommendations, when
voting on legislation.12 That has all sorts of consequences.13 One that has fea-
tured particularly prominently in recent political science discussions is the way
in which this process might produce a ‘structure-induced equilibrium’ in situ-
ations where there might otherwise be a voting cycle among legislators as a
whole.14 But it is the epistemic impact of deferring to committees that is of
concern to us for purposes of this book.
The thing to notice, in that connection, is simply that committees are
small subsets of the legislature as a whole.15 For example, the US House of
Representatives has 435 members, whereas its Ways and Means Committee
(which considers all budgetary legislation) has only thirty-nine members. Or,
again, the US Senate has one hundred members, whereas its Judiciary Committee
has only eighteen members.
The Condorcet Jury Theorem of course teaches us that, as long as its condi-
tions are met, then (all else being equal) smaller groups are epistemically less
reliable than larger ones. But just how big might be the epistemic costs of
the larger body (the House as a whole, or the Senate as a whole) deferring to
the judgement of the smaller body (the Ways and Means Committee, or the
Judiciary Committee)?
To arrive at a rough estimate of those costs, let us suppose that each member
of each chamber is individually pc = 0.55 competent. If all members of the larger
body voted up or down on a piece of legislation,16 the probability that the
majority among the larger bodies will reach the correct decision is given in the
first column in Table 17.3.17 The probability that the majority among the com-
mittees will reach the correct decision is given in the second column of that
table. The latter is the probability that the legislature will reach the correct
12 Matthews 1960. Fenno 1962; 1973. Price 1981. Maass 1983, p. 40. Shepsle and Weingast 1987.

13 Discussed, variously, in: Ferejohn 1974; Shepsle 1978; Weingast and Marshall 1988.
14 Shepsle 1979. Shepsle and Weingast 1981.
15 Except for the ‘Committee of the Whole’, into which it is sometimes convenient for the legis-
lature to constitute itself to evade certain constraints of parliamentary procedure.
16 Considering legislation in the Committee of the Whole was very much the practice in the
early US Congress (Polsby 1968).
17 The calculations reported in Table 17.3 assume that ties are broken randomly in the Senate
committee. By contrast, in the whole chamber of the Senate ties are broken by the Vice President,
whom we assume to be just as competent as the Senators.
Table 17.3 Probability of a majority vote for the correct outcome in

committees and whole legislatures, of varying sizes (pc = 0.55).
Probability of correct outcome in:
whole chamber committee
US Senate 0.844 0.663

– 100 members in whole chamber
– 18 members in committee
House of Representatives 0.982 0.736
– 435 members in whole chamber
– 39 members in committee
decision, supposing (as we do for the purposes of this exercise) that the assembly
as a whole automatically adopts the committee’s decision as its own.
As we can see from Table 17.3, there is a large epistemic price paid for legis-
latures deferring to committees in this way. In the cases reported in Table 17.3,
the assemblies deferring to the committees in question are more than twenty
per cent less collectively competent than they would be deciding the matter
directly for themselves. The reason, once again, is that those small committees
constitute epistemic bottlenecks.
17.2.2 Smaller Upper Houses as Epistemic Bottlenecks
Condorcet advocated a unicameral legislature.18 In this respect, he was follow-

ing his mentor Turgot, and Rousseau before him.19 But unicameralism was not
to the taste of the American Founders. ‘Collecting all authority into one centre’
spooked John Adams. He wrote in reply to Turgot:
A single assembly thus constituted, without any counterpoise, balance, or equilib-
rium, is to have all authority, legislative, executive, and judicial, concentrated
18 He wrote his ‘Letters from a Freeman of New Haven’ largely to lobby for it. At least that is
what he says at the end of Letter 1, in providing an overview (Condorcet 1787/1994, p. 294). As it
transpires, however, he does not actually come back to that issue until the last, short Letter 4
where, after the fashion of Rousseau (1762/1997, bk 2, ch. 2), he argues against multi-chamber
legislatures principally on the grounds that they promote ‘party spirit’ which ‘creates harmful
divisions’ (p. 330).
19 Rousseau 1762/1997, bk 2, ch. 2. Complaining of American state legislatures, Turgot (quoted
by Adams in ‘Preliminary Observations’, Letter I, in Adams 1787) had written:
Instead of collecting all authority into one centre, that of the nation, they have
established different bodies, a body of representatives, a council, and a governor,
because there is in England a house of commons, a house of lords, and a king. They
endeavor to balance these different powers, as if this equilibrium, which in England
may be a necessary check to the enormous influence of royalty, could be of any
use in republics founded upon the equality of all the citizens, and as if establishing
different orders of men was not a source of divisions and disputes.
in it. It is to make a constitution and laws by its own will, execute those laws at its
pleasure, and adjudge all controversies, that arise concerning the meaning and
application of them, at discretion. What is there to restrain them from making
tyrannical laws, in order to execute them in a tyrannical manner?20
The American Founders saw a balance of power as the key to avoiding tyranny.21
That is obviously a hugely important consideration. But for the purposes of this
book, we once again focus purely on the epistemic consequences of dividing
the legislature into two chambers, one larger and one smaller.
There are various forms of bicameralism. Here, however, we shall focus
exclusively on what we shall call ‘strong bicameralism’.22 By that, we mean
institutions involving two independent, directly elected sets of representatives
sitting in separate chambers of differing sizes, each of which has a veto on legis-
lation.23 The US Congress, for example, has a larger House of Representatives
and a smaller Senate, both directly elected and each with a veto on enactments.
From an epistemic point of view, there is of course always something to be
said for taking ‘a second opinion’.24 So epistemically it may seem to be a good
idea to have two (or more) independent bodies expressing their views on any
given question.25 But it would be wrong to jump to that conclusion, in the case
of strong bicameralism.
Recall the rules of strong bicameralism: Each chamber is not merely asked to
express its opinion—it is literally given a veto. So under the rules of strong
bicameralism, no law can be enacted without the independent approval of
both chambers.26 Assuming the votes of the two chambers are independent of
20 John Adams, ‘[Reply to] Dr. Franklin’; Letter XXV in Adams 1787.
21 Madison, Federalist no. 51, 1788/2003. When Condorcet’s publicist petitioned James
Madison for help translating and promulgating his Recherches, containing ‘four well-reasoned
letters sent to me by Condorcet, in which he mathematically upholds a unicameral legislature’,
Madison was firm in rebuffing him: ‘If your plan of a single Legislature etc as in [Pennsylvania]
were adopted, I sincerly believe that it would prove the most deadly blow ever given to republic-
anism.’ Both are quoted in McLean and Urken 1992, pp. 454–5.
22 A weaker form of bicameralism is found in the UK, for example, where since 1911 the upper
house has had only the power to delay but not veto legislation. That makes it more a mechanism
for providing a ‘political cooling-off period’ of the sort that will be discussed in Section 17.4.3.
23 Tsebelis 1995.
24 Vermeule 2011a. But only if the second opinion is genuinely independent of the first, which
in the case of two chambers of the same legislature it may well not be. As Mill (1861/1977, ch. 13)
observes, ‘If there are two Chambers . . . of similar . . . composition, both will obey the same influ-
ences, and whatever has a majority in one of the Houses will be likely to have it in the other.’
See further Waldron 2016, ch. 4.
25 Rogers 2001. There is also an argument, which we will not discuss here further (for reasons
given in Section 3.3.2) above, that multicameralism is the best way of ensuring that the correct
policies are enacted when there are no cycles but not when there are (Riker 1992).
26 Just how realistic is it to assume that the two chambers act completely independently of one
another? Insofar as both chambers are organized along party lines, and the parties are unified
across both chambers, the votes in the two chambers will not be independent of one another; that
would compromise the legislature’s epistemic performance in obvious ways. Another way in
which the independence of the two chambers’ votes might be compromised is by both being
one another, the probability of both approving is, of course, the product of the
probability of the first chamber approving multiplied by the probability of
the second chamber approving.27 Presumably there is some probability of a
majority in each chamber being wrong. And multiplying together two numbers
that are each less than one, of course, yields a number smaller than either of those
original numbers.28 Thus, it is typically true that bicameral legislatures will be
less likely to reach the correct decision than would be a legislature consisting of
just one of that legislature’s chambers alone.29
Making matters epistemically even worse is the fact that the two chambers of
a bicameral legislature are almost invariably of different sizes. That would not
matter epistemically if members of both chambers (or even just the smaller of
the two) predominantly regarded themselves as delegates rather than trustees
free to act on their own best judgement, as we have argued in Section 16.2.3.
It would not matter epistemically, if the smaller chamber simply automatically
endorsed the recommendations of the larger chamber. But insofar as there is a
large proportion of trustee-style representatives in a smaller chamber that
sometimes rejects legislation passed by the larger chamber, giving the smaller
chamber a veto can seriously compromise the epistemic competence of the
legislature overall.
To gauge just how much epistemic damage might be done by giving a smaller
chamber a veto over a larger one, let us consider two examples. The first is
the case of the first US Congress, elected in 1789. That year there were 101
Representatives and twenty-four Senators.30 For a second example, we take the
case of the US Congress elected in 2008. That contained 435 Representatives
and another hundred Senators.
Let us assume that all Representatives and Senators act as trustees, voting on
the basis of their own judgement alone. Let us assume that each member of
Congress has identical individual competence of pc = 0.55. And let us assume
directly elected by the same people using the same voting procedure for elections to both chambers
(Mill 1861/1977, ch. 13). But that is not the case with unelected upper chambers (like the British
House of Lords) or upper chambers elected on a different basis than the lower chamber (like the
US or Australian Senate).
27 Even if the votes are not independent, it remains true that fewer policy proposals will be
adopted when a veto player is added (unless we assume, unrealistically, that the two chambers
are always in perfect sync). That would be a good thing, epistemically, only if keeping the status
quo is the best choice. In Section 17.4.1 we consider settings in which doing nothing (i.e. defaulting
to the status quo) might be correct; in that case, having these vetoes is less bad and the calculation
would have to be done differently than if we count a non-decision as always constituting a deci-
sion failure (as our calculations elsewhere assume it to be).
28 Vermeule (2011a, pp. 1452–6, 1467–9) uncharacteristically misses this obvious point in his
discussion of the epistemic costs and benefits of bicameralism.
29 Unless, of course, there are some other effects that increase epistemic performance. Most
obviously, the existence of a second chamber might improve deliberative processes, as we will go
on to discuss in Section 17.2.3 (see also Chapter 9 and Sections 18.1.1 and 18.3.2).
30 Since in the Senate the Vice President can break ties with his vote, we model the 1789 Senate
as if it had twenty-five and the 2008 Senate as if it had 101 equally competent members.
Table 17.4 The probability of a correct decision from Congress (assuming each legislator
is pc = 0.55 competent).
House of Senate Congress as
Representatives a whole
US Congress 1789 (House 101 seats; Senate 0.844 0.694 0.585

24 + 1 seats)
US Congress 2008 (House 435 seats; Senate 0.982 0.844 0.828
100 + 1 seats)
all other conditions of the CJT are satisfied. The first column in Table 17.4
reports the probability, given those assumptions, that a majority in the House
will reach the correct decision; the second column reports the probability of a
majority in the Senate will do so; and the third column reports the probability
that Congress as a whole (i.e. both the House and the Senate) will reach the
correct decision.
Since each chamber has a veto, the collective competence of Congress as a
whole is equivalent to the product of the competence of both chambers.31 In
consequence, strong bicameral arrangements can carry a substantial epistemic
cost, as we see in Table 17.4. In the first US Congress, the House alone would on
this model have had a competence level of Pn = 0.884. Thanks to both chambers
having a veto, however, the competence of Congress as a whole drops to only
Pn = 0.585. In the 2008 US Congress, the House alone would on this model have
had a competence of Pn = 0.982; but since both chambers have a veto, the com-
petence of that Congress as a whole drops to Pn = 0.828. The drop is less severe
in the 2008 case than the 1789 case, however, because the 2008 Senate is numer-
ically much larger. What drives those results is the smaller upper house’s having
a veto makes it an epistemic bottleneck for the legislature as a whole.
17.2.3 Mitigating Factors: The Selection and Deliberation Effects
Our calculations so far have taken no account the Selection and Deliberation
Effects discussed in Section 16.1, however. The smaller upper house is often
referred to as ‘the deliberative chamber’.32 Being smaller, it is more conversable—
and it may be able to reach the correct decision in that way more often than
suggested by the above calculations, which are based just on aggregative pro-
cesses. That is what we have called the ‘Deliberation Effect’. Furthermore, the
smaller upper house, being more select and more prestigious,33 might be thought
to attract higher-quality members who are more likely to vote correctly than
their lower-house colleagues. That is what we have called the ‘Selection Effect’.
31 Assuming, of course, that the votes are independent conditional on the state of the world.
32 Uhr 1998. 33 Matthews 1960.
We can employ the same procedure as in Sections 16.1.2 and 16.1.3 to explore
whether such effects are likely to compensate for the other epistemic costs of
giving a veto to an upper house with fewer members. Let us assume, once again,
that the Deliberation Effect will make the smaller chamber one percentage
point more likely to vote correctly than it would otherwise have enjoyed, on the
basis of aggregating votes there alone. Then we calculate how much more indi-
vidually competent members of the smaller chamber would have to be, in order
for a majority of them to be as likely to be correct as is the majority among the
larger chamber.
Suppose for the purposes of this calculation that members of the larger
chamber, the House of Representatives, have individual competence of pc = 0.55.
In a legislature as small as the 1789 Congress, the twenty-four members of the
Senate would have to be only pc = 0.595 individually competent in order for the
majority among them to come within one percentage point of the probability
that the 101 members of the House would reach the correct decision. It would
take only a very modest Selection Effect to achieve that. Similarly, in the 2008
Congress, the one hundred Senators would have to be only pc = 0.594 individu-
ally competent in order for the majority among them to come within one
percentage point of the probability that the 435 members of the House would
reach the correct decision.
All of that is merely to say, however, that (thanks to the Deliberation and
Selection Effects) the smaller Senate may well be collectively equally competent
to the larger House. The point remains that, in a strongly bicameral legislature
in which each chamber has a veto over legislation, the probability that the
correct result will be chosen in both chambers is less than the probability that it
will be chosen in one of them alone.34 That is an epistemic cost of strong bicam-
eralism that would not disappear unless one of the chambers were literally
certain to vote for the correct position (or were completely deferential, auto-
matically ratifying whatever legislation was passed by the other chamber; or
were dominated by delegate-style representatives rather than trustees).
The same can, of course, be said with regard to legislatures delegating
powers to smaller committees. If the smaller group is substantially more con-
versable than the legislature as a whole, and gets the boost in its collective
competence that we associate with the Deliberative Effect, then the members
of that committee may not have to be all that much more competent than the
average legislator in the assembly as a whole in order for the collective compe-
tence of the committee to exceed that of the legislature as a whole. And mem-
bers of the specialist committee are indeed likely to be at least somewhat
34 We here treat reaching no decision as a decision failure. The effect of no decision is, however,
just sticking with the status quo; and if that is the correct choice, Congress as a whole would
perform better than either chamber separately. It would do so, however, not so much by reaching
the right decision as by failing to reach a decision.
individually more competent on matters that come before the committee, for
various reasons.35
But even if the committee were just as collectively competent as the larger
assembly, the point remains that for the correct outcome to be enacted into
legislation, it must be passed both by the committee and by the larger assembly.
The probability of that is the probability of it being passed by the committee,
multiplied by the probability of it being passed by the larger assembly.36 If the
larger chamber automatically passes anything that is recommended to it by
the committee, that latter probability would be 1.0. But suppose legislators in
the whole chamber do not automatically and completely defer to the commit-
tee (as we assumed they did for the purposes of our analyses in Section 17.2.1).
Suppose instead that legislators exercise some independent judgement. And
let us charitably suppose that the collective competence of the larger assembly
is identical to that of the committee—each, say, is Pn = 0.80 likely to reach the
correct conclusion. The probability that the correct conclusion will be chosen
in this two-stage process would still be only PnJOINT = 0.64.
Thus, while the Deliberation and Selection Effects might help make up for
the epistemic costs associated with the smaller size of one of the bodies in the
decision-making process (the Senate or the committee), there are still the
epistemic costs associated with the two-stage nature of the process to be
reckoned with.37 Smaller committees or upper chambers with an effective veto
on legislation constitute epistemic bottlenecks, undermining the epistemic
performance of the legislative process.
One (rather demanding) possibility remains as to how the Deliberation Effect
might compensate for the epistemic bottleneck of a smaller upper chamber.
Suppose that each chamber improves its collective competence because the
two-chamber setup induces them to investigate more, negotiate more, and think
more. Then it might be possible that the epistemic performance of both chambers
will be sufficiently enhanced to compensate for the epistemic bottleneck effect.
To continue with our previous example, suppose that each chamber has col-
lective competence PnUPPER = PnLOWER = 0.8 when deciding separately. The fact that
they each can veto the other reduces their joint competence to PnJOINT = 0.64.
But now suppose that the negotiation process between the chambers boosts
the collective competence of each chamber. In our case, to completely
remove the bottleneck, it would have to be the case that PnLOWER × PnUPPER ≥ 0.8.
For example, if both chambers are equally competent that would require
PnLOWER = PnUPPER = √0.8 ≈ 0.894.
35 Variously: because they were selected for the committee on that basis (Masters 1961);
or because they gained in competence as a result of serving on the committee; or because the
committee only has to consider a small subset of all the alternatives that would have to be
considered by the legislature as a whole.
36 Assuming, again, the independence of votes, which is perhaps less plausible when the com-
mittee is composed of members of the larger chamber.
37 As Vermeule (2009a, p. 52) notices, when referring to the ‘serial’ nature of the process.
Perhaps an inter-chamber Deliberation Effect of that magnitude might be

possible. Clearly, however, only a substantial Deliberation Effect could counter
the effect of the epistemic bottleneck of the smaller upper house.38 In fact, the
magnitude of the required effect suggests that we would have to see some fairly
convincing evidence of how inter-chamber deliberation improves decision-
making. We leave the judgement to the reader whether any such effects can be
observed in such political systems.
17.3 PART Y WHIPS AND SMALL, PIVOTAL

PARTIES IN COALITIONS
Political parties can aid the epistemic performance of representative govern-

ment, in ways that we will discuss in Section 18.1. Here, however, let us consider
some ways in which they might hinder it. First we will discuss ways in which
strict party-line voting in the legislature might hinder the epistemic performance
of the legislative process. Next we will discuss ways in which small, pivotal parties
in coalition governments might do likewise.
17.3.1 Party Whips in the Legislature
In the real world, representative assemblies are almost invariably organized

along party lines. In many assemblies, parties are weak things. But in many
they are strong. Where the whip is strong, representatives virtually always vote
in line with the instructions of their party, rather than exercising their own
independent judgement or following the instructions of their constituents.39
Here let us once again take the limiting case, of a ‘strong party whip’ such that
MPs of that party literally always follow the instructions of the party.
In a parliament with strong parties of that sort, there are only as many inde-
pendent sources of judgement at the point of voting on legislation as there are
parties. Of course there are more physical bodies in the chamber. But all those
MPs are not independent sources of judgement: all of them simply vote their
party’s line. Reducing the number of points of independent judgement in that
way can have dire epistemic consequences, the CJT teaches us.
But just how bad the consequences would be depends, of course, upon exactly
how those ‘party instructions’ are themselves generated. At one extreme—that
38 Furthermore, the Deliberation Effect would probably need to be at work in both chambers—
if it is at work in one chamber alone, it would have to be such a very strong effect as to be
implausible.
39 In terms of Section 16.2, we might therefore say they are ‘delegates’—but delegates of their
party rather than of their constituents.
considered in Section 17.1.1—party policy might be dictated by the party leader

and her alone. In that case, the probability that the majority of the governing
party’s MPs is correct is simply the probability that the party leader herself is
correct. And the probability that the parliamentary majority is correct does not
increase whatsoever the more MPs there are in the governing party.
As we have seen in Section 17.1.2, things are only a little better if the govern-
ment is a coalition composed of multiple parties, each with a leader who forms
his policy views completely independently of other party leaders. Obviously,
that is a highly unlikely scenario in a coalition government. But in any case,
having multiple independent leaders whose MPs slavishly follow their lead is
unlikely to make any great epistemic difference to the governing coalition’s
overall epistemic performance.
So much for the case in which the ‘party line’ is dictated by the party leader
alone. Next let us consider the case in which MPs are not purely lobby fodder,
doing as they are told without exercising any independent judgement anywhere
at all. Let us now suppose instead that they do exercise independent judgement—
only in the party room, rather than on the floor of the legislature. That is to say,
rather than exercising independent judgement when voting on legislation, they
instead exercise independent judgement within the party in trying to shape its
policies in ways they think best.40
To model this, we will assume that party policy is chosen purely by a major-
ity vote among the party’s MPs.41 For the purposes of our calculations, we
assume that each MP is individually pc = 0.55 likely to be correct when casting
that vote. When it comes to a vote on the floor of the legislature, we assume that
all MPs always vote strictly in accordance with their party’s line, as determined
by that party-room vote. Since this model (unlike the previous one) is sensitive
to the actual number of MPs, we need to stipulate a number. We will take a case
similar to the UK Parliament, and suppose there are 603 MPs, with 302 votes
being required to enact legislation.
We will begin by considering five different ways into which MPs might be
divided into parties:
1. No parties, with each MP voting purely on the basis of his or her own
independent judgement.42
2. Two almost-equal parties, one controlling 302 MPs and the other 301 MPs.
40 For a more elaborate model of this process see Caillaud and Tirole (2002).
41 Of course, in the real world there may be many extra-parliamentary influences on party
policy as well. And there are alternatives to party-room democracy altogether: party policy might
be decided upon by the entire party conference, for example. If all those other influences satisfy the
standard CJT assumptions concerning Competence and Independence, then the more such input
there is the better from an epistemic point of view. But here we confine ourselves to the case in
which party policy is chosen purely by a majority vote among the party’s MPs in the party room.
42 That is the original Federalist vision (Madison, Federalist no. 10, 1787/2003, p. 41), and was
the older British practice.
3. Three equal parties, each controlling 201 MPs (an approximation to some
European parliaments).
4. Three unequal parties, two controlling 275 MPs each, the other control-
ling fifty-one MPs (an approximation to a recent UK House of Commons).
5. Twenty almost-equal parties, seventeen with thirty MPs each, three with
thirty-one MPs (an approximation of a splintered political landscape).43
Obviously, in this sort of model, the collective competence of a party depends
purely on its number of MPs. The probability that the two-party parliament
will make the correct decision is equal to the probability that the party control-
ling 302 MPs will do so. The probability that the three-party parliament will
make the correct decision is equal to the probability that any two of the parties
or all three parties would do so.
Those probabilities are shown in Table 17.5, where we compare the probabil-
ity of a 603-member parliament making the correct decision in scenarios where
it is structured into well-whipped parties of various sorts.44 The first line of
Table 17.5 provides the baseline calculation of what the collective competence
of the 603-member parliament would be, if each representative voted inde-
pendently of each other, exercising his or her own judgement without refer-
ence to any party. The subsequent lines show what the collective competence
of the same parliament would be if there were strictly whipped parties of
varying sizes.
From Table 17.5 we see that, if there were no parties and each MP were a
completely free, independent epistemic agent, the probability that the majority of
those 603 MPs would vote correctly is Pn = 0.993. With two almost-equal-sized
parties, the probability that the majority across parliament as a whole would
vote correctly is (on the assumptions we are working with here) just equal to
the probability that a majority of the 302-member majority party is correct.
But that is still Pn = 0.954. If there were three equal parties, the probability that
a parliamentary majority comprised of a coalition of any two of them would
reach the correct decision would be Pn = 0.983. And having three parties of
unequal size, none of which is pivotal, is epistemically almost as good as having
three of the same size (Pn = 0.976). Finally, having a great many parties also does
not significantly reduce overall epistemic competence. With twenty parties of
(almost) equal size it goes down to Pn = 0.973. While each single party would—if
pivotal—constitute a tight bottleneck, the fact that there are so many of them
offsets that effect.
43 In all these examples we assume that all votes are independent conditional on the state,
which also implies that the positions of the parties are independent conditional on the state.
44 The calculation is straightforward: we simply sum up the probabilities of all possible winning
coalitions. Note that the winning coalitions possible in the case of twenty parties include those
where ten parties are correct and ten incorrect, as long as two or three of the thirty-one-member
parties are among the correct ones.
Table 17.5 The collective competence of assemblies divided into strictly whipped
parties with party policy chosen by majority vote of that party’s MPs (for a 603-member
legislature, each MP having individual competence of pc = 0.55).
Scenario 1: no parties 0.993
Scenario 2: Two almost-equal parties (302 MPs, 301 MPs) 0.954
Scenario 3: Three equal parties (201 MPs each) 0.983
Scenario 4: Three unequal parties (275 MPs, 275 MPs, 53 MPs), none of which is 0.976
pivotal
Scenario 5: Twenty almost-equal parties (17 × 30 MPs, 3 × 31 MPs), none of which 0.973
is pivotal
Thus, structuring parliament into well-whipped parties whose policy is

decided by a majority vote among that party’s MPs comes at only modest
epistemic cost. Across all the scenarios reported in Table 17.5, the collective
competence of parliament is very nearly the same as if the parliament had no
parties at all. Furthermore, the competence gap is small enough that it could
arguably be closed by higher deliberative quality among the smaller group in
the party room, as compared to the larger group on the floor.45
In all of the scenarios reported in Table 17.5, the epistemic bottleneck asso-
ciated with a strong party leader dictating party policy in Section 17.1.1 has been
substantially overcome by letting party policy be determined instead by a
majority vote among a reasonably large number of MPs of that party. That all
MPs of the same party then automatically vote in whatever way their party has
thus decided, rather than exercising their own independent judgement on the
floor of the legislature, constitutes something of an epistemic bottleneck in
itself, of course. But it is not much of one, as we can see comparing the first row
of Table 17.5 with the rest.
17.3.2 Coalition Government with Small, Pivotal Parties
However, matters are much worse epistemically if there is one relatively small
party that is ‘pivotal’ and must be included in any winning coalition.46 The
small, pivotal party that must be included in the coalition for the correct out-
come to occur then becomes an epistemic bottleneck, and the probability of the
correct outcome winning a majority suffers in consequence.
To get a sense of that, let us simply go back to Scenario 4 in Table 17.5 and
make the small (fifty-three-MP) party there pivotal. Table 17.6 reproduces
Scenario 4 from the previous table (for ease of comparison) and then adds
Scenario 6, which is identical to Scenario 4 except that in Scenario 6 the
45 Per the Deliberation Effect we have discussed in Sections 9.3, 16.1.3, and 17.2.3.
46 As a ‘pivotal middle party’ that must be included in any ideologically connected winning
coalition, for example (Axelrod 1970, ch. 8; Taylor and Laver 1973). Similar problems arise where
there is one justice who is always the ‘swing vote’ on the nine-member US Supreme Court
(Vermeule 2009a, p. 51).
Table 17.6 The collective competence of assemblies divided into strictly whipped
parties with party policy chosen by majority vote of that party’s MPs, with a pivotal
small party (for a 603-member legislature, each MP having individual competence of
pc = 0.55).
Scenario 4: Three unequal parties (275 MPs, 275 MPs, 53 MPs), none of which is 0.976
pivotal
Scenario 6: Three unequal parties (275 MPs, 275 MPs, 53 MPs), the small party is 0.767
pivotal
fty-three-MP party must vote for the correct outcome in order for the correct
fi
outcome to win a majority. Comparing those two rows of Table 17.6 shows the
effect of adding that requirement on the probability that the correct outcome
will secure a majority in the parliament.
Scenario 4 depicts a parliament composed of well-whipped parties with 275,
275, and fifty-three MPs respectively, with each MP being individually pc = 0.55
competent when voting in the party room and with each MP being certain to
vote on the floor the way the majority of his party room decided. On Scenario 4,
where the majority in parliament can be composed of any combination of
parties, the probability that the correct outcome would secure a parliamentary
majority is Pn = 0.976. Scenario 6 imposes the constraint that the winning
majority in the parliament must include the party with fifty-three MPs. In that
case, the collective competence of the parliament drops to Pn = 0.767.47
Again, it is clear what is driving these results. Introducing into the model a
small, pivotal party that has to be included in the winning coalition constitutes an
epistemic bottleneck. The calculations in Table 17.6 enable us to see just how bad
a bottleneck it might be, in real-world political settings. Having to include a
small, pivotal party of the sort described would reduce the epistemic competence
of parliament from over 97 per cent to merely 76 per cent—a drop of fully
21 percentage points. Thus, it is not having small, strongly whipped parties that is
epistemically problematic. It is small parties being pivotal in such a way that
they must be included in the winning majority that does far more damage.
17.4 PRESIDENTIAL VETOES

AND SUPERMAJORIT Y RULES
Political decision-making procedures also create epistemic bottlenecks, inso-

far as they empower a smaller group of people to prevent a decision by a
47 And if MPs were individually less competent than we are here assuming, Scenario 6 would be
much worse yet again. Suppose for example MPs were individually pc = 0.510 competent. Then the
collective competence of a parliament structured as in Scenario 6 would be Pn = 0.482. That is to
say, the parliament in that case performs epistemically worse than would any given MP acting alone!
larger group of people from coming into force. Supermajority rules within
the legislature do that in one way: if a two-thirds vote is required for some
enactment to become law, then any third of the legislature can block that from
happening. Presidential veto powers do that in another way: one individual
(the president) can, through his veto, prevent a piece of legislation supported
by hundreds of legislators from becoming law. And the two mechanisms are
typically linked, insofar as the president’s veto is overridable by a supermajority
in the legislature.
17.4.1 Supermajority Rules
On many matters, we are content to let decisions be taken by a bare majority of

voters. But on other matters we demand much more than that—a two-thirds or
three-quarters majority, or even a unanimous verdict. Politically, there might
be many reasons for demanding extra-large majorities like that. Requiring a
supermajority for one branch of government to overrule the decisions of
another branch underpins the constitutional separation of powers.48 Requiring
a supermajority to amend the constitution entrenches the basic law.49 More
generally, requiring a supermajority to change the status quo slows the rate of
change, which may be a good thing (insofar as change is invariably costly50) or
a bad thing (insofar as requiring a supermajority to alter the status quo simply
entrenches existing power and privilege51).
48 That was the US Constitution’s Framers’ rationale for requiring a supermajority in Congress
to overturn the President’s veto (Hamilton, Federalist no. 73, 1788/2003, p. 358).
49 Some have supposed that the basic law should be literally unalterable. The last clause in
the constitution that John Locke wrote for the Carolina colony reads, ‘These Fundamental
Constitutions . . . shall remain the sacred and unalterable form and rule of Carolina for ever’
(Locke 1669/1983, article 120, p. 232). Thomas Jefferson (1816) thought that constitutions should be
replaced every nineteen years, at which point a majority of those alive when the original was writ-
ten (in the early nineteenth century anyway) would have died (on the Condorcet connection to this
idea, see McLean and Hewitt (1994, pp. 58–9); Sommerlad and McLean (1989, pp. 319–24)). Most,
however, are with Madison (Federalist no. 49, 1788/2003, p. 246) in saying that constitutions ought
be amendable but only on ‘certain great and extraordinary occasions’. Out of ninety-one constitu-
tions that Lagerspetz (2002, p. 269) examined, only a dozen can be amended by simple bare major-
ities, and only a couple require unanimity for amendment; most require supermajorities for their
amendment, with two-thirds being the most common (in fifty-seven constitutions).
50 In Federalist no. 73, Hamilton (1788/2003, p. 359) writes of the ‘mischiefs of . . . inconstancy
and mutability in the laws’ as ‘the greatest blemish in the character . . . of our governments’. He goes
on to say, ‘every institution calculated to restrain the excess of law-making, and to keep things in
the same state, in which they may happen to be at any given time, . . . [is] much more likely to do
good than harm; because it is favorable to greater stability in the system of legislation.’ For a con-
temporary echo see Samuelson and Zeckhauser (1988).
51 Cf. Buchanan and Tullock 1962; Barry 1965, pp. 244, 312–16; Rae 1975, p. 1279. In Federalist
no. 58, Madison (1788/2003, p. 286) before them had warned that, with supermajority require-
ments, ‘it would be no longer the majority that would rule; the power would be transferred to the
minority. . . . [A]n interested minority might take advantage of it to screen themselves from equit-
able sacrifices to the general weal, or . . . to extort unreasonable indulgences’.
In addition to all those pragmatic and political reasons for imposing a

s upermajority rule, there is also sometimes offered an epistemic rationale as well.
Something like that thought was enunciated by Pope Pius II when, reflecting
upon the two-thirds rule by which he was elected pope in 1458, he declaimed:
‘What is done by two thirds of the sacred college [of cardinals], that is surely of
the Holy Ghost, which may not be resisted.’52 It is the epistemic claim in favour
of supermajority rules that will be our concern here.
The sort of supermajority rules ordinarily employed in politics are asymmet-
rical in nature. That is to say, there is some ‘default option’ that wins if the
supermajority is not attained, thus biasing decisions in favour of that option over
all others.53 Suppose 60 per cent of voters favour option φ and only 40 per cent
favour not-φ. If a two-thirds supermajority requirement is in operation and
not-φ is the default option, then not-φ prevails even though it is favoured by
only 40 per cent of voters.
It is logically possible to make any option the default option. But politically,
the status quo is almost invariably taken to be the default option.54 The ques-
tion then is whether there is any good epistemic justification for building such
a status quo bias into political decisions.
Once again, there might be all sorts of good practical reasons for doing so.
The basic law embodied in the constitution, for example, is the sort of thing
that people rely on remaining constant for purposes of their own long-term
planning of their affairs.55 And stability, of course, is one thing that super-
majority requirements biased in favour of the status quo bring with them.
Without denying for a moment the importance of those practical considerations,
for the purposes of this book it is the epistemic implications of supermajority
rules with which we are concerned.
Are there, then, any good epistemic reason to think that political decisions
should be biased in favour of the status quo? Is the status quo more likely to
be the correct outcome, epistemically, than any challenger? Thomas Jefferson
52 Quoted in Colomer and McLean 1998, p. 11.

53 It is logically possible to have ‘symmetrical special majority rules’—Scottish juries for example
have the option, besides judging the accused ‘innocent’ or ‘guilty’, of returning a verdict of ‘case
not proven’. But none of the supermajority rules ordinarily employed in politics are of that nature.
See Goodin and List (2006) on the logical structure of supermajority rules.
54 Supermajorities can, however, be used to drive change as well. Condorcet (1787/1994, Letter 2,
p. 300), for example, suggested allowing members of parliament to be elected by simple plurality
in the first instance, but to be re-elected only with a three-quarters majority, and to be re-elected
again after that only with four-fifths majorities.
55 Goodin 2012, ch. 2. Thomas Jefferson (1816), who in the same letter expressed the thought
that it would be ideal to have a new constitution every nineteen years, concluded that as a matter
of practicality it would be better to bias the constitution against too-easy change: ‘I am certainly
not an advocate for frequent and untried changes in laws and constitutions. I think moderate
imperfections had better be borne with; because, when once known, we accommodate ourselves
to them, and find practical means of correcting their ill effects.’
thought not, even when it came to the basic law of the constitution. He poign-
antly writes,
Some men look at constitutions with sanctimonious reverence, and deem
them . . . too sacred to be touched. They ascribe to the men of the preceding age a
wisdom more than human, and suppose what they did to be beyond amendment.
I knew that age well; I belonged to it, and labored with it. . . . Let us [not] . . . believe
that one generation is not as capable as another of taking care of itself, and of
ordering its own affairs.56
Advocates of small government might, of course, say that having no law on any
given subject is likely to be better than having any law on it at all. Supermajority
rules might facilitate that outcome, were we starting de novo with a completely
blank statute book. But that almost never happens: even after revolution or
conquest, the laws of the ancien régime remain in force until altered or
replaced.57 And insofar as the status quo is one in which there are already lots
of laws on the books, the status quo bias built into supermajority rules would
make it harder to pass legislation to repeal them, much to the chagrin of advo-
cates of small government. So even if those commentators are right about what
the correct political outcome is (something we do not remotely concede),
supermajority rules’ status quo bias would be epistemically very much a mixed
blessing.
Another obvious way to justify a supermajority rule’s status quo bias would
be to claim that the status quo is more likely to be correct than incorrect. But
our arguments in Chapter 10 tell against placing any great faith in ‘the wisdom
of the ages’. Perhaps there is an argument for sticking with the ‘tried and tested’
as a matter of precaution—but building that principle of prudence into any
strong presumption seems excessively cautious and implausibly conservative.58
That said, there may be certain very special circumstances in which we have
good grounds for believing that the status quo is much more likely than any
other alternative.59 Suppose we face a decision between the status quo (SQ)
and an alternative (A), and we know that the ex ante probability of SQ being the
correct choice is much higher than that of A being correct. For example, sup-
pose an activist prosecutor is known to bring too many cases to court.60 Then
we know, ex ante, that acquittal (the SQ) is much more likely correct than
conviction (A). Assuming jurors who are sincere but not extraordinarily
56 Jefferson 1816.
57 As Blackstone (1765, vol. 1, introduction, sec. 4, p. 105) writes, ‘in conquered or ceded coun-
tries, that have already laws of their own, the king may indeed alter and change those laws; but, till
he does actually change them, the ancient laws of the country remain’.
58 Harsanyi 1975.
59 This is in contrast to our assumption above that a non-decision (to remain with the status
quo) is an incorrect decision.
60 Or just suppose that the base rate of offending is known to be low, so there are many more
opportunities for a ‘false positive’ than a ‘false negative’ (Goodin 1985).
competent,61 then it makes sense to demand a supermajority for conviction to

bring down the number of false convictions (at the cost of letting more guilty
defendants walk free).62 Under these specific circumstances it makes sense not
to disregard the ex ante probability and to demand a higher number of votes for
conviction than for acquittal.
A second epistemic reason for preferring supermajorities might be that we
sometimes have reliable knowledge about undesirable common causes influ-
encing some voters. If a subset of voters is influenced by a strong opinion leader,
for example, and we want to prevent the opinion leader from swaying the
decision, then demanding a supermajority may make good epistemic sense.63
The price to pay, of course, is a bias towards the status quo. Sometimes this is a
price we are willing to pay in order to prevent inappropriate influence from
epistemically compromising the electoral outcome.
The best justifications for asymmetrical majority rule typically derive, how-
ever, not from any asymmetry in the probability that one or the other outcome
is correct, but rather from an asymmetry in the cost of erring in one direction
compared to another. Court cases once again provide a good example. It is
arguably much worse to convict and severely sanction an innocent defendant
than to let a guilty defendant go. We are standardly enjoined to bias criminal
procedures in such a way as to err on the side of false acquittal, for that reason.64
In the words of the US Supreme Court, ‘Where one party has at stake an interest
of transcending value – as a criminal defendant his liberty – . . . [we protect him
by] placing on the other party the burden of persuading [the jury] . . . of his guilt
beyond a reasonable doubt.’65
A second reason along those lines relates to the kind of justification we are
required to provide for our decisions. Consider, again, the decision of a jury.
Let us assume that we choose an unusually large and competent jury, say
101 jurors with competence pc = 0.6 who are collectively Pn = 0.979 competent.
This means that ex ante the probability of a correct decision is 97.9 per cent,
which is quite high. However, now suppose that a defendant gets convicted in
a narrow 51:50 vote. Ex post, given that outcome of the vote, we are not very
confident that the defendant is guilty. Indeed, the evidence provided by that
vote is equal to the evidence of one single juror voting to convict, which is
not very good evidence at all.66 Thus, even though we know ex ante that major-
ity decisions of the competent jury are very likely to be correct, the ex post
61 In line with our discussion in Section 4.3.3.

62 Though for epistemic reasons it would be better to define the supermajority requirement not
in terms of a proportion but in terms of an absolute margin; see List 2004.
63 Particularly if that is the only way to prevent that from happening—that is, there is no prac-
tical way of disenfranchising or discounting voters subject to those untoward influences.
64 In Blackstone’s famous formulation, it is better for ten guilty people to go unpunished than
that one innocent person be punished (Blackstone 1765, bk. 4, ch. 27). Others have suggested
ratios between half and double that (Fletcher 1968, p. 881).
65 Brennan 1958, pp. 525–6; see also Harlan 1970, p. 372. 66 See again List 2004.
probability given the specific vote can be much lower. If we have to provide a
strong justification for our decision to depart from the status quo for every
token decision and not just the type of decision, then we should demand super-
majorities to depart from the status quo.67 And indeed, this is how we typically
think about the burden of proof in court cases.
17.4.2 Presidential Vetoes
Another epistemic bottleneck lies in the power of the US President to veto

legislation passed by Congress, thus preventing it from becoming law. If Congress
is no longer in session, the President can simply not sign its enactment, in
which case it does not become law and does not have to be returned to Congress
for any further consideration.68 But if Congress is still in session, the President
must veto it within ten days or else it will automatically become law without
his signature; and if he vetoes it, it must be returned to Congress for further
consideration. If two-thirds of each house of Congress then vote in favour of
the enactment, the President’s veto is overriden and the enactment becomes
law. The rationale that the Federalist gives for imposing that two-thirds vote
requirement is straightforwardly epistemic. It is, Hamilton says, simply ‘far less
probable’ that two-thirds of each house would hold mistaken views than that a
‘bare majority’ would.69
But just how much epistemic confidence remains to be gained, anyway, in a
legislature the size of the current US Congress? Suppose for the sake of argu-
ment that each member of Congress is individually pc = 0.55 likely to vote for
the correct alternative. With 435 Representatives and one hundred Senators,
the probability that a simple majority in both chambers will be correct is at
least Pn = 0.828. This is the probability that both chambers will vote for the
correct decision. If we assume that diverging decisions leave the status quo in
place, and that the status quo is correct in half of the cases, then that number
goes up to 0.913. In short, we already have great confidence that Congress is
correct in its original enactment, in a legislature the size of today’s Congress.70
Ironically, imposing a supermajority rule usually reduces the chances that
the correct alternative will emerge as victorious. Suppose that the two alternatives
67 Tribe 1971. 68 That is what is called a ‘pocket veto’.

69 In Federalist no. 73 again, Hamilton (1788/2003, p. 361) writes:
It is to be hoped that it will not often happen, that improper views will govern so large a
proportion as two-thirds of both houses of the Legislature at the same time; and this too in
defiance of the counterpoising weight of the executive. It is at any rate far less probable, that
this should be the case, than that such views should taint the resolutions and conduct of a
bare majority.
70 See calculations reported in Table 17.4. It would have been otherwise in the much smaller
first Congress: there the probability that a majority of both houses would be correct was only
Pn = 0.585; as we said in Section 17.2.2, the small twenty-four-member Senate in that first Congress
acted as a significant epistemic bottleneck.
are the status quo SQ and one alternative A, and both are ex ante equally likely
to be correct. Then supermajority rules reduce the probability of the correct
alternatives winning. In a Congress with 435 Representatives and one hundred
Senators, each of whom is individually pc = 0.55 likely to be correct, the decision
will be in favour of the status quo with a probabilty of nearly 100 per cent because
voters of that competence almost never reach a two-thirds majority. This means
that the decision will be correct if and only if the status quo turns out to be the
correct choice, which ex hypothesi it is in only 50 per cent of cases. One might
as well just flip a coin, or stick with the status quo without voting. Obviously,
much hinges on the competence parameter here. If the correct alternative is just
‘blindingly obvious’ and everyone will immediately recognize that, then pc will
be higher and two-thirds majorities reached rather more easily. In particular, if the
competence of voters pc is larger than the required threshold, then the extreme
status quo bias disappears and a jury theorem for supermajorities applies.71
To impose a supermajority requirement for overriding the president’s veto
is, in effect, to set a strong presumption that the president is more likely to be
correct than the Congress. What grounds could we have for thinking that
that is so? The president, after all, is only one person, whereas the Congress is
comprised of many members.
Hamilton responds by pointing to some pathologies of collective action in
settings like Congress. In answer to the challenge ‘it was not to be presumed a
single man would possess more virtue and wisdom than a number of men’,
Hamilton writes:
The propriety of . . . [giving the President a qualified veto] does not turn upon the
supposition of superior wisdom or virtue in the Executive, but upon the supposition
that the legislature will not be infallible; that the love of power may sometimes
betray it into a disposition to encroach upon the rights of other members of the
government; that a spirit of faction may sometimes pervert its deliberations; that
impressions of the moment may sometimes hurry it into measures which itself, on
maturer reflexion, would condemn.72
We will consider the ‘in the heat of the moment’ worry in our discussion of
political cooling-off periods in Section 17.4.3. Focus here on the worry about
the spirit of faction and the temptation to encroach upon the prerogatives of
other branches of government.
Hamilton’s thought, presumably, is that no faction is likely to be so strong as
to command two-thirds support in both chambers of the legislature. Anything
that passes both houses of Congress by that margin is thus highly likely to
be in the public interest.73 So far so good. But now remember what it is that
triggers that double two-thirds requirement: the president’s veto.
71 Fey 2003.
72 Hamilton, Federalist no. 73, 1788/2003, p. 358–9. Vermeule 2011a, pp. 1437.
73 Goodin 1996.
What right do we have to assume that the president’s judgement will not be
prey to such distortions as well? Perhaps the solitary person of the president
cannot itself be divided into factions. But the president can be partisan, allied
to one of the factions in the legislature. And encroachments by the ‘imperial
presidency’ are as much to be feared as any from the legislative branch.74
Yet if the president’s judgement is subject to the same kinds of distortions,
why privilege his judgement as expressed in a veto over the judgement of the
Congress, to the extent that a two-thirds vote in both chambers should be
required to overcome it? Perhaps there are informal constraints that prevent
the president from using his veto too often. Or perhaps the president is more
likely to be correct, because he is under enormous scrutiny and has more
epistemic expertise due to the advisors at his disposal. These factors could reduce
the epistemic disadvantage of this bottleneck. Nevertheless, from an epistemic
point of view, the presidential veto is likely to be problematic.
Suppose the alternatives ‘Status Quo’ (SQ) and ‘Alternative’ (A) are equally
likely to be correct.75 Consider the different pathways to a correct decision:
1. A is correct and the majority of both houses and the President agree to it;
2. A is correct and is supported by the majority of both houses, the President
vetoes it, but A is reinstated by both houses;
3. SQ is correct and either:
(a) A fails to win a majority in at least one of the houses, or
(b) both houses vote for A, but the President vetoes A and A is not
reinstated by Congress.
Looking at these pathways we can calculate how likely a correct decision is
with that many veto points. Remember that Congress alone, when assuming
that keeping the status quo is correct half of the time, will get it right in
about 91.3 per cent of cases. That number drops to only 72.7 per cent with the
presidential veto added. And the reason for that is clear: the president vetos
much too often when A is correct, and that cannot be compensated for by the
few additional vetos in the unlikely event that the SQ is true but both houses
have voted for A erroneously.
17.4.3 Political Cooling-off Periods
John Rawls, like Condorcet before him, praises ‘constitutional arrangements

[that] compel a majority to delay putting its will into effect and force it to make
74 Schlesinger 1973.
75 Our default assumption that a decision failure is counted as an incorrect decision does not
make sense here, as such an assumption would render the president’s veto utterly pointless.
a more considered and deliberate decision’. Such arrangements serve, he says,

‘to mitigate the defects of the majority principle’.76
The point of such cooling-off periods, as conventionally understood, is to
prevent premature closure on decisions that may be epistemically unsatisfactory,
and to force the decision maker to ‘think again’.77 One such reason we might
need to ‘think again’ has to do with bandwagon effects. Once a bandwagon is
decisively rolling towards some seemingly inevitable end, everyone wants to
get onto it quick smart.78 Those who were ‘for Kennedy before Wisconsin’ (the
primary election in which he was established as a leading candidate) had a
special kind of political credit with the new administration.79 Likewise with
novel policy proposals: maybe no one wants to be the first to make some
seemingly far-fetched suggestion, but no one wants to be the last to endorse it,
either; and once the proposal gets up a good head of steam, previously reluctant
people rush to get on board. Such a ‘rush to judgement’ can often lead to bad,
ill-considered decisions.
Consumer-protection legislation provides a model for how to avoid such a
rush to judgement. It often provides for a cooling-off period, during which goods
can be returned for a full refund, or contracts annulled, if upon reflection the
consumer comes to regret his initial decision.
Political institutions often make provision for just such cooling-off periods.
Consider these three examples:
1. Since 1911, the UK House of Lords can no longer veto legislation passed
by the House of Commons but merely delay its becoming law (initially by
two years, one since 1949). If the Commons passes the same bill a year
later, it goes directly to the sovereign for royal assent.80 While the House
76 Rawls 1999, p. 201. On Condorcet’s views see Urbinati (2006, ch. 6). It is unclear exactly what
‘defects of the majority principle’ Rawls had in mind. Mandatory cooling-off periods clearly
cannot prevent a steadfast majority from exercising tyrannical power over the minority, if it is
determined to do so. But arrangements slowing down the political process can at least ensure that
the majority action is ‘considered and deliberate’, rather than being rushed and reckless tyranny
practised ‘in the heat of the moment’. Many might well say (with some cause) that ensuring that
majority tyranny is at least ‘considered and deliberate’ is small consolation. Still, it may actually
matter, if some substantial proportion of acts of majority tyranny are ‘in the heat of the moment’
acts that even the majority would regret in a cooler moment.
77 One measure of the rate of ‘legislative errors’ caused by haste is found in ‘star prints’ in the
legislative record in the US, indicating that corrected text has been substituted for previously
erroneous text (Lewallen 2016).
78 Brams 1978.
79 Sorensen 1965, p. 253. ‘For a moment Ted Sorensen suggested a point system [for filling
subcabinet level appointments] – so many points for having been with Kennedy before [the]
Wisconsin [primary election], so many for having been with Kennedy at [the] Los Angeles
[nominating convention], and so on[. B]ut’, (Schlesinger 1965, p. 148) continues, ‘the idea soon
seemed irrelevant’—at least for filling senior posts in the administration, if not for all sorts of
other purposes.
80 Under the long-standing conventions of the constitution, that is never withheld. UK Parliament
Act 1911, 13, Geo 5, as amended by the Parliament Act 1949, 12, 13, & 14 Geo. 6. c. 103.
of Commons is certain to prevail in the end, by rejecting a bill the House

of Lords can nonetheless force the Commons to return and reconsider
the wisdom of that legislation, a year later.81
2. Under the Australian Constitution, if the Senate twice rejects legislation
passed by the House of Representatives—on occasions separated by at
least three months—the Prime Minister may ask the Governor General
for a double dissolution of Parliament, triggering both a new election
for the lower house and a full Senate election (ordinarily only half of
Senators are up for re-election in any given election).82 Again, requiring
the Senate to reconsider the same bill again three months later serves as a
cooling-off period.83
3. Bicameralism, more generally, serves to slow down the process of legisla-
tion. The lower house may well rush to judgement, in approving any given
proposal. But the upper house—often styled the ‘deliberative chamber’84—
considers the proposal at leisure, and sends a revised version of the bill
back to the lower house sometime later for further consideration.85
Of course, the upper chamber might spot some other flaw in the decision process.
Perhaps the upper chamber sees what the lower chamber fails to see—namely
that virtually all of their lower-house colleagues follow the same opinion leader,
or bias, or ideology, or whatever else clouds their good judgement. The view from
outside is sometimes required in order to spot the mistake. Returning the decision
to the lower chamber for another, hopefully more careful, decision could then be
a good idea on those grounds as well. But the distinctively ‘cooling-off-period’
aspect of those arrangements is that the upper house takes its time with the
proposed legislation, and sends its response back to the lower house for a
rethink only after the passage of some time.
81 Similarly, under the short-lived French Constitution of 1791 (ch. 3, sec. 2):
In the case the King refuses his consent [to legislation], such refusal shall be only suspensive.
When the two legislatures following the one in which the decree was introduced have
again successively presented the same decree in the same terms, the King shall be deemed
to have given his sanction.
82 Australian Constitution, section 57.
83 As does the period of time it takes to hold the new election, and the further period of time it
takes for the newly elected House and Senate (first initially, and failing passage in both houses
then jointly) to approve the legislation that triggered the double dissolution.
84 Uhr 1998.
85 In Federalist no. 63, Madison (1788/2003, p. 307) defended the Senate as a ‘temperate and
respectable body of citizens’ designed to ‘defen[d] . . . the people against their own temporary
errors and delusions’. Speaking at the Philadelphia Convention, Madison (1787/1911, p. 421) had
similarly described the ‘ends to be served’ by the upper house as follows: ‘These were first to
protect the people against their rulers: secondly, to protect [the people] against the transient
impressions into which they themselves might be led.’ He went on to extol the virtues of ‘a people
deliberating in a temperate moment’ (quoted in Elster 2000, p. 376). Canada’s first prime minister,
Sir John A. Macdonald, likewise described the Canadian Senate as ‘an institution of “sober second
thought” ’ (quoted in Vermeule 2011a, p. 1437).
The thought behind a cooling-off period is that that passage of time might
prevent a rush to judgement by interrupting a bandwagon or cascade. In the
model we developed in Chapter 10, cascades simply involved sequential voting
in which later votes are a function of earlier ones. In the real political world,
cascades are driven not just by ‘how many’ people have come on board, but by
‘how recently’ as well. The ‘timing’ element is crucial. Lots of recent endorsements
are more compelling than the same number of endorsements collected over a
much longer period of time. That is how bandwagons work, psychologically.
And that is precisely what cooling-off periods hope to break.
Here is one way of modelling the process: Suppose that, when deciding
whether to vote for any given proposal, each voter gives each other’s report
equal credence to his own.86 And suppose that voters vote sequentially, one
after another, with each voter deciding how to vote by combining all previous
votes together with his own private signal as to which is the right way to vote.
Clearly, what is going to happen is that early votes in that sequence will exert
disproportionate influence over subsequent votes. That is just the model of a
‘cascade’ that we developed in Chapter 10.87
Suppose now, however, that voters can remember each others’ votes for
only a limited period of time. And suppose that the institutional arrangements
are such as to call for the same body to re-vote on the same proposition, after
some period longer than that. That provision would function as a ‘reset button’,
ensuring that the second balloting operates independently of the first.88
On that model, imposing a political cooling-off period that is longer than
voters’ memories enables us to make two independent assessments of group
opinion, and this can be useful, even when we know that the group will be
prone to bandwagons and cascades on both occasions. If both cascades are in
the same direction, despite their different starting points, then that provides
reassurance of a sort that the result is the correct one. That, anyway, is the
epistemic thought behind political cooling-off periods.
But are cooling-off periods, in general, really such a great idea? Not neces-
sarily. Central to the political cooling-off-period strategy is a decision in stages.
Assume that the rules of this procedure are such that, to emerge as the overall
winner, the correct proposition has to win in the first stage, and then win again
86 In terms of Section 10.3, each ‘weights’ each other’s vote equally to his own.
87 See similarly Vermeule 2011a, pp. 1455–6.
88 Of course if the order in which votes are taken is fixed, and it is the same in the second
election as the first, we will get the same outcome as before, provided that the voters’ private
signals don’t change. But suppose voting order is random, so that different voters will get to
enjoy the privileged position of being among the very first to vote—and hence it will be different
people whose position other later voters take disproportionately into account when casting
their own ballots. We will still get a cascade in the second election, of course. But since it has a
different starting point (different people vote first), the second cascade may well not be in the
same direction as the first.
in a later stage.89 The probability of the combination of those two victories is

the probability of the first multiplied by the probability of the second; and since
each of those probabilities is less than one, the product of multiplying them is
smaller than either of those separate probabilities alone.90 Hence, we would
have been epistemically better off going with the results of the first ballot alone.91
However low the probability that that first outcome is correct, and however
high the probability that the second is correct, the probability of the correct
outcome prevailing on both occasions is no better than (and most typically
substantially worse than) the probability of the correct outcome prevailing on
the first occasion alone. That is the epistemic case against cooling-off periods.
For cooling-off periods to have any epistemic merit, some fairly specific
assumptions are required. Perhaps the very existence of cooling-off periods
might lead to a substantial increase in epistemic competence across rounds
because it induces the voters to reflect more carefully. Or perhaps upper
chamber triggering the cooling-off period is very good at spotting epistemic-
ally flawed decision procedures in the lower chamber, and the lower chamber
tends to get better when they get prompted to rethink. Or perhaps there is
something special about the status quo, in which case deadlock might be a
good thing, as discussed above.
We here take no stand on the empirical question of whether such assump-
tions might be warranted, either in general or in any particular settings. We
merely note the form of argument that would have to be given in order for
political cooling-off periods to have any epistemic merit.
89 This is extensionally equivalent to the rule Vermeule (2011a, p. 1461) describes operating in
the Italian Parliament prior to 1988, whereby ‘Parliament would . . . hold both an open and a secret
vote in succession on bills designated as issues of confidence, meaning that the government would
fall if the bill were defeated. The background norm was . . . that the government would fall if the bill
failed on either the open or the secret vote.’ Another way the rules of the procedure might work
would be for the first stage to be a non-binding ‘straw poll’ initially, and for the second-stage vote
to be the only binding one. Formally, of course, the first-stage vote would then be of no conse-
quence; but informally it might help in guiding discussions leading up to the second vote (which
is why this is often done in jury deliberations, for example) (Vermeule 2015, pp. 223–4).
90 Assuming independence, which may not hold here. But even with dependence the reduc-
tion in probability holds, though to a lesser extent.
91 Or the second alone: as Vermeule (2011a, p. 1450) observes, ‘the basic puzzle about cooling-
off justifications is whether the first, overheated [stage] has any epistemic value at all. . . . Under
bicameralism, if the upper house supplies the sober second thought, why have a politically intem-
perate lower house in the first place?’
18
Institutional Aids to Epistemic Success
The previous chapter discussed an array of ways in which different institutional

structures can compromise the epistemic performance of the political process.
Those are things to be avoided. Next we will consider how the institutional
structures of representative democracy can be arranged in ways that might
positively enhance its epistemic performance.
A word of warning at the outset. We will be discussing mechanisms that
can—but only ‘can’, not necessarily ‘will’—enhance the system’s epistemic
performance. That is to say, there is no guarantee that the mechanisms we
will be discussing will invariably have the happy epistemic consequences we
contemplate. Typically, various other favourable conditions must also be in
place in order for them to do so; and fully exploring those is beyond the scope
of this chapter, or indeed this book. Our ambitions here are more modest: to
explore possibilities rather than to establish necessities.
18.1 MECHANISMS TO MAKE DECISION

SITUATIONS MORE TRUTH CONDUCIVE
As we said in Section 5.5, decision situations can be more or less truth-conducive.

If for example the correct alternative is not on the agenda, then it cannot be
chosen, not even by our best responder from Chapter 5. Or, again, if the evi-
dence is systematically misleading, then even the best responder would be misled
by it. Here we will discuss briefly some institutional mechanisms that might
make the decision situation more truth-conducive in each of those respects.
18.1.1 Finding New, Better Alternatives
There are various ways by which new alternatives might be discovered and
added to the agenda for decision. One, as discussed in Section 9.2.5, is through
deliberation in groups small enough to be conversable.1
1 Vermeule (2011a, p. 1451) remarks upon how ‘hot deliberation’ can generate lots of new ideas
that cooler heads can then winnow.
Institutional Aids to Epistemic Success 289
Within the formal institutions of representative democracy, there are

three classic examples of smaller and more conversable bodies being used to
expand as well as to explore the agenda. The first example is the smaller upper
chamber of a bicameral legislature, which is conventionally referred to as the
‘deliberative chamber’. The second example is that of a small specialist commit-
tee within the legislature.2 The third example is that of an advisory body within
the executive.
In the first two of those cases, the smaller body plays a decisive role in the
formal political decision-making process. Both the upper house and the
legislative committee can, each in its own way, block or delay legislation from
being enacted by the larger house. And that comes with clear epistemic costs,
as already discussed in Section 17.2. While those costs are (somewhat specu-
latively) quantifiable with the aid of stylized models such as we have been
using throughout this book, the epistemic benefits are not. So it is hard to say
whether those costs are outweighed by the epistemic gains that come from
those smaller deliberative bodies discovering new and better alternatives. But
maybe they are.
Purely advisory committees, whose role is to make recommendations but
not decisions, carry with them no such epistemic costs. Their smaller size does
not compromise the overall epistemic success of the decision-making process,
because they have no veto in the process. Yet thanks to their small size, they can
be genuinely deliberative; and they might in consequence come up with new
and better options for the political agenda, and in that way contribute to the
overall epistemic performance of the decision-making process.
With purely advisory committees, therefore, it is epistemically all gain and
no loss.3 But how much the gain may be depends crucially on the advisory
panel’s recommendations of new items for the agenda being taken up by the
formal decision-making system. Insofar as the advisory committees have been
officially commissioned by other powerful political actors, however, there is
some reason to hope that they might succeed in that.
The same sorts of functions could, in principle, be performed by outside
think tanks or self-organized groups of experts offering advice to governments.
In those cases, however, there is less reason for confidence that decision makers
will act on their recommendations of items to add to the agenda. And the same
2 As Maass (1983, p. 39) says,

The first justification for committees relates to the legislative process. Congress’s role of
overseeing and controlling the process includes the capacity to offer, debate, and adopt
alternatives to the Executive’s proposals. The President initiates, to be sure, but Congress is
not limited to his proposals in its deliberations. Without committees to examine alternatives
and to bring out relevant details pertaining to them, Congress would have to assume,
in effect, that the Executive had examined all alternatives and selected the best one. We
[in the US] have never been willing to assume either of these propositions, necessarily. In
comparison, the British parliament, without an effective system of standing committees, is
forced to assume both.
3 Unless, of course, the new alternatives are unhelpful or even more confusing.
is of course all the more true of the ad hoc ‘minipublics’ whose virtues today’s
‘deliberative democrats’ extol.4 Even if those minipublics perform as deliberative
democrats most deeply desire, there is the very major problem of finding
some way to secure uptake of their recommendations by the central decision-
making institutions of society.5 Many instances of uptake can be found.6 But
the mechanisms by which it occurred seem more of a hodgepodge rather than
anything genuinely systematic.7
Habermas offers a ‘two-track’ model of deliberative democracy that relies on
informal conversations that occur across civil society, rather than intentionally
orchestrated ‘minipublics’, as its deliberative inputs. But once again, the trick
lies in connecting those informal conversations, somehow, to the institutions
that are formally empowered to enact laws and policies.8 Political parties are
his proffered mechanism, although there is of course no guarantee that parties
will necessarily pick up and report systematically all good ideas for the agenda
that might emerge across the entire range of civil society—in dinner-table con-
versations, in barroom arguments, or even at academic seminars.
18.1.2 Weeding Out Bad, Confusing Alternatives
So far we have been focusing on ways of improving the decision situation by

adding new, better alternatives to the agenda. But the converse, weeding out
incorrect alternatives, can help to improve the decision situation as well. If the
agenda is too cluttered—if there are too many alternatives to consider which
are too superficially similar to one another—the decision situation can be
confusing and diminish the capacity of even the best responder to choose
the correct alternative among them. Eliminating confusing alternatives from
the agenda would therefore be another way to improve the decision situation.
In principle there are several ways that that might be done. But the most
common, in representative democracies, all involve devolving that ‘winnow-
ing’ task to some smaller subset of the larger decision-making group.9 The
most familiar example of that is a specialist committee of the legislature, which
4 Fishkin 1995; 2009. Fung 2003. Smith 2009.

5 What is needed, in Dryzek’s (2009, pp. 1385–6) terms, is some ‘transmission’ mechanism to
convey messages from ‘public space’ to ‘empowered space’.
6 Goodin and Dryzek 2006.
7 Dryzek’s (2009, p. 1385) list only serves to emphasize that fact: ‘These means might involve
political campaigns, the deployment of rhetoric, the making of arguments, or cultural change
effected by social movements that come to pervade the understandings of formally empowered
actors, or personal links between actors in the two kinds of spaces.’
8 Habermas 1996, chs 7–8.
9 In principle, advisory panels of experts could play this role. Christiano (2012, p. 42)
recommends precisely that when writing: ‘expertise acts as a kind of external filter on the
deliberations of other parts of the division of labour such as politicians and ordinary citizens.
It rules out certain theories as possible bases of policy-making and permits choice among a
certain small subset of theories for policy-making.’ There would, however, be a clear democratic
is charged with the task of choosing, among the many proposed pieces of
legislation, which of them should be put to a vote in the parent assembly. We
offered some sample calculations to suggest the epistemic value of such arrange-
ments in Section 8.2.3. On the scenario there contemplated, at least, this would
seem to be a highly promising epistemic strategy.
Political parties themselves might play an epistemically valuable role, not
only in searching out new and better alternatives to put onto the agenda (in ways
discussed in the previous section) but also in winnowing the agenda to be
voted on—by the electorate in the first instance, and by the legislature in the
second. Habermas’s two-track model sees political parties primarily as trans-
mission belts between civil society and the legislature; and clearly they are that,
too. But in this section we will be thinking of them primarily as ‘search parties’
of the sort we discussed in Section 8.1.
There is obviously much to be said both for and against political parties and
the partisanship that they induce. As already discussed in Chapter 14, Rousseau,
Condorcet, and the American Founders were all dead set against ‘factions’ of that
sort.10 They thought the narrow pursuit of sectoral interests that they promote is
antithetical to pursuit of the common good for the community as a whole. Yet
others suppose that there is much to be said for a ‘loyal opposition’ in holding
the government of the day to account, and offering voters a well-worked-out
programme of what an alternative government would do instead.11 Those are all
very important considerations. But our focus in this book, remember, is purely
on the effect that political parties might have on the epistemic performance of
the political system.
Political parties serve a great many functions, ranging from mobilizing voters
to dispensing patronage.12 But one of their principal services, from an epistemic
point of view, lies in performing a ‘search’ function for voters.13 We discussed
in Section 8.1 what might be the epistemic advantages of a group sending out
search parties, each charged with scouring some particular locale for the best
option in that neighbourhood and reporting back the results to the group as a
whole. We can think of political parties as performing just such a function.
Political parties, in effect, ‘own’ a certain patch of ideological space.14 They
search, within that space, for the best proposal to offer voters, come the next
(if not necessarily epistemic) cost to devolving the role of restricting the agenda to a democratically
unaccountable body.
10 Rousseau 1762/1997, bk. 2, ch. 3. Condorcet 1787/1994, Letter 2, p. 301; 1789/1994, pp. 178–9.
Among the American Founders, see e.g. Madison Federalist no. 10 (1787/2003) and Washington’s
‘Farewell Address’ (1796/1966, p. 201) drafted for him by Hamilton.
11 Schumpeter 1942. APSA 1950. Simmel 1950. Dahl 1972. Goodin 2003, ch. 7; 2008, ch. 10.
Rosenblum 2008; 2014. White and Ypi 2011; 2016.
12 Katz 1980.
13 Another, as discussed in Chapter 12, lies in providing reliable cues to inform voters as to how
they should vote, from their own point of view.
14 Budge et al. 1987. Klingemann et al. 1994. Hammond and Humes 1993. Seeberg 2017.
election. At that election, voters then choose among those ‘best offers’ that each
party can make within the ideological space that they call their own.
Something like that was at the heart of Joseph Schumpeter’s theory of democ-
racy.15 On his pessimistic view, voters should not realistically be asked to perform
any very challenging tasks. It is the role of political parties to simplify the choice
for them. Given parties, voters merely need to choose among the alternative
‘teams’ and associated policy positions presented to them at periodic elections.
In our own day, Sniderman has proposed a similar role for political parties.16 They
structure choices for voters, reducing the menu for choice to something that can
realistically be coped with by relatively poorly informed and inattentive voters.
Whether the procedure proposed here is epistemically altogether good news,
however, may well be questionable on a couple of grounds. Recall, first, our
discussion in Section 8.1.4. If voters ‘recognized the truth when they saw it’, at
least on the limited sort of menu that parties present to them, that would be one
thing. But if instead we have voters who are imperfect assessors in choosing
among the options presented to them by parties (who are also imperfect asses-
sors themselves), those imperfections are multiplied through that two-stage
choice process. That entails a certain epistemic cost, which will always make
the two-stage process epistemically inferior to a one-stage process—unless (as
in the example in Section 8.2.1) splitting the process into multiple stages makes
agents epistemically more reliable on the component choices.
Furthermore, it may well be thought that what is at stake in competition
between political parties is a clash of values, or anyway of priorities, in which
there is no factually correct answer to be found. We have argued in Chapter 13
that a version of the CJT might still apply in such circumstances. Even if there
is no factually correct answer to the question of ‘which is the right value or
priority?’, at least there is a factually correct answer to the question of ‘what is
the correct option, given the value or priority that I actually hold?’. Assuming
the other conditions of the CJT still hold, the majority in the election is likely
(increasingly so, the more voters there are) to choose the option that is the correct
option from the point of view of the majority (or plurality) value or priority.
In any case, where the political parties winnow the options put to the voters,
whether or not that is epistemically useful depends largely upon how the
15 Schumpeter 1942.
16 Sniderman 2000. As Jackman and Sniderman (2002, p. 214) put it:
Political institutions—above all, political parties—coordinate the alternatives open to citizens
for consideration. This coordination of alternatives through political parties characteristically
imposes a specific set of properties—bipolarity, stability and ideological patterning, among
them—on political choices that citizens are asked as citizens to make. . . . [I]t is precisely
because political institutions, especially political parties, organize and coordinate citizens’
choices that large numbers of citizens are able to take advantage of the judgmental shortcuts
and so choose approximately rationally.
See similarly Ebeling 2016.
parties make those decisions. If it is by majority rule among a large number of

party members, then correct outcomes are likely. If it is by diktat of a single
all-powerful party leader, they are much less so.
18.1.3 Improve the Evidence Base
Another way to make the decision situation more truth-conducive is to

improve the evidence base. The best that the best responder can do in tracking
the truth is, of course, constrained by the evidence that is available. If the
evidence is incomplete or misleading, even the best responder may be unlikely
to decide correctly.
There are various ways in which political institutions can improve the evi-
dence base available in society on any given topic. One is to fund more formal
research, of course. But another is to encourage more informal sorts of social
experimentation. Permitting, yea encouraging, different people and different
jurisdictions to try out different ways of doing things—and then sharing the
results across the wider community so people can learn from one another’s
experiences—can also be a great way to improve the evidence base available
within a society.
That is the point of the ‘laboratory of federalism’, with different jurisdictions
trying out different policies which, if successful, other jurisdictions might
then adopt.17 That is John Stuart Mill’s point, when he advocated tolerance
for various different lifestyles, conceived as ‘experiments in living’.18 That is why
national science foundations fund competing labs, and Franklin Roosevelt put
different sets of advisors to work on the same issue.19 So encouraging institu-
tions that practise and promote toleration and pluralism will be one way to
improve the decision situation by expanding the evidence base in society.
Another way of enhancing the evidence base might be through deliberation,
over the course of which different people’s evidence and experiences might
interact and combine in novel ways that are genuinely informative. Such
deliberations can occur in various institutional settings of the sort ordinarily
found in representative democracies: smaller upper houses of the legislature,
legislative committees, cabinets, advisory bodies, and so on. Those have already
been discussed in Section 18.1.1.
We can also craft new deliberative institutions specifically for this pur-
pose. There are several models for doing that. Consensus Conferences of
laypersons have been convened by the Danish Board of Technology to bring
the views of the lay public to bear in an informed way on the assessment of
17 Brandeis 1932. See our discussion in Section 8.1.2.

18 Mill, 1859/1977, ch. 3, pp. 260–1. See further Anderson 1991.
19 Latour and Woolgar 1979. Burns 1956, pp. 373–5.
new technologies.20 Deliberative Polls have been used to shape public debate and
deliberation, perhaps the most interesting—anyway the most surprising—cases
being those convened on behalf of local governments in China to garner
inputs from the public on choices among infrastructure projects.21 In British
Columbia, a Citizens’ Assembly on Electoral Reform crafted and helped inform
public debate on a proposal for a new electoral system that was narrowly defeated
at a subsequent referendum.22 In Oregon, Citizens’ Initiative Review deliberative
minipublics are convened on referenda topics to go before the voters; and on
the basis of their deliberations they prepare a one-page statement for the
state’s official pamphlet circulated to all voters before the election reporting the
group’s ‘key findings’ and principal arguments for and against the proposal.23
A final way of enhancing the evidence base might be to convene panels of
experts to deliberate in similar fashion on the subject of their expertise. Expert
panels convened to advise regulatory bodies are one example of that.24 Scientific
consensus conferences are another.25 When consensus or near-consensus is
achieved, then (assuming there is no evidence of groupthink at work) there are
strong arguments for decision makers to act on the advice of those experts.26 The
problem comes, of course, when no consensus is achieved and some action is
nonetheless required. When experts disagree, we have to choose which among
them we trust the most; and typically we do that by choosing those with whom
we agree the most, on those matters on which we think we too have some com-
petence.27 The upshot is a sort of expertise analogue of ‘venue shopping’, whereby
each side in an argument has its own band of ‘experts’ testifying to the veracity of
its claims.28 And often, in such situations, the effect is to cancel out any effect of
expert advice, with decision makers using ‘their’ experts as people to hide behind
rather than people from which genuinely useful information can be gleaned.29
18.2 MECHANISMS TO INCREASE INDEPENDENCE
The classic CJT assumes that each vote is independent of each other vote,
conditional on the true state of the world. In Chapters 4 and 5, we have shown
that those assumptions can be relaxed substantially, and qualitatively similar
results will still emerge. They may emerge more slowly with increasing numbers of
20 Grundahl 1995. Nielsen et al. 2006. 21 Fishkin et al. 2010. He and Warren 2011.

22 Warren and Pearse 2008. Warren and Gastil 2015, pp. 568–9.
23 Gastil, Richards, and Knobloch 2014. Warren and Gastil 2015, pp. 569–71.
24 Vermeule 2009c. 25 NIH 2013. 26 Vermeule 2009c.
27 Goldman 2001.
28 On Mercier and Sperber’s (2011) account that puts argumentation at the heart of all reason-
ing, this occurs literally all the time.
29 Lundin and Öberg 2014.
voters than they would were the classic Independence Assumption satisfied; and
collective competence may be upper-bounded at something less than Pn = 1.0,
which is the upper bound in the classic CJT setting. Thus, there is something to
be gained epistemically from reducing the interdependence among votes owing
to their dependence on some common causes, even if it cannot be eliminated
altogether. A successful reduction of dependence might increase the conver-
gence threshold and lead to substantial epistemic improvements. Next we shall
discuss some institutional mechanisms designed to do that.
18.2.1 Restricted Franchise, Secret Ballot
One way to help ensure the independence of votes from one another would be
to make independence a precondition for enjoying the franchise at all. And
that, of course, was what was originally done in the early days of democracy.
The franchise was restricted to those who were independent of the will of
any other.30 Servants and slaves were denied votes for that reason. So too were
women, who were presumed to be subject to the will of their husbands or fathers
or other male protectors. So too were those without enough property to make
them independent of any employer. ‘Papists’ too were excluded, on the grounds
that they took orders from the Pope. Those broad categories of domination and
interpersonal dependence might not pick up every idiosyncratic instance of it;
but excluding those broad classes of people who were subject to the will of
some other was supposed to ensure the independence of a person’s vote in early
democratic electorates.
Note well, however: that strategy is only addressed to preventing one person’s
vote depending directly on another’s. It does nothing to prevent cases in which
there is an indirect connection between two people’s votes owing to them both
depending on some common cause that affects them both. If anything, restrict-
ing the franchise in these ways increases the likelihood of that, assuming such
common causes (shared systems of training, acculturation, etc.) are more likely
to be found within the narrow groups picked out in this way than across the
population at large.
30 Blackstone 1765, bk. 1, ch. 2, sec. 5, subsec. 1, p. 171:

If it were probable that every man would give his vote freely and without influence of any
kind, then, upon the true theory and genuine principles of liberty, every member of the
community, however poor, should have a vote in electing those delegates, to whose charge is
committed the disposal of his property, his liberty, and his life. But, since that can hardly be
expected in persons of indigent fortunes, or such as are under the immediate dominion of
others, all popular states have been obliged to establish certain qualifications; whereby some,
who are suspected to have no will of their own, are excluded from voting, in order to set other
individuals, whose wills may be supposed independent, more thoroughly upon a level with
each other.
See similarly Jefferson 1785, query 13. Cf. Goodin 1993.
The thought of nineteenth-century electoral reformers was that we can safely

enfranchise people who are dependent on the will of others, just so long as we
institute the secret ballot to prevent the superordinate party from learning
how the subordinate party voted.31 Some—notably Jeremy Bentham and John
Stuart Mill—opposed the secret ballot on other grounds, which we will address
in Section 18.4. But while they thought that on balance there was more to be said
against than for the secret ballot, both agreed that the principal thing to be
said in its defence was that it helped ensure the independence of electors.32
It bears reemphasizing here what we said in Section 5.1.1, which is that the
CJT’s Independence Assumption does not rule out all interaction among voters.
Specifically, there is no decisive reason on grounds of independence to rule
out pre-election discussions among voters. (Indeed, there are, as we have said
in Chapter 9, many reasons on other CJT-related criteria to encourage them.)
In the end, whether the votes are independent is a statistical question. To
make them independent voters should exercise their own independent judge-
ment on the evidence and arguments presented by others over the course of
those conversations.33
18.2.2 Public Funding of Elections and Public Broadcasting
Another pair of institutional mechanisms to help promote the independence of

people’s votes from untoward common causes is public funding of political
campaigns and public broadcasting services to provide voters with unbiased
information on which they can base their votes. Where political campaigns
are hyperexpensive and parties are reliant upon private sources of funds,
those who provide those funds have undue influence over the content of the
political messaging.34 Think of the Koch brothers in contemporary America,
for example.35 Likewise, where people are reliant upon one or a few private
sources for their news, the owners of those outlets have undue influence over
31 Rokkan 1961. Przeworski (2015) proves them right in practice, insofar as incumbents lose
much more often under secret than open voting.
32 Thus, Bentham (1791/1999, p. 148) conceded that secret voting should be used in assemblies
when ‘circumstances render a hidden influence suspected’. Mill (1861/1977, ch. 10) agrees:
It may, unquestionably, be the fact, that if we attempt, by publicity, to make the voter respon-
sible to the public for his vote, he will practically be responsible for it to some powerful
individual, whose interest is more opposed to the general interest of the community, than
that of the voter himself would be, if, by the shield of secrecy, he were released from respon-
sibility altogether. When this is the condition, in a high degree, of a large proportion of the
voters, the ballot may be the smaller evil.
Mill simply did not think that that was a major problem in British democracy of his day; the fact
of the matter is, however, much disputed among historians of the period (Nossiter 1975).
33 Goodin 2003, ch. 7. Hawthorne 2014. 34 Lindblom 1977, chs 14–15.
35 Skocpol and Hertel-Fernandez 2016.
the politically relevant information people receive.36 Think of Berlusconi or

Murdoch, for examples.
Of course, if there is only one broadcaster or one newspaper, and that is owned
and operated by the state, then the independence of voters can be compromised
by that as well. All too often, the state having a monopoly on news media is
associated with a totalitarian regime that uses its monopoly deliberately to feed
its citizens distorted information. But even if that is not the case, even if a state
monopoly reported fully and fairly the facts, the sheer fact of the monopoly
would very likely compromise the independence of votes from one another,
which all depend on the same source of facts.37 So for the sake of independence,
a plurality of news outlets is to be sought. And that is precisely what laws against
cross-media ownership are designed to ensure.38
18.2.3 Proliferate Independent Opinion Leaders
Independence, competence, and numbers are what drive the CJT. The more
competent, independent sources of judgement the better (all else being
equal, of course). But remember, the competence of the group as a whole is
a rapidly increasing function of the number of voters; so the difference
between hundreds and millions might not matter all that much to overall
group competence, in certain circumstances. That is what drove our earlier
arguments that overall group performance might be epistemically pretty
good, even if everyone followed opinion leaders slavishly (and all the more
so if they followed them only some of the time), just so long as there are a
multitude of opinion leaders who are both competent and independent of
one another (and in the case of only partially following it helps if voters are
of appreciably better than random competence when they are voting inde-
pendently of any opinion leader) (see Section 11.3).
Institutional mechanisms that attempt to ensure that a multiplicity of inde-
pendent points of view are expressed in the public forum would enhance the
overall epistemic performance of democracy for that reason. Media ownership
laws of the sort just alluded to would be one example. But the right of free
speech itself is another key element in that.39 So too are measures to ensure
robust multi-party democracy and, indeed, ‘associational’ democracy beyond
36 Soroka et al. 2013.

37 At least classical independence in Condorcet’s sense. In our Chapter 5 discussion we show
ways to deal with such situations by suitable conditionalization. Votes would just be ‘independent
conditional on the common source of information’, and the quality of the information provided is
what might be lacking in this case. It is unlikely that even the best broadcaster will never broadcast
confusing or misleading information.
38 Keane 1998. 39 Ladha 1992.
just political parties.40 Toleration of diversity is not just polite multicultural

policy but also a good epistemic investment.
18.3 MECHANISMS TO INCREASE COMPETENCE
18.3.1 Increasing Individual Competence
The CJT tells us that the epistemic performance of a democratic community at

large is a function, among other things, of the mean competence of its individual
voters. But as we observed in Section 6.2, as long as the size of the group is large
and mean individual competence within it is appreciably above random, there
is relatively little to be gained from further improving mean individual compe-
tence. Which, in a way, is just as well, since (as we also observed in Section 6.3)
there may well be little that can be done to improve mean individual competence
among voters. Civic education may pay off in the long term. Providing unbiased
information, through public broadcasting, might be useful at the margins. But
great gains are not to be expected.
The principal way in which relatively uninformed voters manage to cast rea-
sonably accurate votes is, as we argued in Chapter 12, by taking cues. One of the
main cues is ‘party label’. If a person knows which party ordinarily supports
options that she thinks are best, and she knows that this option has her preferred
party’s endorsement in this election, then she can with some confidence vote
whichever way ‘her’ party tells her.
Just how accurate voters can be, when cueing on party labels, is upper-bounded
by just how reliable party labels themselves are. One way to improve the per-
formance of voters who cue on party labels is therefore to make the party labels
more meaningful, in the sense that they correspond more closely and reliably
to some particular set of policy options. That would be one epistemic reason for
preferring parties to occupy distinctively different positions from one another,
rather than being ‘catch all’ parties that stand for everything and nothing.41
18.3.2 Increasing Collective Competence
There is also scope for increasing the collective competence of groups, without
increasing mean individual competence within them. The way to do that is by
increasing diversity, in the relevant respect, within the group (as discussed in
Chapter 7).
40 Dahl 1972; 1982. Hirst 1994; cf. Amin 1996. Cohen and Rogers 1995. Skocpol and Fiorina 1999.
Skocpol, Ganz, and Munson 2000.
41 Schattschneider 1950. Kirchheimer 1966.
It is an empirical question what sorts of institutional designs best do that.

But empirically it seems likely that that is better achieved by having an exten-
sive franchise, rather than limiting it to one narrow group of people that is
likely to be pretty homogeneous. And it seems likely that it is better achieved by
institutions of tolerance and multiculturalism, which are welcoming to people
with differing perspectives and interpretive frames.
18.4 MECHANISMS TO INCREASE SINCERIT Y
A final assumption required for the CJT to go through is that voters vote
sincerely, for the alternative that they truly believe to be the correct one, rather
than voting for some other alternative for an ulterior strategic motive. There are
various institutional mechanisms that can assist in promoting that outcome.
All of them involve introducing more openness, transparency, and accountability
into the voting system.
As we foreshadowed in our discussion of secret ballots in Section 18.2.1,
both Jeremy Bentham and John Stuart Mill favoured open voting as a way of
ensuring people vote sincerely for what they believe to be in the public interest
rather than in their own narrowly private interests.42 As we have already said,
that comes at a cost, in terms of creating scope for vote-buying and electoral
intimidation, and hence a violation of the CJT’s Independence Assumption.43
But insofar as open voting induces sincerity, it helps underwrite another of
the CJT’s key assumptions.
Deliberation, involving the open sharing of reasons for voting as one does,
logically should and empirically has been shown to reduce the incidence of
insincere, strategic voting (as discussed in Section 9.2.3). When you do not just
cast a vote, but also have to explain your vote to others, that reduces your free-
dom of manoeuvre. Your votes have to tally with your proffered reasons, and
there is simply less scope for your vote to misrepresent strategically what you
truly believe to be the best option.44 Institutions requiring deliberation and the
42 Bentham 1791/1999, pp. 135 ff. Mill 1861/1977, ch. 10. Tocqueville similarly writes (quoted in
Elster 2015, p. 10):
One should not be fooled if a political assembly preferred the secret regime by citing the need
to avoid the surveillance by the head of the State: it would only be a pretext. The real motive
for this behavior would rather be the desire to submit oneself to his influence without
exposing oneself too much to public blame.
43 It may come at a cost even in terms of promoting the public rather than narrow sectoral
interests: representatives who might have engaged in more cooperative public-interest-seeking
behaviour behind closed doors might, when casting a vote they know will be public, be more
inclined to vote instead for the sectoral interests of those they are supposed to be representing. See
Stasavage 2007; Naurin 2007.
44 See Vermeule (2015) on various systems of ‘open-secret voting’.
giving of reasons in a discursive setting (either prior or even subsequent to a vote)

serve this sincerity-enhancing function. That advantage is not unique to the
minipublics, which are the focus of today’s deliberative democrats. Instead, it is
an advantage that is shared by deliberative assemblies of much more traditional
sorts—smaller upper houses of the legislature, legislative committees, and so on.
Accountability-forcing mechanisms more generally can help serve to increase
the sincerity of people’s votes. Having to stand for re-election has that effect: the
incumbent has to explain how he voted and why; and saying that he voted one
way purely strategically in the hope it would have the opposite effect is likely
to prove electorally unpersuasive. Institutional rules that serve to promote
transparency—and thus maximize the number of votes that incumbents have
to account for—have the same sincerity-inducing effect. Submajority rules
are often used for just that purpose; to allow a minority of a legislative com-
mittee to force consideration of a bill by the full chamber or to force a roll call
rather than just an unrecorded voice vote on a proposed piece of legislation.45
45 Vermeule 2005; 2007, pp. 85–114.

Part V
Conclusions
Perhaps in earlier times factually correct answers to important policy questions

were easier to identify. Nowadays, the complexities created by the global nature
of politics—global markets, global environmental challenges, global security
challenges, and so on—seem to make it harder to come up with straightfor-
ward answers. We’re uncertain whether things ever used to be easy, or whether
this is just a false perception born of hindsight. But one thing is clear: when
the answers to political problems are not obvious, the epistemic demands on
democracies are high. We come to expect our system of government to perform,
epistemically speaking, where we as individuals struggle to find answers. In
this book we argue that democracies can do surprisingly well on that account,
and almost certainly better than the obvious alternatives.
We have investigated how and under which conditions democracies can
track ‘The Truth’. Our investigation revealed a surprising variety of democratic
decision-making practices and procedures, each with distinct epistemic effects.
Democracy, in reality, is a mix of arrangements and institutions, with different
epistemic benefits and burdens. It is partly deliberative and partly aggregative.
It can operate in centralized or distributed fashion, sometimes dividing epi-
stemic labour, and sometimes asking everyone. Democracy is sometimes
direct and sometimes indirect. It creates institutions and sometimes delegates
decisions to subgroups. It can take decisions simultaneously, or, less often,
sequentially. It often depends not only on who votes, but also on who sets the
agenda and how the best alternatives for the agenda are identified. And it relies
on a number of mechanisms for learning and information transmission,
including, but not limited to, respecting traditions, following cues, using heur-
istics, listening to leaders, and other forms of learning.
In concluding the book, we return in Chapter 19 to the question of whether
‘The Truth’ to be tracked exists and how to interpret it. The relation between
politics and The Truth is seen, by some, to be fraught in various respects. Some
say we should not seek truth through political means, either because there is no
truth to be found or else because there are some truths that, even if found, ought
not to be acted upon politically. Others question ‘by what right’ some people
impose their vision of The Truth on others—some among them expressing this
as a fear that pursuit of The Truth will justify tyranny and the stifling of free
expression.
Chapter 20 collects our headline findings about the Condorcet Jury Theorem,
its various extensions, and how it can be applied to the diverse institutional and
social practices of democracy. We end that chapter with some proposals as to
how the epistemic performance of democracy can be improved.
In the Epilogue, Chapter 21, we explore at some length the challenges posed
for an epistemic theory of democracy by the victories of Trump in the 2016
US Presidential Election and of Brexit in the 2016 referendum on the UK’s
membership in the European Union. Both constitute genuine challenges for an
epistemic theory of democracy, for in both cases campaigns based on palpable
falsehoods nonetheless electorally prevailed. But as we shall show, they did so
for reasons that are themselves explicable in terms of the epistemic theory of
democracy that we have here been developing. In both of those cases, some of
the key assumptions required for the Condorcet Jury Theorem to work were
absent—or, indeed, intentionally circumvented.
19
The Relation between Truth and Politics,

Once Again
19.1 THE LIMITATIONS OF TRUTH-SEEKING

IN POLITICS
There is a certain moral seriousness about playwright Harold Pinter’s Nobel

Lecture that we would commend to anyone tempted by outlandishly sceptical
positions as regards the role of ‘The Truth’ in politics. Pinter said:
In 1958, I wrote the following: ‘There are no hard distinctions between what is real
and what is unreal, nor between what is true and what is false. A thing is not neces-
sarily either true or false; it can be both true and false.’
I believe that these assertions still make sense and do still apply to the exploration
of reality through art. So as a writer I stand by them—but as a citizen I cannot. As
a citizen I must ask: What is true? What is false?1
We desist from comment on the truth status of artistic fictions, but we endorse
wholeheartedly Pinter’s proposition as regards politics.
The nature of the truth is of course a central topic in philosophy. Vast amounts
have been written about it, and many subtly different positions advanced over
the course of those debates. We have no hope of canvassing all of them—but
happily, neither do we have any need to do so for the purposes of our present
discussion. Here we will focus instead upon two positions that are of particular
relevance to the political realm.
First, social constructivists are right to say that some things are true only
because most people believe them to be true. Here is one such proposition: ‘The
US Constitution is the effective law of the land in the US.’ The truth of that
proposition depends on people in the US (or anyway US judges2) believing it to
be true. But to say that some things are true only because people (most people,
enough people, the right people) believe them to be true is not to say that that
1 Pinter 2005/2012, p. 9. 2 Hart 1961, p. 113.

is true of all things. Some things are true regardless of anyone’s beliefs. H2O
boils at 100 °C at sea level, whatever anyone’s belief about that matter.3 Anyone
acting in denial of that fact is on a hiding to nothing. Likewise anyone who
denies that there is overwhelming evidence for anthropogenic climate change,
or for the evolution of species. And even insofar as the social constructivist is
right about the constructed truths of some claims, the question of ‘what is that
constructed truth around here?’ can itself have a true or a false answer.
A second position, associated with political liberals in general and Rawls in
particular, urges ‘epistemic abstinence’.4 The claim here is that there are some
things the truth of which we ought not to insist upon, politically—even if they are
true. People’s ‘comprehensive doctrines’ concerning The Right and The Good are,
on Rawls’s view, classically among them. So too are propositions that affect no one
but the believer of that proposition; which is to say, arguments against paternalism
might well apply in the realm of belief as well as in the realm of action.
The deliverances of this god or that are also typically regarded as something
towards which we ought to adopt some such stand-back attitude. But the issue
there seems to be less one of epistemic abstinence than one of epistemic access
(if there is anything to access).5 If your god told you one thing and mine told me
another, and that’s all that either of us can say about the matter, then there is no
way of adjudicating which of those contradictory revelations (if either) is correct.
‘Agreeing to disagree’, however illogical a stance in other circumstances,6 might
actually make sense where contradictory private revelations constitute the only
reasons anyone has for believing one proposition rather than another.
Where there is other evidence to be taken into account apart from divine
revelation, however, then that can constitute the decisive tie-breaker. The
sheer fact that there are some people who believe, because of perceived divine
revelation, that water boils at something other than 100 °C at sea level (contrary
to all other evidence) provides no reason whatsoever for the rest of us in their
political community to suspend (even just for practical purposes connected
with politics and public policymaking) our otherwise well-founded (yea, correct)
conviction that H2O boils at 100 °C at sea level—whatever our deluded fellow
citizens’ god might tell them.7
The best epistemic grounds for epistemic abstinence, in our view, are not that
‘comprehensive doctrines’ are all embracing or deeply held. (From a purely
3 This used to be true by definition. More recently, the Celsius scale has been pegged to different,
more precisely defined physical events.
4 Rawls 1993. Raz 1990.
5 Habermas (2008, p. 129; cf. Cooke 2013) has a similar thought: ‘Religiously rooted existential
convictions, by dint of their if necessary rationally justified reference to the dogmatic authority of
an inviolable core of infallible revealed truths, evade the kind of unreserved discursive examin-
ation to which other ethical orientations and worldviews, i.e. secular “conceptions of the good”,
are exposed.’
6 Aumann 1976.
7 This is even more obvious when tautologies are involved. As Grotius (quoted in Arendt
1967/1977, p. 240) says, ‘even God cannot cause two times two not to make four’.
The Relation between Truth and Politics, Once Again 305
practical point of view, it may be politically prudent not to press the matter if
someone is deeply wedded to falsehoods that he would fight hard to defend; but
that is purely matter of pragmatism, not of principle.) The best epistemic grounds
for epistemic abstinence are, instead, that there are no independent standards
outside of those doctrines that we can access for assessing the truth of the claims.
The truths that we suppose that democratic institutions can effectively deliver
are what Bernard Williams terms ‘everyday truths’ (about what J. L. Austin
dubbed ‘moderate sized . . . dry goods’).8 They are not grand truths of the soul,
but rather plain brute facts of the world. They involve, at the very least, facts
about how things are in the world, how they are causally connected, and ‘what
we should do’ given a certain goal, in an instrumentally rational sense. These
truths may or may not extend to include the value statements to determine
what is Right or what is Good or what justice requires.9 But there are, it seems
to us, plenty of purely factual matters of major political consequence.
Those plain truths might of course still be subtle or hard to ascertain. The
CJT’s claim to applicability does not depend upon everyone knowing all the
truths in question, but rather upon many people being better than random at
identifying them and the views of those who are not better than random not
being systematically biased. That is to say, the wisdom of the multitude does not
depend on each and every member of the multitude being correct, so much as
it depends on there being an expected majority of correct votes among that
multitude, which in turn depends upon a certain level of competence and an
absence of (or anyway a limit in) the systematic bias in any errors.10
19.2 IS PURSUIT OF THE TRUTH

DANGEROUS IN POLITICS?
As Bernard Williams rightly recalls:

It is a familiar theme of contemporary criticism, one that has been inherited from
some members of the Frankfurt School, that the Enlightenment has generated
unprecedented systems of oppression, because of its belief in an externalized, object-
ive, truth about individuals and society. This represents the Enlightenment in terms
8 Williams 2002, p. 9 ff. Austin 1962, p. 8.

9 Recall our discussion of alternative views in relation to that question in Sections 1.6.3 and
4.1. Nervousness about ‘The Truth’ as regards the latter sort of propositions is what animates many
discussions of ‘truth and politics’ (Rawls 1993; Estlund 1993; 2008, ch. 2; Cohen 2009). Those seem
one-eyed discussions, failing to separate out facts from values. As we said in relation to ‘moral
separability’ in Section 4.1.3, even if you do not think that there are facts about values that can be
established or acted upon in politics, you can surely agree there are facts such as those about the
boiling point of water that can be and should be.
10 Although, following the large literature on ‘motivated reasoning’, it may well be that people’s
values systematically induce them to adopt certain beliefs (Stich 2014; Bullock et al. 2015).
of the tyranny of theory. . . . [T]here is . . . a question whether the Enlightenment’s

models of scientific understanding do lead to the denial of political freedom and,
if they do, by what social and intellectual routes.11
There is no denying that those who think themselves in possession of The
Truth sometimes do some terrible things. They do them unhesitatingly and
unashamedly, precisely because they think they are in possession of The Truth.
And that is one of the principal reasons for some people’s opposing talk of
The Truth in politics.12
People who are Bayesian rational should never assign probabilities of zero
or one to any proposition based on empirical evidence.13 In terms of Bayes’s
theorem, assigning a probability of zero or one would preclude any updating
in light of subsequent evidence—and that simply has to be wrong. Likewise,
even in its boldest and most classic form, the CJT says only that, as the number
of better-than-random voters increases, the probability of the majority being
correct approaches certainty. It approaches, but never reaches, a probability of
one. So nothing in the apparatus we have here been using justifies arrogant
claims of certainty, even with very large numbers of very competent voters. Add
to that the qualifications that we have entered in Chapter 5 about the probability
that the majority will be right being upper-bounded by the best available
evidence—which will inevitably be less than perfect—and our Best Responder
Corollary truly is a message of epistemic modesty.14
Hannah Arendt warns of ‘tyrannies of “truth” . . . which . . . politically speak-
ing are as tyrannical as other forms of despotism’.15 Cautioning against the
pursuit of The Truth in politics, she writes:
Truth carries within itself an element of coercion, and the frequently tyrannical
tendencies so deplorably obvious among professional truthtellers may be caused
less by a failing of character than by the train of habitually living under a kind of
compulsion.
. . .
The trouble is that factual truth . . . peremptorily claims to be acknowledged and
precludes debate, and debate constitutes the very essence of political life. The modes
of thought and communication that deal with truth, if seen from the political
perspective, are necessarily domineering; they don’t take into account other
people’s opinions, and taking these into account is the hallmark of all strictly
political thinking.16
11 Williams 2002, pp. 3–4. 12 Arendt 1967/1977.

13 Analytic truths, tautologies, and such like are a different matter of course.
14 Beerbohm (2012, pp. 155–8) commends ‘epistemic humility’ as a virtue. We commend it,
instead, as a deliverance of pure practical reason.
15 Arendt 1967/1977, p. 246.
16 Arendt 1967/1977, pp. 230–41. See similarly Nozick (1981, pp. 4–6) on ‘coercive philosophy’.
But that is not at all what we are here proposing to do. Our proposal, built on
Condorcet’s jury theorem and certain extensions of it, is to take everyone’s
opinion into account in coming to an overall assessment of what is true. We
further point to the epistemic advantages of discussion and deliberation among
those with diverse opinions.17 Our proposal is merely that (assuming the pre-
conditions of the CJT or related jury theorems are broadly satisfied) the overall
assessment we come to on the basis of all those inputs should then be taken
as the best indicator of truth that we have available, for the purposes of subse-
quent policymaking. This strategy depends crucially upon the evocation, rather
than the suppression, of people’s own private information and opinions.
19.3 ARE THERE MAT TERS THAT SHOULD

NOT BE PUT TO A VOTE?
There are many questions that should not be put to a vote. As a matter of pure
logic, there is at least one thing that cannot be decided by voting, which is who
in the very first instance gets to vote.18 There are several more policies that can
only succeed if there is an element of surprise (like sneak attacks and currency
devaluations), thus precluding general public discussion and hence open
voting.19 There are yet more decision situations requiring confidentiality for
one reason or another, to do with personal privacy or commercial interests
or other states’ sensitivities. Finally, there are many matters on which no vote
should be taken because doing so would infringe people’s protected rights or
liberties. We take all that as read.
Might there be some further matters that ought not to be decided by demo-
cratic ballot? Some argue that if there is no correct answer, then at least one
argument for majoritarianism—the epistemic one—is mute. There may be
other arguments for making social decisions by majority voting, but tracking the
truth would then not be one of them. This is to raise once again the question of
facts and values, first introduced in Section 1.6.3. Condorcet-style jury theorems
assure us that, so long as their basic assumptions are met, large groups of people
are as likely to be correct as anyone can be about matters of fact. But as regards
value judgements, there are (it is often said) no right or wrong answers.
We offered reasons in Section 4.1 for thinking that that distinction is over-
drawn. There can be facts of the matter about values—at least sometimes if not
always, and at least on some plausible views of metaethics. But what lessons
17 In Chapters 9 and 7 respectively.

18 Once an electorate has been constituted, that electorate can then decide whom else to
include. But that initial electorate cannot be constituted through a vote, prior to anyone being
empowered to vote (Whelan 1983; Goodin 2007a, p. 43).
19 Bok 1982.
might this book hold for those who believe that that most politically relevant
questions are questions of value that do not have true or false answers?
Here is one thing we can say to them: Most of our imagined interlocutors
would happily agree that any given issue on the political agenda involves a
combination of factual judgements and of value judgements, and that different
issues do so in varying proportions. Furthermore, they would also surely agree,
any given political discussion might focus more heavily on the factual aspects
of the issue or on more emotive aspects of the issue. So we would invite our
imagined interlocutors to interpret the lessons of this book like this: When the
political campaign surrounding some particular issue has been more firmly
focused on the facts of the matter, the outcome of the election can be taken to
have more epistemic value. However, when the campaign has largely proceeded
on the basis of emotive value claims with minimal references to facts, the outcome
of that election will have little epistemic value. Either type of campaign may be
democratically permissible, by some larger standards. All we care to claim
here is that, insofar as you want to extract more justified epistemic judgements
from the outcomes of democratic elections, political campaigns ought to focus
on the factual premises underlying people’s differences of opinions—and
perhaps, as an adjunct to that, the question ought to be framed in a way that
facilitates that focus.
19.4 WHO SHOULD DECIDE WHAT IS TRUE?
Some express the fear that questing for The Truth may lead in anti-democratic
directions. William Galston expresses that worry in these terms: ‘Truth is one
thing, legitimacy another. The people have a right to be wrong, a right they often
exercise.’20 We are ourselves more tempted by the view attributed (approvingly)
to Senator Daniel Patrick Moynihan by President Obama when recalling the
following anecdote:
Moynihan was in a heated argument with one his colleagues over an issue, and the
other senator, sensing he was on the losing side of the argument, blurted out: ‘Well,
you may disagree with me, Pat, but I’m entitled to my own opinion.’ To which
Moynihan frostily replied, ‘You are entitled to your own opinion, but you are not
entitled to your own facts.’21
In any case, our claims regarding ‘authority’ are quite minimal. We only maintain
that if a system of rule systematically ‘gets it wrong’, the legitimacy of its authority
20 Galston 2012, p. 142. We are less inclined than Galston to suppose they exercise it all that
often. See similarly Estlund (1993, pp 74–5; 2008, ch. 2) and the remark from Rawls’s doctoral
dissertation quoted in the former.
21 Obama 2006, p. 126.
will be undermined by its sheer incompetence. A system of government that

consistently fails to ‘deliver the goods’ provides less convincing (perhaps entirely
unconvincing) content-independent reasons to obey the government.22 Epistemic
competence may confer legitimacy of a sort (Scharpf ’s ‘output legitimacy’23),
and indeed authority of a sort (‘epistemic authority’). But we do not claim that
even perfect epistemic performance suffices for legitimate political authority,
understood as the right to give orders to others who have a duty to obey them.24
On that question we remain insistently agnostic. All we claim is that competence
deriving from good epistemic performance is a good-making feature of a system
of government. However, that is not the only (and may not be the principal)
thing that is required in order for it to have legitimate political authority.
The question posed by those who question truth-seeking as a goal of politics
might be phrased, ‘By what right do you get to do what’s right?’ On its face,
that question may seem absurd. But, disambiguating, the question that is really
being posed is: ‘By what right (authority) do you get to do what’s right (correct)?’
It would indeed be wrong to suppose that the second ‘right’ (correctness) in any
way automatically gives rise to the first (authority). There are plenty of cases
where someone who is correct in his assessment of what should be done
nonetheless lacks legitimate authority to do it. Critiques of paternalism are
rife with such examples. Within the protected sphere carved out by her rights,
each person has the right to be wrong, so long as she infringes no one else’s
rights or interests in the process.
But the ‘by what right?’ question can—and in the argument about ‘epistocracy’
typically does—arise in another form. The question there concerns not the ‘right
to act’ but rather the ‘right to judge’. The question is ‘by what right (authority)
do you get to decide what we collectively deem to be right (correct)?’ This is a
different form of authority: not an authority that leads to obedience in action,
but an authority that leads to an acceptance of the putative authority’s epistemic
judgements. In the epistemic case, the fact that someone (be it a group or an
individual) tends often to be correct is prima facie an authority-conferring
fact, because being a good truth tracker in a particular domain simply is what
normally confers epistemic authority in that domain.
If a group has epistemic authority—and it is plausible that it does have
such authority in democratic majorities, where the assumptions of the CJT are
satisfied—then there exists a content-independent reason for the minority to
accept the majority’s judgement as the judgement of the collective. Individual
voters are still free to hold their own private judgements, of course. But they
22 Which is Raz’s (1986) definition of authority.

23 Sometimes legitimacy depends on ‘getting it right’. Vermeule (2010, p. 49) recalls, for example,
that: ‘The claim of epistemic superiority is itself one of the main ways in which courts and legal
theorists attempt to legitimate the courts’ countermajoritarian role.’
24 Cf. Estlund 2008; Arneson 2016; Viehoff 2016.
have good epistemic reasons to agree that the position of the collective should
be as determined by majority vote.
Of course, if there happens to be a dual consensus—first on the fact that the
best judgement is what is wanted in the case at hand, and second on whose
judgement is the best (i.e. the majority’s, on the CJT)—then this epistemic
authority to judge on behalf of the collective will not be questioned in practice.
But even if there is consensus on whose judgement is the best, the question of
‘by which authority’ might still arise among those who do not think that being
the best judge automatically authorizes someone to decide on behalf of us all.
The broadly inclusivist conclusion towards which Condorcet-style jury the-
orem reasoning points us helps to avoid that problem arising. Assuming the
whole population is, at least on average, better than random at assessing which
is the correct option, there is no compelling epistemic reason to exclude anyone
from being among the assessors of the truth.25 And if no one is excluded, each
person is exercising authority over herself.
Perhaps she may still ask ‘by what right are those other people involved?’.
Rugged individualists may question the authority of the whole to make deci-
sions on their behalf, when their own personal judgement points in a different
direction. But if the conditions of the CJT are satisfied, and the numbers are
large, the chances that she will be right and the majority wrong are vanishingly
small; so the majority at least has epistemic authority on its side. Someone who
insists on her right to judge for herself nevertheless is insisting on a right to be
wrong; and if the judgement affects only herself, there may be good grounds for
conceding that claim. But if it is a decision that will unavoidably affect in a
serious way a great many people at the same time, arrogating unto herself any
claim to impose her quite-probably-wrong view on so many others who will
also be gravely affected is a claim not to be seriously entertained.26
There are of course, famous cases in which small upright minorities have
held firm against misguided or deluded majorities. It is perfectly possible that
the minority—or even a lone dissenter—is right after all.27 This is normally a
rare occurrence. But CJT reasoning should lead us to expect it to be a much
more likely scenario when others are either not competent or not independent
in their judgements. If a great many voters are either not competent or not
independent, then their epistemic authority should indeed be regarded as
diminished or even non-existent.
25 As pointed out in Section 15.4, collective competence could be increased marginally by
weighting more competent voters’ votes more heavily. But in any large electorate, the increase
would be so marginal, and the probability of the majority being correct would be so high already,
that it would be hardly worth the bother.
26 Weinstock 2016.
27 All the CJT says, after all, is that the probability the majority is correct under favourable
circumstances asymptotically approaches 1—it never quite reaches it.
When we should, and when we should not, assume that a majority has
epistemic authority is a difficult empirical question. That is an issue with which
we have grappled throughout the book, and we have encountered some scen-
arios in which a minority of dissenters has good reasons to doubt the epistemic
authority of the majority. But note well that the argument against granting
epistemic authority must be rooted in a violation of jury theorem conditions.
In a tolerably well-functioning democracy, the conditions of the CJT—or at
least the conditions of less demanding jury theorems we have introduced—
should normally be met. In flawed democracies, however, it is an open question
whether or to what extent that is the case. Consequently, the argument for the
epistemic authority of the majority is conditional on favourable epistemic
circumstances.
The upshot is this. For a democracy to succeed, it needs (among many other
things) to have at least some measure of epistemic authority. But to have epi-
stemic authority, it needs to have the features that make democracies succeed,
epistemically. Our Epilogue (Chapter 21) is devoted to a pair of dramatic cases
in which those conditions were not met.
20
Headline Findings, Central Implications
In concluding, let us recall the comment from Condorcet that serves as the
epigraph for this book:
In general, a law which has not been voted unanimously involves subjecting men
to an opinion which is not their own, or to a decision they believe contrary to their
interest. It follows that a very great probability of the truth of this decision is the only
reasonable and just grounds according to which one can demand such submission.1
We concur with the broad thrust of Condorcet’s claim. He surely exaggerates in
saying ‘the only’. That they have considerable epistemic virtues is not the only,
and maybe not even the principal, thing to be said in favour of democratic
procedures. Still, that is definitely a major good-making feature of democracy—
if only because getting our facts right helps us achieve the outcomes that we seek,
and in that way confer ‘output legitimacy’ on the political system as a whole.2
We thus endorse what we shall call a minimal competence requirement for good
political rule. Insofar as government decisions depend on facts, the government
must be minimally competent at establishing the facts. The competence to get
the facts right helps the government to select broadly correct means to chosen
ends—an ability that we insist is a necessary condition of legitimate rule.3 A form
of government that fails, time and again, to establish the facts correctly, will fail
in the provision of most basic services. It will not be responsive to what its
1 Condorcet 1785/1976, p. 44. Jason Brennan (2011c; 2016) sets out a similar ‘competence

principle’, although he misapplies it in extrapolating from ‘competent people’ (2011c, p. 700) to
the competence of the ‘deliberative body’ (2011c, p. 704) composed of them. The lesson of the
Condorcet Jury Theorem is that, assuming its conditions are met, the collective competence of
the group can be substantially higher than the competence of the individuals comprising it.
Elsewhere, Brennan (2011b, pp. 169–75) argues—for reasons we have queried in Chapter 6—that
one of those conditions is not met.
2 Scharpf 1999.
3 It may not be all that matters for legitimacy. A decision-making procedure may, for example,
also be more legitimate if it offers explanations and reasons for its decisions, and not merely an
ironclad guarantee that they are indeed correct (Estlund 2008, pp. 105–6). But notice that the
giving of reasons—and having them challenged and errors in them corrected—should be com-
mended on epistemic as well as on legitimation grounds (Fuerstein 2013).
Headline Findings, Central Implications 313
citizens want because it will not be responsive to how the world is, and its
policies will be misguided most of the time.
When Condorcet insists that the probability of correctness must be high, we
understand that as a claim about the class of decisions that are capable of having
correct or incorrect answers. That is to say, if there is a truth to be tracked then
a system of government ought to be good at tracking it. A citizen subjected to
the opinion of others is entitled to expect, insofar as facts are involved, that the
government tends to get these facts broadly right.4 This ‘minimal competence’
criterion must of course be applied to the overall performance of a government.
A government does not become illegitimate if it errs occasionally. Nor does it
become illegitimate if it fails to track truths that are very hard or impossible to
track, in the given situation. But it does lose legitimate authority if it errs most
of the time. Most obviously, if a government time and again produces decisions
that are epistemically worse than the decisions of any of its citizens chosen at
random, it cannot be legitimate in the long run.5
The larger claim of this book is that democratic institutions of broadly the
ordinary sort can help to satisfy the minimal competence requirement.6 They
can, just so long as presuppositions of the Condorcet Jury Theorem are tolerably
well satisfied and there is a truth of the matter to be tracked. Anyone who denies
that latter proposition will be unmoved by our argument. Still, to be completely
unmoved, someone would have to deny that premise fully. Doing that seems to
us highly implausible.7 No doubt there are some politically important proposals
that are not subject to any external standard of truth. And no doubt there are
some other proposals that, even if true, ought not to be politically imposed. But
it is wildly implausible that, between them, those two categories wholly exhaust
every proposal that is of political interest.
Even where there are truths to be found and properly to be pursued politic-
ally, democratic verdicts are not invariably correct. Aggregating the votes of
many may or may not help much if the crucial presuppositions of the CJT are
not met.8 In particular, if the available evidence is just not good enough, even
4 As far as is reasonably possible of course. Even the epistemically best government is not
omniscient.
5 Whereas our minimal competence requirement is a criterion of legitimate political authority,
for Jason Brennan violation of his ‘competence principle’ constitutes an ‘injustice’ (2011c, p. 704) or
a rights violation (2011c, p. 700). For objections to his formulations see Arneson (2016, pp. 169–73).
6 Sen (2009, p. 342–5) points to the ‘informational role of democracy’ as the principal explan-
ation for why ‘no major famine has ever occurred in a functioning democracy’.
7 Williams (2002, pp. 2–3) delightfully bursts the sceptics’ bubble as follows:
some ‘unmasking’ accounts of natural science . . . aim to show that its pretensions to deliver
the truth are unfounded, because of social forces that control its activities. . . . They . . . typically
depend on the remarkable assumption that the sociology of knowledge is in a better position
to deliver truth about science than science is to deliver truth about the world.
8 ‘May’, because (as we have shown in Chapters 3–5) those assumptions can be relaxed in
many ways.
the largest group might not be able to make sense of it. Similarly, if there are
other misleading factors such as a strong but incompetent opinion leader, the
majority of even a large number of voters has little chance of being correct.
Still, understanding the systematic sources of epistemic failure can help us to
safeguard against them.
If you are attracted by the epistemic promise of democracy, as we are, then
you should value the CJT for helping to show how to improve it. Our claim is
not that democracy is epistemically perfect. Our claim is merely that, unless the
presuppositions of the CJT (or its related renderings) are grossly violated, dem-
ocracy is likely to be a pretty reliable truth-tracker—better than any other sys-
tem of government, and capable of further improvement in ways that we
indicate in what follows.
20.1 HEADLINE FINDINGS
Before turning to its central implications for political practice, let us summarize
the principal findings of our analyses in just a few nutshells.
20.1.1 As Good As It Gets, As Bad As It (Probably) Gets
The most plausible way in which the competence of individual voters might on
average be worse than random is through the influence of one or more common
causes (opinion leaders, and such like) that systematically mislead them. Such
common causes violate the classic CJT’s Independence Assumption.9 But the CJT
conclusions are reasonably robust against moderate violations of Independence,
and the Best Responder Corollary can deal with interdependence via common
causes by conditionalizing on those common causes.10
Viewed through the lens of the Best Responder Corollary, the conventional
wisdom about the Condorcet Jury Theorem (indeed, Condorcet’s own conclu-
sions) are quite probably both too optimistic and too pessimistic.
• Pace Condorcet, and virtually everyone following him, the probability of
a large group of people’s majority verdict being correct does not normally
converge either to one or to zero (depending on whether individual voters
are on average of better- or worse-than-random competence).
• In the best-case scenario, it converges to the probability that the available
evidence and other aspects of the decision situation are truth-conducive.11
9 Section 4.5.1. 10 Sections 4.5 and 5.3.

11 More precisely, the probability that, given the state of the available evidence and other
aspects of the decision situation, its best responder would judge correctly. See Sections 5.3 and 5.4.
• In the most realistic worst-case scenario, it ordinarily converges to the

(im)probability that the opinion leader (or other common cause) that
guides voters is itself pointing in the right direction.12
Where the opinion leader (or other common cause) is systematically
misleading—less likely to be right than random—then the collective epistemic
performance of a group can be better if voters follow that lead less than perfectly.
That will be the case insofar as, when deviating from that lead, voters do so
randomly or better (i.e. deviating more often when the lead is in error than
when it is correct).13
20.1.2 All’s Well So Long As There Are Sufficient,

Numerous, Competent, Independent Influences
at Work Somewhere
The probability that the majority will be correct will converge to the upper
bound set by the probability that the available evidence and other aspects of the
decision situation will be truth-conducive, just so long as there are sufficient
numerous, competent, independent influences at some suitable place in the
decision-making process. This can happen in any of several ways.
• First, it is fine for all voters slavishly to follow opinion leaders (or other
common causes such as cues) just so long as there are sufficiently many
independent ones that are themselves competent and independent of one
another and not too unbalanced in their influence on voters.14
• Second, it is fine for all voters to follow even just one single opinion leader
(or other common cause), just so long as they do not follow it too often
and, when not following it, the voters are both individually competent and
independent of one another in how they cast their votes.15
• Third, it is fine for most voters slavishly to follow polarized opinion lead-
ers (or other common causes) that deliberately direct their followers in
ways diametrically opposed to one another (and hence are neither inde-
pendent of one another nor of better than random competence), just so
long as the opposing groups of ‘led’ voters are roughly equal in size and
there is a large enough group of voters who remain independent of any of
those influences, are individually competent, and who vote independ-
ently of one another.16
12 Section 5.4. 13 Chapter 11. 14 Sections 5.4.3 and 11.3.
15 Section 4.5.3. 16 Section 11.2.2.
20.1.3 There Are Ways of Coping with Incompetent Voters
Even if there are substantial numbers of incompetent voters in the electorate,

there are ways of coping epistemically with that.
• First, if incompetents (particularly, or exclusively) vote on the basis of a
set of cues that are epistemically pretty good guides, then the epistemic
performance of the group as a whole may be very similar to what it would
have been had its more competent members alone voted.17
• Second, insofar as democratic theorists are right in their speculations
about the educative effect of voting and experiencing the consequences of
one’s vote, the practice of voting might make people increasingly individu-
ally competent in how they cast their votes.18
20.1.4 The Case for Large Groups
If the voters are individually competent and independent in the minimal sense
required by the Best Responder Corollary, and if the evidence is truth-conducive
most of the time, then, ceteris paribus, a majority among a larger group of
people is more likely to be right than among a smaller group.
• In particular, large electorates with better-than-random average
competence19 are likely to outperform smaller groups of experts with
higher individual competence, and increasingly so as the electorate
grows larger.20
• In a large group of voters with average individual competence better than
random,21 weighting the votes of more competent voters more heavily
will make little difference to the group’s overall epistemic performance,
decreasingly so as the size of the group increases.22
On its face, that would seem to constitute an epistemic argument for direct
democracy among a mass electorate. But representative democracy can be
almost as good epistemically, just so long as even a moderately small fraction of
the representatives see themselves as ‘delegates’ and follow the majority of their
constituents in their own vote.23
17 Section 12.5.5. 18 Section 15.6.

19 Competence here understood in terms of competently tracking the best responder to the
available evidence and other aspects of the decision situation.
20 Section 15.1. 21 And individual competence pci > 0.5 for all voters i.
22 Section 15.4. 23 Section 16.2.3.
20.1.5 Smaller Groups to Deliberate and Winnow the Options
Furthermore, small groups can sometimes have an epistemic edge over very
large ones. This is for four reasons:
• First, they have the chance to deliberate in a way that is unlikely in a larger
group. Deliberation can help to increase individual competence; it can
also reduce the influence of misleading evidence biases, opinion leaders,
and so on; and it can lead to the discovery of new and better items to put
on the decisional agenda.24
• Second, if there is a risk that individual competence might decline with
an increase in the number of options being considered, there can be an
epistemic advantage in different smaller groups being assigned responsi-
bility for making recommendations over a smaller set of options which the
larger group can then consider.25
• Third, sometimes a small group of more competent voters may be able
to process the evidence better or avoid the traps of misleading com-
mon causes more effectively than a large group can. However, this
positive effect must be weighed against the epistemic costs of risking
the introduction of other biases when selecting a small group to make
the decision.26
• Fourth, breaking up into smaller search parties can be helpful in the
explorative stage of the decision process, when the relevant alternatives to
be put on the agenda need to be found.27
Such arguments can be used to support representative rather than direct
democracy, with the choices of the government of the day then being
accepted or rejected by the mass electorate at the next election.28 They also
support the delegation of legislative work to committees, whose recom-
mendations are then judged by the legislature as a whole.29 Division of
labour can be very useful in search processes on ‘rugged landscapes’ where
starting the search in many different places increases the chance to find the
best alternatives.30
Does this mean that, given the edge of smaller groups, we should generally
delegate decision-making to them altogether? Not at all. Competent experts in
smaller groups should advise, especially when deliberation in smaller groups
can be helpful to reduce biases by informing the larger group and by narrowing
down the agenda to a manageable set of relevant alternatives. But it remains the
24 Section 9.2. 25 Sections 8.2.3 and 8.2.4.

26 Section 5.5. 27 Section 8.1.2. 28 Section 16.1.
29 Sections 8.2 and 18.1. 30 Section 8.1.
case that larger electorates should usually decide, given the epistemic advan-
tage deriving from their superior size.31
In rare instances, we might face a decision situation in which the electorate
is likely to be misled but a select smaller group would not be. Imagine, for
example, a whole nation under the spell of a mad but persuasive dictator or a
whole population enamored with a convincing but false pseudo-scientific
theory. In such cases it is at least theoretically possible that (i) a smaller group
of less biased voters might do better, epistemically, and (ii) that they have to
make the decision without the electorate if there is no realistic hope of breaking
the spell of the biases that disqualify the electorate. But note how specific the
assumption for such a setting would have to be: there needs to be convincing
evidence that the smaller group is less biased, and that there is no way for the
smaller group to de-bias the larger group.
There may be cases like that, but those really are exceptional circumstances.32
Under more normal circumstances, the place for smaller groups should be not to
decide, but to clarify the alternatives, to prepare a sensible agenda for decision,
to inform, and to remove any misperceptions of the existing evidence.
20.1.6 The Decision Situation Is Crucial
Even if all the other conditions of the CJT are met perfectly, there is effectively
an upper bound—lower than the probability of 1 standardly given by the CJT—
to the probability that a majority vote will be correct. That is set by the decision
situation, and just how truth-conducive it is. If the evidence is systematically
misleading, then people (even a majority among a very large group of people)
will be systematically misled.33 If the correct answer is not on the group’s menu
for choice, it cannot be chosen even by an otherwise highly reliable majority
among a very large group of people.34
Standard CJT-inspired discussions concentrate on ways of improving
epistemic performance by working on dimensions internal to the CJT, making
individuals more competent or their votes more independent for example. That
helps somewhat, but in most cases not all that much. It is important to go
outside the strict framework of the CJT to find ways to improve the decision
situation itself—because that sets the effective upper bound to just how
epistemically good the outcome can be.
31 Section 8.2.4.

32 ‘States of exception’, to use a phrase coined by Schmitt (1922/2005; Agamben 2005) in this
very different context.
33 Section 5.4.2. 34 Section 5.5; see also Sections 4.6, 5.3, and 9.2.5.
20.2 CENTRAL IMPLICATIONS

FOR POLITICAL PRACTICE
Now let us highlight some of the central implications of our analysis for
political practice.
20.2.1 Avoid Epistemic Deference
On the face of it, it might seem that the epistemically responsible thing to do is
to defer to the judgement of someone else, if you think that person is more
likely to be right than you are about the matter at hand. That is why we take
our doctor or lawyer’s advice, rather than keeping our own counsel on those
matters, after all. In such settings, deferring to the wisdom of others makes
perfectly good sense.
In collective-choice contexts, however, deferring in that way can be epistem
ically the wrong thing to do. Imagine a set of 101 individuals, all of whom are
independent of one another. Suppose that each of them correctly assesses the
likelihood that he will be right on the question under discussion as pc = 0.55;
and each of them correctly assesses the likelihood that the one expert available
to them will be right on that question as pc = 0.70. If each of them defers to the
expert, adopting the expert’s conclusion as his or her own, then the probability
the majority among them will be right is just the probability that the expert is
right, 0.70. But if each of them disregards the expert and votes on the basis of
his or her own assessment, the probability that the majority among them will
be right is 0.84.
20.2.2 Pluralism Is Good
Multiple independent sources of competent judgement maximize epistemic

performance. That need not involve each person being independent of each
other in the way contemplated in the classic CJT, however. It may just involve
instead lots of common causes (cues, parties, etc.) that are independent of one
another. Taking the epistemic function of democracy seriously should lead us
to create safeguards against overpowering opinion leaders and biases. The epi
stemic performance of a democracy is compromised if all information passes
through tight bottlenecks or is controlled by a small number of actors. Similarly,
a democracy is unlikely to perform well, epistemically, if everybody blindly
follows someone else.35 Citizens should be epistemically empowered to develop
35 Although it performs well if each follows some different other person, and if each does so
less than blindly; see Sections 11.3 and 11.4.
their own views, or, to paraphrase Kant, to have the courage to use their own
understanding.36 This has implications from school curricula to media regula-
tion, from promoting places for citizen deliberation to limiting the influence of
money on politics.
20.2.3 More High-Quality Evidence Is Good
The quality of the evidence upper-bounds what anyone can do epistemically.

Getting more high-quality evidence raises that upper bound. Notice, how-
ever, that this is not a matter of relying more heavily on experts: they’re
limited by the available evidence, just like anyone else. Rather, what we need
is improvement in the evidence itself, while reducing the influence of mis-
leading biases. The important point is that the citizens obtain evidence that
is truth-conducive. There is no (or not much) harm done if citizens rely on
cues, but it matters that citizens are informed enough to take into account
many cues. Formal education has an important role to play here, of course.
However, apart from education in the form of schools, colleges, universities,
and so on, education can also be improved by lifting the standard of public
reasoning and by widening participation.
20.2.4 Small-scale Deliberative Conclaves to Advise

the Electorate Are Good
Small-scale deliberative conclaves can play an important epistemic role

in advising the mass electorate. These might be conclaves of experts to
help the mass electorate weed out many of the possible but implausible or
confusing alternatives for decision, and help them focus on the choice
among those more likely to be correct. They may be conclaves of political
leaders (constituting a political party, for example), to help voters in similar
ways by highlighting some items deserving of particular attention on the
voters’ agenda.37 Alternatively, they may be conclaves of ordinary folk,
designed to show others what they would think had they devoted as much
time to the matter38 or to help them overcome ‘motivated reasoning’ by
having to explain and justify their reasoning to others who start from dif-
fering motivational premises.
36 Kant 1784. 37 White and Ypi 2011.

38 This is the role Fishkin (1995, pp. 163, 173) sees for mini-public deliberations.
20.3 GET TING IT RIGHT MAT TERS
Many political commentators, and some political scientists, have become

remarkably pessimistic about the future of democracy. These concerns are
often couched in epistemic terms. We are told that voters are (perfectly under-
standably, indeed perfectly rationally) uninformed; that the electorate bends
whichever way the wind blows; that there are no facts of the matter, anyway, on
which to decide; that even if there are facts, they are drowned out by emotions;
that the electorate is manipulated by the media, big money and populists; and
so on, and so forth.39 There is, as always, a grain of truth in these complaints.
But note well that democracy, like any other form of government, must be
measured against its alternatives. Do we seriously believe that a dictatorship or
oligarchy will fare better, epistemically? Admittedly, we sometimes welcome
governments of experts40 (Mario Monti’s two-year stint as prime minister of
Italy comes to mind)—but we generally only do so against the backdrop of
dysfunctional governance structures, when the ‘old’ system is in deadlock, cap-
tured by special interest, corrupt, or perhaps all these at the same time. From an
epistemic perspective, there is a high risk that the technocrats will all be influ-
enced by the same limited number of common causes—expertise and group-
think can often go hand in hand.41
Democracy can also be dysfunctional, of course. But even if a somewhat
dysfunctional democracy may sometimes be beaten epistemically, a major
argument for democracy remains: there is no other form of government that
has as much epistemic potential as democracies do. The combination of deliber-
ation and information pooling, the division of epistemic labour and the use of
multiple sources of information—if exercised properly by minimally compe-
tent and independent citizens who then vote together—cannot be beaten epi-
stemically. Democracy, for all its perceived or real flaws, comes with the greatest
epistemic expectations.
Therefore, even the diehard pessimists will have to agree with us at least on
this: If we can make democracy better, then it will meet the minimal compe-
tence requirement for democratic legitimacy, and it will do so far more easily
than any of its competitors.
39 Posner’s (2003, p. 107) remarks are typical:

With half the population having an IQ below 100 . . . , with issues confronting modern
governments highly complex, with ordinary people having as little interest in complex policy
issues as they have aptitude for them, and with the officials whom the people elect buffeted
by interest groups and the pressures of competitive elections, it would be unrealistic to
expect good ideas and sensible policies to emerge from the intellectual disorder that is
democratic politics by a process aptly termed deliberative.
40 In Dahl’s (1967, p. 21) terms, an ‘aristocracy of experts’. 41 Tetlock 2005.
21
Epilogue
What about Trump and Brexit?
Our analysis in this book has been a conditional one. Assuming certain
conditions (about competence, independence, and sincerity) are satisfied, the
pooling of votes by majority rule has epistemically beneficial properties. The
conditions can be weakened and many of those epistemic benefits still follow.
However, at some point, when the conditions are violated too severely, the epi-
stemic benefits of majority voting break down. Our argument can thus also be
taken as an analysis of why democracy does not produce epistemically super-
ior outcomes, when it does not, in epistemically radically non-ideal circum-
stances. That is the subject of this epilogue.
The day after Trump was elected, our colleagues inevitably poked fun at us
for our work asserting ‘the wisdom of crowds’. Joking aside, the election of
Donald Trump, and the vote for ‘Brexit’ (Britain’s exit from the European
Union), raise inevitable questions that cannot be ignored about the epistemic
function of democracy.
We are confident that the analysis offered in this book has internal validity. The
mathematics are as they are. Given the assumptions as specified, the conclusions
are as we report. Whether those assumptions and the implications drawn from
them correspond to the real world is, however, something else again.
Surely they correspond to something in the real world. That is to say, surely
there are cases (quite a lot of them, actually) in which there is some fact of the
matter and people engage in good-faith efforts to pool their information with
one another’s to find out what those facts actually are. But there are also cases
where influential actors deliberately, and successfully, mislead others. Take, for
example, the concerted campaign by the tobacco industry to discredit mount-
ing evidence that tobacco smoking causes cancer.1 In the 1950s tobacco com-
panies came together to establish
1 The best analogue in our own day might be organized climate change denial.
Epilogue: What about Trump and Brexit? 323
the Tobacco Industry Research Committee, a sham organization designed to

spread corporate propaganda to mislead the media, policymakers and the public
at large. Their goal was not to convince the majority of Americans that cigarettes
did not cause cancer. Instead, they sought to muddy the waters and create a second
truth. One truth would emanate from the bulk of the scientific community; the
other, from a cadre of people primarily in the employment of the tobacco industry.
The ruse continued for almost five decades, until lawsuits against the industry
forced the closure of the ‘research institute’ and the public release of its internal
documents. Now anyone with an Internet connection can read the full details of
the tobacco industry’s expensive efforts to create an alternate set of facts about its
products.2
The extent of such behaviour in politics has traditionally lain somewhere in
between the two polar cases of honest information-pooling and utter deception.3
The ‘big lie’ is a long-established technique (but one historically not all that oft
used) for securing political power.4 And the frequency, if not necessarily
effectiveness, of negative (and often not altogether truthful) advertisements
2 Rabin-Havt 2016; see similarly Harford 2017. It is perhaps no coincidence that it was the busi-
ness world out of which Donald Trump emerged (Barstow 2016). Indeed, his confidant Roger
Ailes, sometime CEO of Fox News, served as a secret operative for the tobacco industry resisting
the Clinton healthcare reforms in 1993 (Dickinson 2011). But in terms of the influences on Trump,
the impact of his mentor Ray Cohn—sometime chief counsel of Senator McCarthy’s communist-
witch-hunting committee—must not be underestimated (O’Harrow and Boburg 2016).
3 Arendt (1967/1977, p. 227) cynically says, ‘no one has ever doubted that truth and politics are
on rather bad terms with each other’. Still, the last time before 2016 that a US presidential election
could have plausibly been said to have been won on the basis of a literal falsehood was in 1960,
when the foreign policy centerpiece of Kennedy’s campaign was the claim that the previous
Republican administration had allowed a ‘missile gap’ to arise between the USSR and the US. That
was subsequently shown to be untrue. At the time, however, it was reasonably thought to be true
by Kennedy and the wider public, on the basis of what were still the official US estimates of Soviet
military capacity. Only those privy to top-secret briefings based on U-2 reconnaissance knew
otherwise, and they could not reveal that information without giving away that still super-secret
technology (Atlantic 2013; CIA 2013). Mind you, wars have subsequently been started on the
basis of falsehoods (Vietnam, on the basis of false reports of attacks on US warships in the Gulf of
Tonkin; Iraq, on the basis of false reports about Sadam Hussein’s weapons of mass destruction);
and elections elsewhere have been won on the basis of barefaced lies (in 2001 in Australia, on the
basis of lies about refugees throwing their children overboard to force the navy to pick them up
and take them to Australia (Australia Senate Select Committee 2004)).
4 The technique was enunciated most famously by Orwell (1949) in Nineteen Eighty-Four—
which rose to the top of Amazon’s bestseller list the week after Trump’s inauguration, when his
former campaign manager started talking about ‘alternative facts’ (Charles 2017). But it was known
to Abraham Lincoln, who in 1854 warned of the demagoguery of his nemesis Stephen Douglas in
these terms (quoted in Blumenthal 2016):
It was a great trick among some public speakers to hurl a naked absurdity at his audience,
with such confidence that they should be puzzled to know if the speaker didn’t see some
point of great magnitude in it which entirely escaped their observation. A neatly varnished
sophism would be readily penetrated, but a great, rough non sequitur was sometimes twice
as dangerous as a well polished fallacy.
Or as Press Secretary Larry Speakes said a propos President Reagan’s apocryphal stories, ‘If you tell
the same story five times, it’s true’ (quoted in Marcus 2016).
designed to undercut the credibility of one’s opponents has increased over the
past decades.5 But at least until recently, politicians conspicuously caught in
lies tended to be punished by voters, and in consequence politicians were
historically very wary of lying.6 While voters may have long been prone to a
certain amount of bias, selective perception, and motivated reasoning, they
have traditionally been broadly concerned with promoting the truth at least
as they see it.
In terms of our interest in the epistemic performance of democracy, we
must distinguish two different questions. First is the question about the
correctness of recent decisions of the electorate. Second is the question about
the truth-conduciveness of the campaigns and processes that led up to those
decisions. The first question is the one that triggered the snarky comments of
our colleagues. Don’t we agree that Brexit and Trump were surely the wrong
choices? As citizens we certainly do. But as social scientists we have to tread
more cautiously.7
Our concern in this epilogue is, therefore, with the second question alone. Our
concern is not that voters made what we ourselves think were catastrophically
incorrect decisions in supporting Brexit, and Donald Trump for US president.
We do—but that is beside the present point. Our point here is that in these two
prominent cases voters continued to lend their support to those campaigns,
even after they were clearly shown to be based on blatant falsehoods. Large
numbers of voters seemed to be impervious to the truth of the central claims of
those campaigns. Small wonder that the Oxford Dictionaries named ‘post-truth’
the ‘word of the year’ for 2016.8
Such apparent indifference of voters towards the truth should be highly
worrying for an epistemic theory of democracy. Certainly, anyway, it should be
5 Lau et al. 1999.
6 Nyhan and Reifler 2015. As James Fallows describes the old rules, ‘public figures would at
least try to tell the truth most of the time and they would recognize it as a significant penalty if
they’re shown not telling the truth’ (quoted in Rehm 2016b). Traditionally, ‘knowledge of the
risks of being caught has encouraged most politicians to minimize provable lies’ (Fallows 2016a,
emphasis in original). Traditionally, straight-out lying has not been a particularly successful strat-
egy of political manipulation in the long term (Goodin 1980, ch. 2). Of course, in Keynes’s famous
riposte, ‘in the long run we’re all dead’, and it may well be that the lies will not be caught in time to
do any good. That was Jonathan Swift’s (1710) worry:
[I]t often happens that if a lie be believed only for an hour, it hath done its work, and there
is no further occasion for it. Falsehood flies, and truth comes limping after it, so that when
men come to be undeceived, it is too late; the jest is over, and the tale hath had its effect:
like a man who hath thought of a good repartee when the discourse is changed, or the
company parted; or like a physician, who hath found out an infallible medicine, after
the patient is dead.
7 After all, we might be wrong in our own assessments—as we think are others (Caplan 2007;
Somin 2013; Brennan 2016) who bemoan ‘voter ignorance’ based purely on the fact that voters
disagree with neoliberal economists, who we too think to be often in error. See similarly Killick
(2017).
8 Flood 2016. Wang 2016. Oxford Dictionaries 2016.
if that were a pattern that is likely to persist, rather than being an aberration
limited to those two campaigns alone.
The purpose of this epilogue is to offer our best guesses as to what, exactly,
was going on with the votes for Brexit and Trump. In that way, we hope to
provide at least a set of reflections (our evidence is of course no better than
anyone else’s, so reflections are the most they can claim to be) as to the implica-
tions of the events of 2016 for the general applicability of an epistemic theory of
democracy.
21.1 THE POLITICAL LIES OF 2016
We begin by substantiating our claim that the Brexit and Trump campaigns
were based on lies that voters could and should have known to be false on the
basis of evidence that was readily available at the time they voted.9 Whether
voters actually knew what they could and should have known, or whether they
actually believed the falsehoods to be true, is an issue to which we will we return
in Sections 21.5 and 21.7.
21.1.1 Brexit Lies
In the UK referendum on the EU, the ‘Leave’ campaign made many tenden-
tious claims.10 But the one upon which we will focus here is the slogan embla-
zoned in huge letters on the side of the Vote Leave Battle Bus in which leaders
9 A lie is, by definition, a falsehood that the speaker utters with the intent that the hearer
believe it, the speaker knowing it to be untrue. Of course no outsider can really be sure what
someone else knows, believes, or intends. Some journalists hesitate to call Trump’s falsehoods
‘lies’ for that reason (Baker 2017; Baker in NBC Meet the Press 2017). Others, after the fashion of
a jury in a criminal trial, judge the weight of evidence to be such that beliefs and intentions can be
ascribed and lies attributed accordingly (Fallows 2016a; Dean Baquet 2016). Ironically, that is
precisely the approach the alt-right website Breitbart suggested in relation to ‘alternative facts’
(discussed in Section 21.9): Breitbart insisted that that is ‘a harmless, and accurate, term in a legal
setting, where each side of a dispute will lay out its own version of the facts for the court to decide’
(Gabbatt 2017; see similarly Hughes in Stelter 2016). It is just worth mentioning however that any
attorney who literally fabricated evidence—which is what the ‘alternative facts’ in question
amount to—would of course be in contempt of court (Goodin 2010).
10 Many of the Leave campaign’s claims about immigration were highly dubious—particularly
Nigel Farage’s poster picturing long lines of would-be immigrants who were actually nowhere
near Britain’s borders. So too was the claim that Turkey was about to join the UK. And many of the
‘sovereignty-undermining’ court cases of which Leave campaigners complained involved deci-
sions of the European Court of Human Rights, which is not part of the EU anyway (Grice 2017).
Advocates of Leave claimed that the Remain campaign was built on Project Fear, which they
claimed were lies about the economic consequences of Brexit. Evidence on that so far is mixed: in
the first year since the referendum the stock market performed strongly (Financial Times 2017),
but Sterling dropped precipitously to a thirty-one-year low against the US dollar in the immediate
aftermath of the referendum and has only very partially recovered (Allen et al. 2016).
of that campaign (including Boris Johnson and Michael Gove) travelled up and
down the country. The slogan read: ‘We send the EU £350 million a week. Let’s
fund our NHS instead. Vote Leave.’11
That claim was literally a lie in one respect, and it was seriously misleading in
another. That was the official finding of UK Statistics Authority, ‘an independ-
ent body operating at arm’s length from government as a non-ministerial
department directly accountable to Parliament’. And that finding was released
fully two months before the referendum vote and was widely reported during
the rest of the campaign.12
The Leave campaign’s claim was literally a lie in the sense that, while it is
true that
in 2014 the UK’s official gross payments to the EU amounted to £19.1 billion [the
basis for the ‘£350 million a week’ claim], this amount of money was never actually
transferred to the EU. Before the UK government transfers any money to the EU a
rebate is applied. In 2014 the UK received a rebate of £4.4 billion. This means
£14 billion was [all that was] transferred from the UK government to the EU in
official payments.13
That brings the amount actually transferred down, from £350m per week to
just under £270m per week.
Even that sum is seriously misleading, however, because ‘£4.8 billion came
back to the public sector in 2014. . . . Given these figures, . . . the UK government’s
net contribution to the EU— . . . the difference between the money it paid to the
EU and the money it received—was £9.9 billion in 2014.’14 That brings the
actual net transfer down to just over £190m per week—just over half the £350m
per week emblazoned on the side of the Vote Leave Battle Bus and featured
prominently on its website and in its leaders’ stump speeches throughout the
referendum campaign.
After continued harassment from Vote Leave correspondents, the Chair of
the UK Statistics Authority, Sir Andrew Dilnot, issued a further statement say-
ing, ‘The continued use of a gross figure in contexts that imply it is a net figure
is misleading and undermines trust in official statistics.’15 And even the chief
funder of the Leave.UK campaign, Arron Banks, agreed in response that ‘it’s
not smart to lie’.16 Leaders of Vote Leave nonetheless persisted in these gross
11 Of that claim, one Financial Times writer remarked, ‘It is hard to think of a previous example
in modern western politics of a campaign leading with a transparent untruth, maintaining it
when refuted by independent experts, and going on to triumph anyway’ (Harford 2017).
12 Dilnot 2016a, b; UK Statistics Authority 2016. BBC 2016; Islam 2016a; Ship 2016.
13 UK Statistics Authority 2016, emphasis added.
14 UK Statistics Authority 2016, emphasis added.
15 In what was, in the coded language of the British bureaucracy, a particularly stern rebuke to
the Vote Leave harassers, Dilnot (2016b) upgraded this from ‘potentially misleading’ in his earlier
announcement (Dilnot 2016a).
16 Quoted in Islam 2016a.
misrepresentations all the way to the end of the campaign17—only to repudiate

them promptly thereafter.18
21.1.2 Trump Lies
In the case of the Brexit Leave campaign, we have focused on one central lie. In
the case of the Trump campaign, there is a plethora to choose among. Table 21.1
contains a pot pourri of some of Trump’s more outlandish whoppers during the
campaign, which he typically continued repeating even after they had been
revealed as such.19
Beyond the particulars, the sheer frequency and brazenness of Trump’s lying is
utterly astonishing. Here is the Washington Post’s 2016 end-of-year assessment:
There has never been a serial exaggerator in recent American politics like the
president-elect. He not only consistently makes false claims but also repeats them,
even though they have been proven wrong. He always insists he is right, no matter
how little evidence he has for his claim or how easily his statement is debunked.
During the campaign, Trump earned 59 Four-Pinocchio ratings [‘whoppers’—the
highest rating], compared with 7 for Hillary Clinton.20
17 Gove, quoted in Islam 2016a. Boris Johnson insisted, ‘We think it’s relevant to keep people
focused on the global figure, because that is the figure over which we have no control’ (quoted in
ITV 2016). Throughout the rest of the campaign that claim remained on the website of Vote Leave
(Griffin 2016), the organization officially recognized by the UK Electoral Commission (2016a) as
‘represent[ing] those campaigning for that outcome to the greatest extent’.
18 That claim, along with everything else, was wiped from the Vote Leave website within days
of the referendum (Griffin 2016). Gove, in launching his abortive bid for the Conservative Party
leadership immediately after the election, reduced the sum promised to the NHS to £100m per
week: ‘Gove insisted he was not retreating from the slogan that implied all £350m would go to the
NHS, but said that was likely to be the impression given’ (Asthana and Mason 2016). Others—
such as UKIP leader Nigel Farage immediately, and David Davis after he became Minister for
Brexit—asserted they themselves had never made any such claim (Stone 2016b; Sparrow 2016a).
19 One is reminded of Jonathan Swift’s (1710) description of an English politician of his
generation:
his genius consists in nothing else but an inexhaustible fund of political lies, which he plentifully
distributes every minute he speaks, and . . . forgets, and consequently contradicts, the next half
hour. He never yet considered whether any proposition were true or false, but whether it were
convenient for the present minute or company to affirm or deny it. . . . I think he cannot with any
justice be taxed with perjury . . . because he hath often fairly given public notice to the world that
he believes in neither.
20 Kessler 2016a; for a summary of each see Kessler et al. 2016. George W. Bush, in contrast
(Fritz, Keefer, and Nyhan 2004, p. 4),
subtly and systematically attempted to deceive the nation about most of his major policy pro-
posals . . . while generally avoiding obviously false statements. Instead, Bush consistently uses
well-designed phrases and strategically crafted arguments to distract, deceive and mislead.
The result is that all but the most careful listeners end up believing something completely
untrue, while proving the President has lied is usually impossible.
Table 21.1 Fact-checking Trump’s lies

Trump claim Fact Check21
‘Our real unemployment rate is Actually, it is 5.3%. ‘Yes, . . . there are . . . [42%] “not in the
42 percent’ (18 August 2015). work force,” but the vast majority of those people do not
want to work. Most are retired or simply are not interested
in working, such as stay-at-home parents.’22
‘We can save as much as $300 Actually, ‘total spending in Medicare Part D [prescription
billion a year’ on prescription drugs] in 2014 was only $78 billion’.
drugs purchased by Medicare
(18 February 2016).
‘On November 1 . . . new numbers ‘This is a classic Trump claim. He cherry-picks the
are coming out which will show most extreme examples, applies them to the general
40, 50, 60 percent increases’ in population. . . . He says rates will increase by 40, 50, 60
premiums for health insurance percent—but the most common plans in the marketplace
under the Affordable Care Act will see an average increase of 9 percent. The vast majority
(26 September 2016). of marketplace enrollees receive government premium
subsidies and will be protected from premium increases.’
‘There are scores of recent migrants ‘The claim may be a . . . reference to a list from the office
inside our borders charged with of Sen. Jeff Sessions (R-Ala.) of 30 foreign-born
terrorism’, and ‘dozens and dozens individuals who were arrested on charges relating to
more’ per each case known terrorism in recent years. . . . The majority of the 30 cases
publicly’ (2 May 2016). involved naturalized US citizens—people who came to
the US as children or had arrived before 2011. We
reviewed similar lists of cases from 2014 and 2015,
involving 76 people charged with activities relating to
foreign terrorist organizations. Of them, 57 were US
citizens, [either] naturalized [or] natural-born . . . , and
many of the naturalized citizens had arrived in the
country as children.’
‘No, you’re wrong’ that stop-and- ‘In 2013, US District Judge Shira A. Scheindlin, in the
frisk was ruled unconstitutional Southern District of New York . . . , issued a 195-page
(28 September 2016). ruling . . . [holding] the city liable for violations of the
plaintiffs’ rights under the Fourth and 14th
amendments. . . . The federal appeals panel denied the
city’s request to overturn Scheindlin’s ruling.’
‘There were people over in New Actually, ‘Jerry Speziale, the police commissioner of
Jersey, a heavy Arab population, Paterson, which has the second-largest Muslim
that were cheering as the population in the United States, [said], “That is totally
buildings came down on 9/11’ false. That is patently false. That never happened. There
(22 November 2015). were no flags burning, no one was dancing. That is
bullshit.” ’ In an attempt to defend his claim, ‘the Trump
campaign posted snippets of video clips from a local
CBS New York City newscast at the time that reported
on the arrest of “eight men”—not “thousands and
thousands”—who were reported by neighbors as having
celebrated the attack.’
21 All from Kessler et al. (2016, emphasis in original) and links from that URL, unless other-
wise stated. Other fact-checkers tell basically the same stories (Yuhas 2016).
22 Kessler 2015.
‘I was totally against the war in Actually, in an interview on 11 September 2002, Howard
Iraq’ from the beginning Stern asked him, ‘Are you for invading Iraq?’ Trump
(23 February 2016). replied, ‘Yeah, I guess so.’ The fact checker continues:
‘Trump clearly was outspoken about his opposition
starting in 2004 . . . But by then—17 months after the
invasion—many Americans had turned against the war,
making Trump’s position not particularly unique. Trump
has repeatedly cited his remarks in [an] August 2004
story to support his claim that he was “totally” against
the war. In light of his repeated false claim citing this
article, Esquire added an editor’s note to [the online
version of] its August 2004 story [that] reads: “The
following story was published in the August 2004 issue
of Esquire. During the 2016 presidential election,
Donald Trump has repeatedly claimed to have been
against the Iraq War from the beginning, and he has
cited this story as proof. The Iraq War began in
March 2003, more than a year before this story ran,
thus nullifying Trump’s timeline.” ’
Expressed in another way, ‘63 percent of the 91 Trump statements that the
[Washington Post’s] Fact Checker has checked were given a Four-Pinocchio
rating—meaning they were . . . totally false’. To put that in perspective, a ‘typ-
ical candidate gets Four Pinocchios somewhere between 10 and 20 percent
of the time’, and only ‘14.2 percent of Clinton’s claims have been given Four
Pinocchios’.23
Ordinarily fact checkers just focus on major claims, often ones referred to
them by readers.24 But the magazine Politico undertook to fact-check literally
every one of Trump’s statements for a week. They found that, during the four
hours forty-three minutes worth of speeches and interviews that Trump gave
over the course of that week, he made eighty-seven ‘misstatements, exagger-
ations or falsehoods’. That is a rate of one every 3.25 minutes.25
Trump’s campaign responded to that report with characteristic bluster:
There is a coordinated effort by the media elites and Hillary Clinton to shame-
lessly push their propaganda and distract from Crooked Hillary’s lies and flailing
campaign. All of these ‘fact-check’ questions can be easily verified, but that’s not
23 Cillizza 2016; his numbers do not quite tally with Kessler’s (2016a), because his article was
written just before the election, while a few fact-checks were still underway. In keeping with his
past practice, as president Trump earned Four Pinocchios for an interview with Time magazine
about his lies (Kessler and Lee 2017). Similarly, in his first formal meeting with congressional
leaders after becoming president, Trump insistently repeated four-Pinocchio fictions about mil-
lions of illegal voters—and as president Trump launched a formal government investigation into
that matter (Johnson and Zapotosky 2017), even after being called on the lies once again by fact
checkers (Kessler 2017a; Lee 2017).
24 Kessler 2013. 25 Cheney et al. 2016.
what blog sites like Politico want people to believe. Mr Trump is standing with the
people of America and against the rigged system insiders, and it’s driving the media
crazy. We will continue to speak the truth and communicate directly with the
American people on issues they care most about, and we won’t let the dishonest,
liberal media intimidate us from speaking candidly and from the heart. A Donald
J. Trump presidency will make America great again.26
But the assertion that the facts bear out Trump’s claims is, of course, just the big
lie at work.27 It certainly is true that Politico is a left-wing outlet. But facts are
facts, and fact-checking protocols are well established.28
21.1.3 How Lies Undercut the CJT
Now, in one way, all those lies might be neither here nor there from the point
of view of the Condorcet Jury Theorem. Strictly speaking, all that the CJT says
is that (as long as its assumptions are met) the majority among a large elector-
ate will vote for the right outcome. And of course, it is logically possible that
leaving the EU was the correct outcome of the British referendum, whether or
not the UK paid £350 million per week to the EU, and that electing Trump was
the correct outcome of the American election, whether or not 11 September
2001 saw thousands of Muslims in New Jersey celebrate the collapse of the
Twin Towers.
Formally, that response is perfectly appropriate. Epistemic theories of dem-
ocracy are not public reason theories. Their emphasis is upon ‘getting the out-
come right’, not ‘giving one another good (true, honest, sincere) reasons’. Those
two styles of democratic theory would respond very differently to lying in
politics. Theories of public reason would see it as wrong in itself.29 Epistemic
theories of democracy see it as wrong only insofar as it is likely to compromise
voters’ capability to choose the correct outcome.
While it is possible for you to end up voting for the correct outcome even
though your reasoning is based on false facts, it is unlikely that you will. If your
reasoning is valid but based on false premises, your conclusions are more likely
26 Jason Miller, Trump’s senior communications advisor, quoted in Cheney et al. 2016.
27 As was Trump’s (2016a) assertion in his acceptance speech at the Republican National
Convention that ‘here, at our convention, there will be no lies. We will honor the American people
with the truth, and nothing else.’ A Washington Post fact-check identified twenty-five statements
in that very speech as being either false or misleading (Kessler and Lee 2016).
28 Poynter 2017. Trump continued making an average of 4.92 false or misleading statements a
day for the first hundred days of his presidency, according to the Washington Post Fact-check (Lee,
Kessler, and Shapiro 2017).
29 When ‘public reason’ theorists sometimes talk of relaxing the ‘sincerity’ requirement, they
do not mean to endorse uttering falsehoods but merely the giving of reasons for a course of action
that would be genuinely good reasons for others to endorse it even if those are not the speaker’s
own reasons for so doing (Schwartzman 2011).
to be wrong than if they had been based on true premises, all else equal. Choices
based on false reasoning are not necessarily incorrect, of course. But if they do
end up being correct, that will be coincidental—a fluke. Philosophically we
must not rule out flukes, but politically we should not count on them.30
False information designed to alter political attitudes is likely to undermine
the reasoning of otherwise competent reasoners, leading them to incorrect
conclusions and to vote in incorrect ways. Political lies, after all, attempt to
change the way people behave in the voting booth. If those people are ‘other-
wise competent reasoners’ (i.e. voters who would otherwise be likely to vote
correctly31), the lies changing their votes would most often change them for the
worse, epistemically speaking.
21.2 IN THE US ANYWAY, THE BIG

LIAR ACTUALLY LOST
Before we turn to the epistemically bad features of the campaigns themselves,

it is worth having a closer look at the actual results of the popular vote.
In the UK, Leave clearly won the majority of votes in the Brexit referendum.
Furthermore, it almost certainly would have done so even if turnout had not
been suppressed on the day by flooding that stranded many commuters and
closed some Underground lines in strongly pro-Remain London.32
In the US, however, the plain fact of the matter is that Donald Trump lost the
popular vote in the 2016 presidential election, and he did so by a relatively size-
able margin. Hillary Clinton beat him by almost 3 million votes. Out of a total
of more than 136 million votes cast, that represents a margin of 2.10 per cent in
favour of Clinton.33 Trump won the presidency due only to the vagaries of the
archaic Electoral College. But from a CJT point of view, that is irrelevant—or so
the argument might go.34
What that argument does not appreciate, however, is the fact that each of the
states and territories represented in that Electoral College itself has a very large
30 A fluke is just that—a ‘lucky stroke, an unexpected success, a piece of good luck’ (Oxford
English Dictionary, q.v. ‘fluke’ (n3)). Gettier (1963) showed philosophers why they matter.
31 I.e., assuming the standard CJT Competence assumption applies to them.
32 Forster 2016. ‘Leave’ scored a clear popular majority of 1,269,501 votes; even if London
turnout had been as high as the nationwide average, there would only have been around 100,000
extra London voters, not all of whom would have voted ‘Remain’ in any case (UK Electoral
Commission 2016b).
33 US FEC 2017.
34 It is also interesting—but likewise irrelevant, from a CJT perspective—that if some other
vote-aggregation procedure had been used instead of plurality rule Trump would quite likely have
been defeated in the Republican Primary Elections before ever getting into the General Election.
Maskin and Sen (2016) discuss the alternative of Condorcet pairwise comparison, but the same
would be true of a Borda count.
number of voters. Even the smallest, Wyoming, had over 250,000 people voting
in the 2016 election.35 With that number of voters we should expect the CJT to
take full effect, not just in the electorate nationwide but also at the level of each
of those state electorates. If the less truthful candidate were the wrong one to
win, and people’s votes tracked the truth with better-than-random accuracy,
Trump should have lost in every (or virtually every) state—in which case the
Electoral College would have been virtually unanimous in favour of Clinton.
Needless to say, that did not happen in 2016.
Of course, it is perfectly standard in most elections for one candidate to win
some states and the other candidate to win others. That presumably just reflects
the fact that the interests, values, and priorities of people in those states differ.
We showed in Chapters 13 and 14 how the CJT might be modified to take
account of such differences; the 2016 election is discussed in terms of that
model in Section 21.4.
For now, the crucial fact to note is simply this. The sheer fact the more
truthful candidate won the majority of votes nationwide does not, in itself,
vindicate the epistemic merits of the 2016 US presidential election. Some other
explanation is required to account for the fact that she lost the majority of
votes in so many places that, if voters were competent truth-trackers and
truthfulness were all that mattered, she should have won easily.
21.3 SENDING A STRATEGIC SIGNAL
Maybe the Brexit and Trump outcomes do not really represent the sincere will
of the majority in another way. Maybe those outcomes represent, instead,
attempts at strategic signalling that went wrong.
Here is one anecdote along those lines. A person who worked closely with
Michael Gove and the Vote Leave campaign is reported as having said after the
referendum,
We weren’t meant to win. That line, ‘you were only meant to blow the bloody doors
off ’36—it’s true. The plan was to run the Remain side close enough to scare the EU
into bigger concessions. None of us thought we were ever going to win. . . . It’s all
such a mess. I want a second referendum now.37
Just how common such sentiments might be among those who voted to Leave
can be surmised from a large-scale sample survey undertaken by the British
Election Study after the referendum. In that survey some 6 per cent of Leave
35 Wasserman 2017.
36 The line from the film The Italian Job that Michael Gove’s wife reportedly said to him the
morning after the referendum (Vine 2016).
37 Quoted in Sparrow 2016b.
voters reported regretting voting the way that they had. Furthermore, and
tellingly for the purposes of detecting strategic voting, the probability of voters
saying they regretted voting to Leave was strongly associated with a voter’s
reporting that ex ante she or he did not believe that Leave would win. Fully one
in ten Leave voters who thought ahead of the referendum that Leave had no
chance of winning said, in that post-referendum survey, that they now regretted
voting Leave.38
In the US, too, Trump was not expected to win, either by himself or by a great
many of his supporters.39 Many (probably most) of those who voted for Trump
did so despite their perception that he had little chance of winning. The strategic
voting question is this: just how many of those Trump voters voted for him
precisely because they thought he was not going to win, and hence that voting
for him would be a ‘safe’ protest vote?
Of course, a protest vote (‘sending a message’) is an expressive act as well;
that is the topic of Section 21.6. But in the circumstances here in view it counts
as a type of strategic voting, too. Unlike standard strategic voting where voters
vote as if they were pivotal because they care about the outcome, in the case of
‘expressive strategic voting’ voters vote as they do precisely because they believe
that they are not pivotal. The counterfactual test for identifying this type of
strategic voting is that the voter would have voted otherwise if that vote would
have determined the outcome of the election.40 In that case, the vote is clearly a
false indication of the voter’s own true judgement of who is the best candidate.
We have little solid evidence of how many Trump votes were strategic in
that sense.41 We must largely rely on evidence that is anecdotal and circum-
stantial. But there is a fair bit of that sort of evidence. For one thing, Trump
went into office with historically high ‘unfavourable’ ratings in the polls—and
furthermore, those ‘unfavourable’ ratings actually increased as the date of his
inauguration neared.42 Anecdotal evidence suggests that at least some voters,
distrustful of Clinton and fearing that she might win by a landslide, voted for
Trump to deny her too great a mandate.43 There is also reason to believe that
Trump’s ‘movement’ was always at least as much a protest movement as it was
38 British Election Study 2016. Economist 2016.

39 Jacobs and House 2016. Kahn 2016.
40 Of course in a large electorate it is almost never the case that any one person’s vote will
actually be decisive in this way. Still, that counterfactual constitutes the proper test to decide
whether the vote is ‘strategic’ rather than ‘sincere’.
41 Anecdotal evidence can be found in the compilation of tweets at ‘@Trump_Regrets’
(Kassam 2017).
42 Saad 2017.
43 There is for example Lu’s (2016) report of
David Marcus [who] has recently argued that at least for him (as a resident of New York
State), a Trump vote is the most reasonable form of protest vote. He doesn’t like Trump. But
he knows his state has no chance of going red, and he doesn’t want Hillary Clinton to run
away with a landslide popular vote.
a movement that aimed at actually seizing power (protesting is one thing,

actually governing is quite another).44 None of that is remotely conclusive, but
all of it is consistent with the possibility that at least some Trump voters might
have voted for him purely strategically to send a protest message and would
not have voted for him had they foreseen that he would actually win.
Further evidence along those same lines can be found in Trump voters’ own
assessment of his fitness for office. According to exit polls, an astonishing
23 per cent of those who said they voted for Trump also said that they regarded
him as ‘not qualified to serve as president’; and 27 per cent said that they did not
think he ‘has the temperament to serve effectively as president’.45 Of course, it
is perfectly possible that they voted genuinely intending to install him as presi-
dent, notwithstanding his unsuitably, on the grounds that they thought that the
alternative candidate was more unsuitable. But another interpretation, equally
or more plausible (particularly when set against the background fact that so
many of his supporters did not expect Trump to win), is that at least some of
those Trump voters were voting strategically, intending to send a message, not
intending to send their candidate to the White House. It seems likely that at
least some of them (who knows how many) might have voted otherwise, had
they foreseen that Trump might actually win.
Here is one final bit of evidence (admittedly, circumstantial once again) of
‘buyer’s regret’ surrounding Trump’s election. The Republican campaign focused
heavily on a promise to repeal the Affordable Care Act (Obamacare). In the
Kaiser Foundation Tracking Poll just before the election, 69 per cent of intending
Republican voters said they wanted to ‘repeal the entire law’; yet in the month
after the election, that had dropped precipitously to 52 per cent.46 Nothing
much had happened over the course of that month except the election. So that
looks very much like a case of ‘be careful what you wish for’—that is, Republican
voters regretting, if not necessarily their vote, anyway the consequences of
their vote for their health insurance. Of course, their original vote may have
Or for another example, Rhonnie Enterline (28, Sacramento, CA) explained to the Washington
Post, ‘If I weren’t in California where my presidential vote doesn’t count for much, I might not
have voted for [Trump]. But, I thought, why not be part of sending a message to Washington?’
(quoted in Kelly 2016).
44 In his column the day after the election, Garrison Keillor (2016b) wrote:
The Trumpers never expected their guy to actually win the thing, and that’s their problem
now. They wanted only to whoop and yell, boo at the H-word, wear profane T-shirts, maybe
grab a crotch or two, jump in the RV with a couple of six-packs and go out and shoot some
spotted owls. It was pleasure enough for them just to know that they were driving us wild
with dismay—by ‘us,’ I mean librarians, children’s authors, yoga practitioners, Unitarians,
bird-watchers, people who make their own pasta, opera-goers, the grammar police, people
who keep books on their shelves, that bunch. The Trumpers exulted in knowing we were
tearing our hair out. They had our number, like a bratty kid who knows exactly how to make
you grit your teeth and froth at the mouth.
45 CNN 2016. 46 Kirzinger et al. 2016.
been completely sincere and the regret may have set in only afterwards. But
this pattern is also consistent with at least some Trump voters having voted
strategically, intending only to ‘send a message’ and not sincerely intending that
he should be elected president.
Who knows how many Trump voters were actually thinking like that.47 It
may have been only a relatively small number. Still, even a small number of
strategic voters might have made all the difference where the margin of victory
is even smaller.
Furthermore, the number of voters engaged in this sort of ‘expressive strategic
voting’ is likely to be larger than the number engaged in strategic voting of the
more ordinary sort. Ordinary strategic voting is designed to change the result
of the election, and the chances of succeeding in that are usually pretty slim.
Expressive strategic voting, in contrast, is designed to send a protest message,
and the success of that messaging does not depend on changing the result of
the election. Knowing it is unlikely that her vote will change the outcome of the
election dissuades a voter from engaging in strategic voting of the former sort
but liberates her to engage in strategic voting of the latter sort, by voting for a
candidate for whom she would not have supported had she expected her vote
to be pivotal.
Notice finally that, given the logic of such expressive strategic voting, such
votes are invariably concentrated on one side of politics—namely, the side that
was generally expected to lose ahead of the voting. So it could well be true that,
had everyone voted sincerely, neither Brexit nor Trump would have won.
21.4 DIFFERING PRIORITIES
A highly plausible explanation of the—to many of us, surprising—outcomes of the

UK EU referendum and the 2016 US presidential election is that voters on oppos-
ing sides simply had different interests, priorities, or values from one another.
Empirically, that certainly seems true. The standard analysis of both cases is
that the priorities of metropolitan elites simply differed from those of voters
in the deindustrialized hinterlands. That much is plain from the electoral maps
that reveal sharply geographically differentiated bases of support for Trump
and Brexit respectively.48 And it is confirmed by surveys showing that voters
for the two opposing sides identified very different issues as being the ‘most
47 Polls taken around his one-hundred- day anniversary in office show that ‘just 4 percent of
Trump’s supporters say they would back someone else if there was a redo of the election’ (Blake
2017d).
48 For the US see <http://www.nytimes.com/elections/results/president> and for the UK see
<http://www.electoralcommission.org.uk/find-information-by-subject/elections-and-referendums/
past-elections-and-referendums/eu-referendum/eu-referendum-result-visualisations>.
important’.49 This also explains why, despite winning the popular vote quite
decisively, Clinton lost many swing states, especially in the Rust Belt.
Normatively, the question is simply whether even the more modest claims
for the epistemic merits of majority outcomes sketched in Chapters 13 and 14
can be sustained with respect to the Trump and Brexit majorities. As we observed
in those chapters, where people have different interests, priorities, or values,
each voter is voting on the basis of what he or she believes to be the correct
outcome from the point of view of his or her own interests, priorities, or values,
which differ from those of other voters. In such circumstances, the most that
the CJT can claim (which is still quite a lot, if it is true) is that the majority
winner will be the outcome that is correct from the point of view of the interests,
priorities, or values of the majority of voters—assuming that each voter is better
than random at choosing the correct outcome for furthering his or her own
interests, priorities, or values.
But is that Competence Assumption warranted in the case of Brexit or Trump
voters? Perhaps it is in terms of the values and priorities manifested by Leave
voters in the Brexit referendum—Leave presumably is indeed the correct out-
come if, as post-referendum polling suggests, the priorities of Leave voters
were to restore British sovereignty over laws that applied in Britain, to reduce
immigration (at least from within the EU), and to reduce or eliminate costly
British contributions to the EU.50 Whether Leave is in the objective economic
interests of those who voted for it in other senses is another question, however.
EU regional funds go disproportionately to regions that voted in favour of
Leave, which would of course result in those funds being cut off, for only the
most salient example.51 But if this referendum was less about the pay cheque
and more about identity, perhaps Brexit is just what the majority wanted.
Similarly in the US, it may be that Trump’s diffuse slogan, ‘Make America
Great Again’, resonated with his voters’ values and priorities. Maybe his vague
promises to prevent jobs from moving abroad and restricting immigration
resonated with their values and priorities, too. Or anyway maybe it sounded as
if voting for Trump was the right way for them to promote those values and
priorities. Given how little detail he offered, however, it would have been hard
49 In the 2016 US election, exit polls showed Trump voters were far more likely to say that
‘immigration’ or ‘terrorism’ were the most important issues, compared to Clinton voters (by 64%
to 33% and 57% to 40% respectively); Clinton voters were far more likely to nominate ‘foreign
policy’ (60% to 33%) and, oddly enough, ‘the economy’ (52% to 41%) (CNN 2016). In the UK Brexit
referendum, an eve-of-poll survey found Leave voters were far more likely to say that ‘immigration’
and ‘Britain’s right to act independently’ were ‘most important to you in deciding how to vote’ (by
35% to 2% and 45% to 18% respectively); Remain voters were far more likely to say ‘jobs, investment
and the economy generally’ (by 60% to 8%) (YouGov 2016a). See further Hobolt and Leeper 2017.
50 Luck 2016.
51 Dean 2016. EU funds might be replaced by ones from the UK central government, of course;
but if they were, then that would reduce the ‘£350 million per week’ savings from leaving EU that
would be available for the UK government to spend elsewhere (on the NHS, as Leave promised,
for example).
to say for sure. Indeed, given how much he lied about everything else during
the campaign, it would have been hard to say whether there was even a better-
than-random chance of that being true.
Whatever doubts we might have about whether voting for Trump was the
right way for his supporters to promote their own values and priorities, those
doubts are redoubled when it comes to the question of whether voting for him
was the right way for his supporters best to serve their own objective interests.
Again, Trump’s policy proposals were so sketchy and incompletely specified at
the time of the election that no voter could have had the remotest way of deter-
mining whether they were genuinely in his or her own objective interests.
Given how much he lied about everything else, voters could not even be confi-
dent that there was a better-than-random chance that he would do what he
promised, insofar as he did make any specific promises. And indeed, on many
topics the balance of evidence available to them should have suggested that the
policies Trump was most likely to pursue were not in their objective interests.
Consider for example the Republicans’ promise to repeal Obamacare, one of
the central and most specific planks of their campaign and one that Trump
partially acted upon in one of the first Executive Orders he signed upon
assuming the presidency. The fact of the matter is that it is only because of
Obamacare that a great many of Trump’s poor and unemployed Rust Belt sup-
porters have insurance to protect them against the extremely high costs of US
medical care. The non-partisan Congressional Budget Office estimates that, on
the initial Republican plan for repeal-and-replacement (H.R. 3762), ‘the num-
ber of people who are uninsured would increase by 18 million in the first . . . year
following enactment of the bill’, and ‘premiums . . . would increase by 20 percent
to 25 percent’.52 Assuming that estimate is correct, the interests of a great many of
Trump’s voters in affordable health insurance will be ill served by that policy.53
That is the case not only objectively but subjectively as well, judging from
focus groups with Trump supporters in Rust Belt states in December 2016. In
those discussions,
Several [participants] described their frustration with being forced to change plans
annually to keep premiums down, losing their doctors in the process. But asked
about policies found in several Republican plans to replace the Affordable Care
Act—including a tax credit to help defray the cost of premiums, a tax-preferred
savings account and a large deductible typical of catastrophic coverage—several of
these Trump voters recoiled, calling such proposals ‘not insurance at all.’54
There is one final way to try to make sense of how people might have seen vot-
ing for Trump or Brexit as being in their objective interests. Maybe they were
52 CBO 2017.
53 In certain clear respects, at least: perhaps in other respects they think of themselves as having
a ‘dignity’ interest in being able to afford to pay for health insurance without a state subsidy.
54 Altman 2017.
thinking ‘nothing could be worse than the status quo, and at least they will
shake things up’. Trump put the point precisely like that in attempting (largely
unsuccessfully) to appeal to black voters, asking, ‘What the hell have you got to
lose?’55 Interviews with many of Trump’s white working-class supporters indi-
cate that they, too, were thinking along those lines.56 In Britain, many backed
Brexit having based their decision on similar reasoning.57
Again, that may well be an accurate characterization of the thinking of some
(perhaps many) Trump and Brexit voters. They voted for change simply out of
deep despair with the status quo. But for the Chapter 14 version of the CJT to
apply, it must be the case that such voters were actually correct in thinking that
‘nothing could be worse’ than the status quo.58 When terminal cancer patients
volunteer to participate in trials of new drugs, bioethicists worry that the quality
of their consent might be compromised by the ‘therapeutic illusion’, leading them
to suppose that what the researchers intend only as an experiment with a new
drug might actually have positive therapeutic effects.59 From the point of view of
willing research subjects, however, it is not at all a bad bet: after all, there is some
chance (however small) that the drug will work; and terminally ill cancer patients
genuinely do have nothing to lose. Does anyone seriously believe that Trump and
Brexit voters can say the same with anything like the same confidence?60
Thus, it may well be that their distinctive priorities, values, and interests drove
a majority of voters to support Trump in the US and Brexit in the UK, contrary
to the priorities, values, and interests of the rest of the electorate. But whether
the moderately happy CJT conclusion we adduced in Chapters 13 and 14 follows
is in doubt. Certainly, as regards their interests, and perhaps even their values
and priorities, it is far from clear that voting for Trump or Brexit really was
indeed the correct way for those people to best further their own objectives.
55 Bump 2016. Newt Gingrich summarized the ‘Principles of Trumpism’ to the Heritage
Foundation a month after the election in the phrase ‘he repeated again and again: “Donald Trump’s
gonna kick over the table” ’ (Gingrich 2016, quoted in Godfrey 2016).
56 Kelly 2016.
57 In the postmortems on Brexit, one sixty-two-year-old London jobseeker was quoted as say-
ing, ‘Leaving the EU might make my life shit, but it’s shit anyway. So how much worse can it get?
I’ve got nothing to lose . . .’ (Martin Parker, quoted in Ryan 2016).
58 A more sophisticated, and more plausible, version of this line of thought would be couched
in terms of ‘the probability is sufficiently high that they will do something sufficiently better for
me than the status quo does’. Maybe that is what some supporters of Trump and Brexit were
thinking; but the more extreme formulation in the text is how they actually put it.
59 Casarett 2016.
60 Bump 2016. Ehrenfreund 2016. A former George W. Bush speechwriter summarizes the
strategy thus:
Because poor neighborhoods can’t get any worse, why not try something new? Because
America is already a jihadist battleground, why not take a radical and discriminatory new
direction on immigration? Because the planet is in chaos, why not entirely reorient American
foreign policy toward alliances and great power rivals? Things, after all, can’t get any worse.
Michael Gerson then pointedly adds: ‘The problem is: Things can get a lot worse, and quickly’
(quoted in Tumulty and Nakamura 2017).
21.5 OPINION LEADERS LIED,

AND VOTERS BELIEVED THEM
Let us now turn our attention to the campaigns and their epistemic flaws.
Trump and leaders of the Brexit Leave campaign lied; we know that from
Section 21.1. We also know that a lot of people paid attention to what they said.
Surveys found, curiously enough, that Boris Johnson (who had twice previ-
ously been fired for lying) was the political leader most trusted by Britons who
voted to Leave the EU.61 And Donald Trump, whose preferred mode of messa-
ging is the 140-character tweet, has over 20 million Twitter followers and ‘an
audience attentiveness score of 75%’.62 In short, in both the UK and US cases
there were strong opinion leaders who lied.
But did the voters actually believe their lies? In the UK, we have survey evi-
dence showing that they did. A poll conducted shortly before the referendum
found that 47 per cent of all respondents believed to be true the claim that
‘Britain sends £350 million a week to the European Union’, and only 39 per cent
believed it to be false.63 Furthermore, in polling after the referendum, nearly
one in five Leave voters named that claim as their primary reason for voting to
Leave the EU.64
We do not have systematic polling in the US pertaining to each of Trump’s
many lies. But we do have anecdotal evidence that at least some of his voters
genuinely believed at least some of the ‘fake news’ promulgated by Trump and
his supporters. Perhaps the most famous example concerns the ‘Pizzagate’ myth,
generated by social media and subsequently fuelled by a tweet from Michael
G. Flynn, a member of Trump’s transition team and son of Trump’s national
security advisor designate at the time. That story linked ‘Hillary Clinton, her
campaign chairman and the owner of [the pizza shop] Comet Ping Pong to [an]
alleged sex-slave conspiracy’.65 A North Carolina man, Edgar Maddison Welch,
became so fixated on [that] fake news story that he drove [to Washington] . . .
determined to take action. . . . For 45 minutes . . . Welch, cradling an AR-15 assault-
style rifle, roamed the Comet Ping Pong pizza restaurant looking to prove an
Internet conspiracy theory. . . . With D.C. police amassing outside . . ., Welch finally
walked out with his hands up—but not before he finished his search. He had come
to rescue the children, court papers say he later told police, and now was convinced
that none was being harmed there.66
61 YouGov 2016b. In previous incarnations, Johnson had ‘been sacked twice, and on both
occasions . . . for dishonesty: once by the Times for making up a quote, and again by the former
Tory leader Michael Howard, for lying to his face about an extramarital affair’ (Freedland 2016;
Major 2016).
62 According to TwitterCounter (2017). 63 Ipsos MORI 2016, p. 6.
64 Luck 2016, p. 8.
65 According to one report, ‘despite being widely debunked and described by the police as
“fictitious” [that story] was still believed by 9% of registered voters’ (Naughton 2017).
66 Hermann et al. 2016. Washington Post Editorial Board 2016.
To such anecdotal evidence, we can add a certain amount of polling data. Some
pertain to specific statements on which Trump backers insistently endorse his
version of the facts despite clear evidence it is false. Here are two examples. In
a post-election poll, 52 per cent of Republicans said that Trump really won the
popular vote.67 And in another post-inauguration poll, respondents were shown
two photos of the Washington Mall, one taken during Obama’s 2009 inaugur-
ation and the other taken during Trump’s 2017 inauguration; when asked in
which photo there were more people, 15 per cent of Trump supporters denied
the clear evidence of their own eyes and asserted that the 2017 photo showed
more people on the Mall.68
We can add to that other polling data on the more general question of
whether voters consider each of the candidates as ‘honest and trustworthy’.
Neither Trump nor Clinton scored highly on that among the electorate as a
whole.69 But each did well among his or her own voters.70 Fully 94 per cent of
Trump voters reported believing Trump to be ‘honest and trustworthy’, despite
his having been shown to have been lying so repeatedly.71 Perhaps respondents
did not believe Trump’s specific assertions, but they nonetheless thought him
to be speaking some ‘deeper truth’. We will discuss that possibility in Section
21.7.2. Still, the responses taken at face value seem to constitute at least prima
facie evidence that a large proportion his supporters actually believed Donald
Trump’s many lies.
In Chapter 11 we discussed the phenomenon of opinion leadership from a
CJT perspective. There we argued that, if voters follow opinion leaders rather
than exercising their own independent judgement, those voters are (at best)
collectively only as likely to be correct as their opinion leader is. If an opinion
leader is wrong (whether innocently or intentionally so) in the guidance he
provides, voters following him will be misled accordingly and the wrong out-
come may well win a majority as a result.
In terms of the Condorcet Jury Theorem as it is standardly construed, the
fault lies with a failure of the Independence Assumption as applied to voters
67 Oliver and Wood 2016.

68 Schaffner and Luks 2017. Whether people really believed that, or whether they were just
saying that in order to support Trump, is of course an open question. Other evidence suggests that
‘partisan bias’ (and motivated reasoning and confirmation bias more generally) is overcome with
only rather modest ‘accuracy incentives’ (Prior et al. 2015).
69 Clinton by 64% to 36%, Trump by 64% to 33% (CNN 2016). But when other polls asked which
candidate they regarded as more honest, around 45% persistently said ‘Trump’—astonishingly
enough, given the evidence reported in Section 21.1.2 (Cillizza 2016).
70 Ninety-four per cent of Trump voters considered him ‘honest and trustworthy’, as did the
same percentage of Clinton voters her. When exit polls asked ‘which candidate is honest?’ only
29% of all voters responded ‘neither is’ (CNN 2016).
71 CNN 2016. A month into his presidency, 78% of Republicans still said they trusted Trump
rather than the media ‘to tell the truth about important issues’ (Sargent 2017a). As Jonathan Swift
(1710) quipped in an earlier era, ‘as the vilest writer hath his readers, so the greatest liar hath his
believers’.
who follow opinion leaders. But in the case of opinion leaders who deliberately
lie, there is a second failing as well—a second-order failure of Sincerity. In our
previous discussions, we have implicitly been assuming that opinion leaders at
least make a good-faith effort to lead their followers to what they themselves
genuinely believe to be correct outcomes. Opinion leaders might be wrong
about that and accidentally mislead their followers in consequence. But when
they deliberately lie, they intentionally mislead their followers—certainly in
what to believe, and quite probably in how to vote as well.
Formally, nothing changes in the Chapter 11 analysis of opinion leadership,
of course. The probability that the majority of an opinion leader’s faithful fol-
lowers will support the correct outcome is still fixed by the probability that the
opinion leader himself will support the correct outcome. The only difference is
that, with lying opinion leaders, that probability is likely to be lower than with
truthful opinion leaders. The analytics are identical—the upshot is merely
more depressing.
In Chapter 11 we were relatively sanguine about the dangers of opinion lead-
ership. Our grounds then were that the overall effects of opinion leadership
would be likely to cancel out if: (a) there are multiple, independent opinion
leaders commanding different segments of the electorate as followers; or
(b) there are many independent, lower-level opinion leaders mediating the
influence of top-level opinion leaders. But those happy predictions fail when
top-level opinion leaders have direct, unmediated access to voters, via Twitter
and Facebook and such like, or when too many opinion leaders are intercon-
nected with one another and send the same message. Both were the case, defin-
itely with the 2016 US presidential election and arguably with the UK EU
referendum as well—again, making the implications for Chapter 11’s analytics
more depressing as applied to those two cases.
21.6 AFFECTIVE EXPLANATIONS
In Section 21.3, we already briefly looked at strategic protest voting—voters

hoping to send a ‘warning shot’ message without changing the outcome. But
what precisely do these voters intend to communicate? Here we will examine a
suite of ‘affective’ explanations for the outcomes of the two 2016 elections under
discussion.
These explanations are sometimes lumped together under the heading of
‘expressive voting’.72 The essence of that claim is that a voter votes to ‘express’
something (which can be accomplished by the sheer act of voting in itself)
rather than for any consequentialist reasons that depend on his or her vote
actually changing the outcome of the election.
72 Brennan and Lomasky 1993.

For CJT purposes, however, it matters less whether people are voting for
consequentialist or non-consequentialist reasons. What matters is instead
what voters are trying to express—namely, whether or not they are expressing,
through their votes, choices that they think to be tracking the truth. As we
argued in Section 4.3.3, it is perfectly coherent for a voter to vote perfectly
truthfully and sincerely for what he or she believes should be the correct out-
come of the election, without being under any misapprehension whatsoever
that his or her vote has any realistic chance of causing that outcome to prevail.
Hence in our discussion of affective voting we will be sensitive to what affects
were involved and what truth value, if any, they might have.
21.6.1 Expressing Emotions
In naming ‘post-truth’ its word of the year for 2016, the Oxford English Dictionary
defines it as ‘relating to or denoting circumstances in which objective facts are
less influential in shaping public opinion than appeals to emotion and personal
belief ’.73
The appeal to emotion, and a corresponding indifference to facts, is far from
new in politics, of course.74 But in the 2016 US presidential election, it was carried
to new heights. As President Obama remarked of his successor, ‘Trump under-
stands the new ecosystem, in which facts and truth don’t matter. You attract
attention, rouse emotions, and then move on. You can surf those emotions.’75
Trump spokespeople sometimes forthrightly admitted that they were
playing on emotions that were not only ungrounded in facts but indeed flatly
contrary to them. Take for example this CNN interview with Newt Gingrich.
When the interviewer challenged Trump’s false claims about soaring crime
rates, Gingrich replied,
‘The average American—I’ll bet you this morning—does not think crime is down,
does not think they are safer.’
‘But we are safer and it is down’, says [the interviewer], citing FBI data to that effect.
‘No’, says Gingrich. ‘That’s your view. . . . What I said is also a fact’, Gingrich con-
tinues, as if patiently explaining something obvious to a child. ‘The current view is
that liberals have a whole set of statistics that theoretically might be right, but it’s
not where human beings are.’
73 Flood 2016.
74 As Elizabeth Drew commented on Ronald Reagan’s first tilt at the presidency in 1976, his
appeal had ‘to do not with competence at governing but with the emotion he evokes. Reagan lets
people get out their anger and frustration, their feeling of being misunderstood and mishandled
by those who have run our government, their impatience with taxes and with the poor and the
weak, their impulse to deal with the world’s troublemakers by employing the stratagem of a punch
in the nose’ (quoted in Rich 2016).
75 Obama 2016.
Confronted with the fact that the crime statistics cited come from the FBI—hardly
a ‘liberal’ organization—Gingrich makes it clear that he doesn’t care. ‘No, but what
I said is equally true. People feel more threatened. As a political candidate, I’ll go
with how people feel, and I’ll let you go with the theoreticians.’76
In the UK, Brexit campaigners deliberately took a leaf from the early Trump
campaign in that respect. Arron Banks, the chief financial backer of the Leave.
EU campaign, attributed its success to hiring a Washington firm that taught
them that ‘facts don’t work’. He went on to say, ‘The Remain campaign featured
fact, fact, fact, fact, fact. It just doesn’t work. You have got to connect with
people emotionally. It’s the Trump success’—both in the primary elections of
which Banks was speaking and in the general election that followed.77 Trump’s
strategy in that respect was mimicked by the Leave campaign in the Brexit
referendum.
How ought we to evaluate those emotional appeals in terms of the Condorcet
Jury Theorem? One approach, along the lines of Chapter 13, might be to say that,
so long as people are more likely than not to be correct about what worries
them and what it would take to ease those concerns, the vote of the majority is
highly likely to point to the correct outcome for the purposes of easing the
majority’s concerns. But when people’s worries are based on false facts, then
there seems little reason to believe that people will indeed be more likely than
not to be correct (or even better than random) about what actions would be
best suited to easing their concerns.
21.6.2 Expressing Identity
Certainly one aspect of identity—‘party identity’—ended up being central to

Trump’s victory. During the campaign it did not look as if it was going to be.
Republicans, both elite and mass, had been wary of supporting Trump after
what many regarded as his ‘hostile takeover’ of their party. Former Republican
presidents refused to endorse him; Barry Goldwater’s daughter denounced
him. Yet, in the end Republican voters came around. In the wake of the dual
gratuitous interventions by the FBI Director late in the campaign, undecided
Republican voters reluctantly got behind their party’s candidate—and that
seems to have been the main reason the polls (which Clinton had been leading
comfortably throughout the previous month) closed so dramatically in the
closing days of the campaign.78 Come election day, exit polls showed the same
proportion of self-identifying Republicans voting for Trump as of self-identifying
Democrats voting for Clinton.79
76 Loofbourow 2016, emphasis in original. 77 Booth et al. 2016.

78 Silver 2016. 79 CNN 2016.
That self-identifying Republicans should identify with the Republican

candidate is one thing. That economically hard-pressed Rust Belt voters should
identify with the occupant of a gold-plated penthouse atop a Manhattan mini-
skyscraper bearing his own name is quite another. Personal style is a large part
of the story, perhaps. Whatever personality disorder drives Trump, it leads
him to behave in ways reminiscent of a short-order cook in a Tuscaloosa
greasy-spoon restaurant—behaviour that clearly resonates with a large swathe
of voters in flyover America.80 That is how an eighty-one-year-old Pennsylvania
restaurant owner explains Trump’s appeal to so many of his formerly Democratic
customers: ‘With the majority of them, I think it was his ordinary man’s con-
versation. It wasn’t rehearsed. He said it like he felt it was. They all identified
with the guy.’81
In addition to Trump’s presentational style—his limited vocabulary, fractured
grammar, persistent misspellings—there is something else that attracted Rust
Belt voters.82 That is what sociologists term the ‘hidden injuries of class’.83
Trump’s Twitter rants – their explosive and unmediated primal fury – tap into a
deeper wellspring than just economic anxiety. His in-the-moment, consequence-
free, grandly unedited Twitter style is a potent fantasy for working-class people
who have to step cautiously through the daily discouragements of their lives. . . .
These are people who lack agency. Who are resigned to a bite-your-tongue-and-
take-crap relationship with their world; a battery of daily demeanments. These
come from a brew of horrendous bosses; credit-stealing and slothful coworkers;
disconnected and oblivious senior management; overbearing in-laws; demanding
children; idiot foremen; and non-responsive insurance company bureaucrats
(who themselves have no agency).
Enter Trump. Every time he responds to a big attack or a micro-slight, however
undisciplined and dramatically over-aggrieved it might be, the cathartic joy meter
lights up like the Christmas tree in the lobby of Trump Tower.
Good for you, Donald. You don’t take any shit. If I were a billionaire neither would I.84
Needless to say, the sheer fact that Donald Trump’s Twitter tantrums make left-
behind voters feel good—the fact that he does outrageous things that they wish
80 McAdams 2016. James 2016.

81 Quoted in McCarthy 2017. Similarly, a woman at a Trump rally explained that ‘he’s down to
our level. He’s not like past presidents who prepare a beautiful speech but it’s not coming from the
heart. He speaks it like it is’ (quoted in Pilkington 2016).
82 Milbank 2017. Sclanfani 2017.
83 Sennett and Cobb 1993; Jütten 2017. The president of the American Enterprise Institute
explains it in terms of how ‘people are stripped of their sense of dignity . . . when they feel super-
fluous to society, when they feel that they are not needed . . . Donald Trump was talking to people
in the parts of America that have been truly forgotten and left behind now for generations in a way
that . . . helped people understand that he understood . . . ’ (quoted in Capehart 2017). For more
in-depth analyses of the current American malaise in these terms, see Hochschild (2016) and
Isenberg (2016).
84 Hanft 2016.
they could do—in no way proves that he is the correct candidate to support,
even just in terms of promoting their own values and priorities, much less
their interests.
Some political theorists say that it is good, from a democratic point of view,
if people elect representatives who are ‘just like themselves’. The reason they say
this is that, if voters do so, their representatives in the legislature are likely to
vote just the same way as their constituents would have done had they been
there.85 There is no reason to think that that is true of Trump, however. Strongly
though his supporters might identify with Trump’s behaviour and his personal
style, that is no guarantee whatsoever of any identity of interests, values, policy
preferences, or priorities. If voters use these as cues (in ways we discussed in
Chapter 12), they misfire spectacularly.
At best, any association is random. And it might be worse than that, insofar
as whatever it is in Trump’s behaviour that his supporters see as akin to their own
actually has a much different source (if, for example, he is just a psychologically
disturbed rich kid). Hence a Chapter 13-style CJT defence of the epistemic
merits of a majority in favour of his election seems on shaky ground in this
respect as well.
21.6.3 Having Fun
Another broadly affective explanation of the Trump success may be just this.
Perhaps his supporters did not believe, or even care, what he was saying. Maybe
they were just having fun. Maybe it was purely entertainment, and in voting for
Trump they were saying nothing more than they ‘want the show to go on’—just
as they might ‘vote’ for a reality television contestant who amuses them to
remain on the show for another week. And maybe the same was true to a lesser
extent with the UK Brexit referendum, where everyone remarks upon how one
of the leading Leave campaigners, Boris Johnson, insistently plays the part of a
‘clown’ and a ‘buffoon’.86
Again, the blurring of ‘news’ and ‘entertainment’ is far from new. Just recall
the ‘yellow journalism’ that sold all those Pulitzer and Hearst newspapers in the
1890s. Or, in more recent times, think of the sorts of ‘newspapers’ that you see
at supermarket checkout counters.87 No one buys them because they believe
the truth of their stories of alien abduction or Martians landing in Soldier’s
85 Miller and Stokes 1963. 86 Frayer 2016.

87 ‘An army of crazed monkeys. John Belushi’s drug dealer. Lee Harvey Oswald’s autopsy photo.
The contents of Henry Kissinger’s trash cans. A woman who used her son’s face as an ashtray. The
presidential candidacy of Donald Trump. . . . Over the years, an array of jaw-dropping oddities has
drawn readers to the National Enquirer. . . . Now it is the real estate developer’s turn on top’
(Gillette 2016). But the National Enquirer is the semi-respectable face of supermarket tabloids.
For an even more extreme case, consider the Weekly World News (Heller 2014).
Field. People buy them for their entertainment value, not their news value.
Or anyway, most people do.
Arguably the Trump phenomenon was largely about entertainment.
Undoubtedly that was so in its early stages.88 Right throughout the campaign,
however, ‘cable news networks routinely broadcast Trump rallies not for their
civic content but for their ratings boost’, as Trump’s surrogate Newt Gingrich
boasted in an address on ‘The Principles of Trumpism’ to the right-wing Heritage
Foundation a month after the election.89
Much about Trump’s rallies did indeed suggest that they were about enter-
tainment more than anything else. The pulsating old rock anthems and the
call-and-response chants remind one of nothing more than a small-town high
school football pep rally (an impression strongly reinforced by the fact that
Trump’s rallies were indeed typically held in pretty small venues in pretty small
towns). Most tellingly, perhaps, is the fact that in his seemingly extemporaneous
rambles at those events Trump himself repeatedly made much of the fact that
everyone was ‘having fun’.
Trump often opened rallies with a rhetorical question, ‘Is there anyplace more
fun to be than at a Trump rally?’.90 He repeated it often. As a protestor was being
roughly evicted from one of his rallies, Trump famously said, ‘Try not to hurt
him, [but] if you do I’ll defend you in court’—immediately adding, once again,
‘Are Trump rallies the most fun? We’re having a good time.’91 Recalling the rally
in which he famously instructed a mother with a crying baby to leave the room,
Trump conjured up this image in his mind’s eye: ‘Everyone’s having fun, we’re
smiling, I’m waving. Everyone’s having fun. . . . ’92
Of course, it’s perfectly possible for political rallies to be genuinely enjoyable
without their being nothing but pure entertainment. Still, one highly plausible
way of reading Trump’s rallies, and his campaign more generally, would be as
pure entertainment. To adapt a pithy phrase from a former Australian prime
minister, Trump simply ‘threw the switch to vaudeville’.93
88 Here is the report of CNN interviews with 150 Trump supporters early in the campaign (Lee
et al. 2016):
When he hit the campaign trail [immediately after announcing his candidacy], the crowds
quickly swelled. Thousands were soon turning up at school gymnasiums, auditoriums and
local event halls to see Trump in person, forming long, winding lines that often spilled
into overflow rooms. In the first weeks and months of Trump’s campaign, plenty of
attendees admitted they were there to catch a glimpse of the former host of the ‘The
Apprentice’ – maybe even shake hands with the TV star.
89 Vyse (2016), glossing Gingrich (2016). See also Halloway 2016. By one estimate, Trump
received $2 billion worth of free media coverage during the first half of the primary election
campaign alone (Confessore and Yourish 2016).
90 Kizenko 2016. 91 Moyer, Starrs, and Larimer 2016. Blake 2016.
92 Flores 2016. 93 Paul Keating, quoted in Kelly (2009, p. 35).
Sometimes entertainment can have cognitive content. It does in the case

of satirical skits or jokes, for example.94 Perhaps certain of the entertainment
aspects of Trump’s campaign, too, contain the kernels of some genuine truths.95
Clearly, participants’ sense of identity can sometimes be accurately represented
and reinforced (as well, of course, as being sometimes created de novo) through
pep rallies, whether high school football teams or for presidential candidates.
If the entertainment aspects of the Trump campaign somehow accurately
conveyed the interests, values, or priorities between the candidate and his audi-
ence, then perhaps a CJT story of sorts can be told about them for that reason
as well (along the lines of Chapter 13 and Section 21.4).
Insofar as people embraced the Trump campaign ‘just for fun’, however, they
were simply not being serious.96 And from an epistemic point of view, their
votes for their favourite performer should not be taken seriously, in deciding
who should occupy a far more consequential role than that.
21.7 EPISTEMIC INSOUCIANCE
Why did voters let political actors get away with falsehoods and lies? Quassim
Cassam has coined the helpful term ‘epistemic insouciance’. ‘Insouciance in the
ordinary sense [suggests] unconcern, carelessness or indifference.’ Epistemic
insouciance, ‘the form of insouciance to which’, Cassam believes, ‘some politi-
cians are prone, is indifference or unconcern with respect to whether their claims
are adequately grounded in reality or in the best available evidence.’97
Harry Frankfurt more prosaically calls it ‘bullshit’. Here is how he distin-
guishes that from lying (what Cassam would call ‘epistemic malevolence’, the
subject of the Section 21.9):
When an honest man speaks, he says only what he believes to be true; and for the
liar, it is correspondingly indispensable that he consider his statements to be false.
For the bullshitter, however, all these bets are off: he is neither on the side of the
true nor on the side of the false. His eye is not on the facts at all, . . . except insofar
as they may be pertinent to his interest in getting away with what he says. He does
not care whether the things he says describe reality correctly. He just picks them
out, or makes them up, to suit his purpose.98
94 People are counting on that heavily in the wake of Trump’s victory: as the cover of Atlantic
Monthly put it, ‘Can Satire Save the Republic?’ (C. Jones 2017).
95 That is to say, maybe they were akin to the ‘true fictions’ discussed in Section 21.7.2.
96 Mel Brooks describes Trump as ‘just a song-and-dance man’ who ‘didn’t expect to win’ the
election. ‘He didn’t take it seriously. Three hundred million Americans didn’t take it seriously.
Now they do’ (quoted in Queenan 2017).
97 Cassam 2016, p. 2. 98 Frankfurt 1988p. 131.
Donald Trump displays this attitude in spades.99 He simply ‘doesn’t seem to

care whether [his assertions] can be proven false five minutes later’.100
Donald Trump either cannot tell the difference between truth and lies, or he
knows the difference but does not care. Tiniest example: On a single day during
the campaign, Trump claimed that the National Football League had sent him a
letter complaining that the presidential-debate schedule conflicted with NFL
games (which the NFL immediately denied), and then he said the Koch brothers
had begged him to accept their donations (which they also flat-out denied).
Most people would hesitate before telling easily disprovable lies like these, much as
shoplifters would hesitate if the store owner is looking at them. Most people
are fazed if caught in an outright lie. But in these cases and others, Trump never
blinked. . . . David Fahrenthold (and Robert O’Harrow) of The Washington Post
offered astonishing documentation [from his testimony in a 2007 lawsuit he had
brought against an unflattering biographer] of Trump being caught in a long string
of business-related lies and simply not caring.101
Here of course we are concerned with explaining the outcomes of the Brexit
referendum and the 2016 US presidential election. Hence, we are concerned
with the attitudes not only of leaders but also of their followers. To what extent
might the success of campaigns based on lies in those two cases reflect ‘epistemic
insouciance’ on the part of voters, as well as of their leaders?102
Consider what it would mean, for the Condorcet Jury Theorem, if that were
thoroughgoingly true. If voters were completely indifferent to the truth when
casting their ballots, then no epistemic claims can be made on behalf of the
outcome of the voting. The fact that the majority voted one way or another
would be of epistemically no moment, if voters were not even trying to track
the truth in the way that they voted.
To foreshadow: we shall show that voters may well have displayed a fair bit
of epistemic insouciance in both the British referendum and the American
election of 2016. As we shall also show, however, there are various different
ways of and reasons for being indifferent to the truth of politicians’ utterances.
If voters displayed epistemic insouciance towards some facts but not others,
then there might be ways in which some modest CJT-style epistemic claims for
the merits of the majority could be vindicated. We doubt that such a vindication
could succeed in relation to the two campaigns here in view; but we acknowledge
it as possible.
99 Frankfurt (2016) himself supposes Trump more often to be lying, on the grounds that he
either knew or could and should have known his statements were untrue. But by Frankfurt’s own
definition, someone can remain a bullshitter whilst saying all sorts of things knowing them to be
untrue, just so long as he does not say them because they are untrue.
100 Fallows in Rehm (2016b). See similarly Swift’s (1710) description of English politicians of
his day.
101 Fallows 2016a. Fahrenthold and O’Harrow 2016.
102 We analyse the related phenomenon of ‘epistemic agnosticism’ among the mass public in
Section 21.10.
21.7.1 Voters Were Indifferent on the Topics of the Lies
One version of the epistemic insouciance argument connects with our dis-
cussion of people’s ‘different priorities’ in Section 21.4, and we can dismiss it
equally quickly.
The speculation here is that voters may have been prepared to overlook cer-
tain of politicians’ lies because they were indifferent to the things about which
the politicians were caught lying. Perhaps people just did not care about those
things; and they thought the politicians were actually telling the truth on those
matters that genuinely concerned them. As one commentator speculated,
‘Who cared if Trump denied sexually harassing women, when he was so boldly
telling the truth about the fear, rage, racism, xenophobia and misogyny that
many of his supporters felt but had hesitated to voice?’103 This is an argument
often advanced by Trump’s surrogates (albeit sometimes in the face of over-
whelming evidence to the contrary).104
That explanation does not particularly ring true of the 2016 American elec-
tion, however. The reason is simply that Trump lied about so very many things,
it is hard to believe that any given voter simply did not care about any of them.
Indeed, impartial fact checkers caught him lying about virtually every issue
central to his campaign.105 Trump did not lie just about things of peripheral
interest to his supporters. His lies would have undercut all the central messages
of his own campaign, or anyway they would have done so for anyone prepared
to accept incontrovertible evidence from impartial fact checkers.
Neither does this explanation ring particularly true of the Brexit referendum.
According to post-referendum polling, the lie upon which we have here been
focusing—that the UK was sending £350 million per week to the EU—was the
principal reason behind the votes of fully one in five of Leave voters, and pre-
sumably at least a secondary consideration for a great many others.106 Again,
what was being lied about was hardly a peripheral issue for Leave voters.
103 Quoted in Prose 2016. Another woman, asked ‘about the impact on her daughter of poten-
tially having someone in the White House who brags about groping women’s genitals’, replied: ‘I’ll
teach my own daughter to be independent and stand up for herself; that’s my job, not the presi-
dent’s’ (quoted in Pilkington 2016).
104 Reneging on Trump’s campaign promise to release his tax returns as soon as his IRS audit
was over, his spokesperson Kellyanne Conway announced two days after his inauguration that ‘he’s
not going to release his tax returns’ explaining, ‘People don’t care’. A Washington Post-ABC poll just
the week before ‘showed that Trump’s continued refusal to release his tax returns continued to be
an unpopular decision, with 74 percent of Americans saying he should make the documents pub-
lic, including 53 percent of Republicans’ (Wagner 2017). Perhaps more plausibly, Newt Gingrich
(quoted in Baker et al. 2017) said in response to Trump’s disastrous first week in office,
The average American isn’t paying attention to this stuff. They are going to look around in
late 2019 and early 2020 and ask themselves if they are doing better. If the answer’s yes, they
are going to say, ‘Cool, give me some more.’ . . . There are two things he’s got to do between
now and 2020: He has to keep America safe and create a lot of jobs. . . . If he does those two
things, everything else is noise.
105 Kessler et al. 2016. 106 Luck 2016.
21.7.2 True Fictions
A second version of the epistemic insouciance argument turns on a notion of

‘true fictions’.
Ronald Reagan—to whom Donald Trump bears many biographical and
behavioural similarities, despite their stark ideological differences107—was a
master of ‘true fictions’. Reagan was forever couching his political points in
terms of anecdotes.108 Often they were made up or half-remembered plots of
movies he had once seen.109 The stories were apocryphal—‘of doubtful authen-
ticity; spurious, fictitious, false; fabulous, mythical’.110 But they spoke to a deeper
truth. They were stories that ‘should be true’, even they were not. Like ‘myths’,
Reagan’s false stories evocatively encapsulated some generalities that were
arguably true, even if the particular anecdotes themselves were not.111
Trump’s falsehoods are importantly different in many ways from Reagan’s
anecdotes, however.112 The latter were offered purely as illustrative of some
more general principles that were supposed to stand in their own right. The
truth of those general principles did not depend in any way on the truth of the
anecdotes. The propositions espoused in Trump’s lies, in contrast, purport to
provide evidentiary support for the positions he espouses, in which case it
genuinely does matter if (as is so often the case) they are palpably false.
Another version of that sort of argument might be offered to vindicate
Trump’s lies, after a fashion. That version is based on the notion of ‘truthful
hyperbole’—a non-sequitur that Trump (or his ghostwriter) coined in his book
The Art of the Deal.113 ‘The . . . key to the way I promote’, Trump writes, ‘is bra-
vado. I play to people’s fantasies. . . . People want to believe that something is the
biggest and the greatest and the most spectacular. I call it truthful hyperbole. It’s
an innocent form of exaggeration. . . . ’114 As Aristotle says in the Poetics, when
appealing to ‘the vulgar’ it is necessary for a person to ‘overact his parts’.115
107 Rich 2016. 108 As did, famously, Lincoln before him (Masur 2012).

109 As in the case of a story he told to a 1983 meeting of Congressional Medal of Honor win-
ners, of a World War II pilot who remained in his crippled aeroplane as it crashed rather than
letting his injured gunner die alone. It was in fact the storyline of the 1944 film A Wing and a
Prayer. When ‘asked if anyone bothered to check the accuracy of accounts presented as factual in
presidential speeches’, Reagan’s press secretary replied: ‘If you tell the same story five times, it’s
true’ (Cannon 1991, p. 39–40).
110 As in the Oxford English Dictionary definition.
111 Hanska 2012. For an insightful appreciation of the role that ‘fanciful imagination’ might
play in ‘the growth of empirical knowledge’ see Novitz (1980). See Wilson and Sperber (2012) on
the related phenomenon of ‘loose uses of language’.
112 Among them, ‘Reagan’s stories were often about the distant past or unspecified people illus-
trating themes he wanted to stress. Trump’s tweets, by contrast, often include falsehoods about
recent, clearly specified events. And, unlike Trump, the more upbeat Reagan didn’t use his words
as retribution for personal slights’ (Decker 2016).
113 Mayer 2016. On the interpretation of hyperbole see Wilson and Sperber (2012, pp. 50–1).
114 Quoted in Lozada 2015. 115 Aristotle 1965, 1461b–1462a.
At the traditional Harvard post-mortem on the 2016 US Election, Trump’s

former campaign manager Corey Lewandowski explained that ‘the problem
with the media’ is that ‘you guys took everything that Donald Trump said so
literally. The American people didn’t. They understood it.’116 As another com-
mentator elaborates:
When Donald Trump says he wants to build a huge wall, the media and his critics
seem to think he is imagining something like the Great Wall of China stretching
from Tijuana to Brownsville. But Trump’s supporters interpret his words differ-
ently. They hear him saying that he’s going to take a hard-line approach to border
security and illegal immigration. He’s not going to mess around. So when his sup-
porters hear him walking it back a bit—for instance, saying it could be a fence not
a wall at places—they knew what he meant all along. They understood he was
speaking figuratively about the wall.
When he talks about ripping up trade deals, he’s not saying that he is going to shred
the 741-page North American Free Trade Agreement and 348 pages of annexes.
His supporters take him to mean that he is going to take a much tougher approach
to NAFTA and other trade deals, that he is going to enforce trade agreements
much more rigorously, and that the US wasn’t going to be a chump any longer.117
As one commentator pithily puts it, ‘The press takes him literally, but not ser-
iously; his supporters take him seriously, but not literally.’118 Trump’s supporters
understand that he may not do everything he says, but they think he will none-
theless ‘try to stay in the spirit of the original statement’.119 Reinforcing that
message, Trump pledged to his legions in his unprecedentedly partisan inaug-
ural address, ‘I will never let you down.’120
The trouble with ‘truthful hyperbole’ from an epistemic point of view, of
course, is that one can never know what is the ‘truthful’ bit and what is the ‘hyper-
bole’. Does the defence of ‘truthful hyperbole’ as a form of ‘true fiction’ amount to
a claim that the direction of the vector is as described, and only its length is exag-
gerated? Or is the claim that its length is as stated, but the direction might be a
little off? Or is the suggestion that both might be off? And in all cases, by how
much? It is anyone’s guess—and when the ‘truthful hyperboles’ are surrounded
by a tissue of other lies, one’s best guess could well be worse than random.
21.7.3 Actions, Not Words, Are What Matter
A final version of the epistemic insouciance analysis might build on the ‘trust
me’ motif discussed above. In a catchphrase associated with Richard Nixon,
a previous US president driven from office for duplicity, ‘Watch what we do, not
116 Tumulty and Rucker 2016. 117

Cook 2016. 118 Merriman 2016; Zito 2016.
119 Cook 2016. 120 Trump 2017a.
what we say.’121 It is a phrase that Trump himself obliquely invoked during

the campaign in order to deliberately contrast himself with ‘politicians who are
all talk and no action’.122 He set about earning the ‘Man of Action’ sobriquet
conferred on him by the Speaker of the House of Representatives after their
first post-election meeting by signing a flurry of Executive Orders during his
first days in office.123
Something like that might also have been at work behind the Brexit campaign.
After all, referendum voters had no way of knowing what exact terms, if any,
might be negotiated as the terms of divorce between the UK and the EU.124
Referendum voters might have perfectly reasonably discounted the propositional
content of Leave campaigners (on the grounds that ‘they had to say that’) and
instead have been trusting what its leaders would actually do, once the referen-
dum campaign was over and their victory won.
A similar story might be told about at least some of Trump’s backers. Here is
one telling piece of evidence. The Kaiser Foundation convened post-election
focus groups involving Trump voters in Rust Belt states to discuss their views
on healthcare plans. Participants were initially asked what they disliked about
Obamacare and what they wanted to see in any replacement plans. Then con-
versations turned to actual Republican proposals for replacing Obamacare.
When told Mr. Trump might embrace a plan that included these elements [of
which they disapproved], and particularly very high deductibles, they expressed
disbelief. They were also worried about what they called ‘chaos’ if there was a gap
between repealing and replacing Obamacare. But most did not think that, as one
participant put it, ‘a smart businessman like Trump would let that happen’.125
Much of Trump’s rhetoric had a ‘trust me’ character to it, and clearly many of his
voters did.
Asked to explain her vote for Trump the day after the election, one of his
supporters said, ‘My vote was my only way to say: I am here and I count.’126
121 The actual words were those of John Mitchell, Nixon’s law partner and later his Attorney
General (Safire 1988).
122 Trump 2017a.
123 Trump 2016b. S. Jones 2017. Much of that action was more symbolic than real, at least in
the first instance, insofar as many of those Executive Orders require the action of others in order
to be implemented, and it was far from certain that that would be forthcoming (Parker and
Sullivan 2017).
124 Furthermore, as we said in Section 21.3, some might have voted strategically to Leave
merely to strengthen the UK’s hand in negotiations to remain in or anyway affiliated with the EU.
125 Altman 2017.
126 Diana Maus (61, Suffern, NY) in Kelly 2016; Garrison Keillor (2016a) quipped in reply,
‘People who shoot up theaters may feel the same way.’ À propos the Brexit referendum, a sixty-
two-year-old London jobseeker (Martin Parker, quoted in Ryan 2016) explained that he ordinarily
would not have voted:
I couldn’t really care less about the EU. [But] people are sick and tired of being ignored. I don’t
suppose I’m the only one to use this opportunity. It was a chance to kick the whole establishment
where it hurt, for us to send pain the other way. And we took it.
In both his speech accepting the Republican nomination and in his inaugural
address, Trump appealed to that old Roosevelt–Nixon trope, ‘the forgotten
men and women’, promising that they ‘will be forgotten no longer’.127 But what
exactly he would do, having remembered them, was always left pretty radically
unspecified.128 His appeal was always substantially, ‘trust me!’.129
In a pre-inauguration interview Kellyanne Conway, Trump’s former cam-
paign manager and counsellor designate, urged people to think about Trump
in just that way. ‘Why is everything taken at face value?’ she asked. ‘You always
want to go by what’s come out of his mouth rather than look at what’s in his
heart.’130 The interviewer rudely but rightly interjected, ‘How do I know what’s
in his heart except by what comes out of his mouth?’ But Trump’s supporters
think they know.131
Again, there is a rational gloss that could be put on that sort of claim. After all,
candidates always promise many things, but as president they inevitably have to
face situations no one could have anticipated during the campaign. Therefore, it
is only sensible for voters to assess candidates as much, or more, on their ‘char-
acter’ as on their specific policy proposals. Of course their assessment of a can-
didate’s ‘character’ is adduced, in no small part, from what specific policy
proposals that candidate makes during the campaign. But the voters’ real task is
to elect a person whom they can trust to do what they would have wanted in
circumstances neither they nor the candidate could have foreseen.132
Of course, in the case of Trump the standard political science term ‘charac-
ter’ is unfortunate, since he is a man of reprehensible character (as even many
of his most ardent supporters might concede).133 But on the analysis just
127 Trump 2017a. Schrag 1969.

128 In his RNC acceptance speech Trump (2016a) said, ‘I have visited the laid-off factory work-
ers, and the communities crushed by our horrible and unfair trade deals. These are the forgotten
men and women of our country.’ He said he would strike better trade deals and bring manufactur-
ing jobs back to America. But how, exactly, would he do that? That is what is always left awfully
vague.
129 As Joe Lockhart, President Clinton’s former press secretary, observed, Trump’s message in
every post-election interview was, ‘People out there, trust me. Don’t trust what you read or you
see’ (in NBC Meet the Press 2017).
130 Blake 2017a.
131 As one woman at a Trump rally replied when asked about some of his more questionable
statements: ‘words come out in the wrong way at times; you put your foot in your mouth’ (quoted
in Pilkington 2016)
132 Barber 1972. Hardy 2017.
133 As Swaim (2017) has it:
[T]here is something brutally, refreshingly realistic about Trump’s manner, or about the
whole Trump persona. He is a deeply flawed man, but he doesn’t try very hard to pretend
otherwise. Even his most enthusiastic supporters, or many of the ones I’ve talked to, are
happy to acknowledge Trump’s failings. . . . [T]hey did not vote for him because they thought
him scrupulously honest or because they believed his character to be unimpeachable.
Indeed, there must be very few people on either side who believe Trump to be a thoroughly
good man. Effective in his way, maybe. Capable of disrupting what ought to be disrupted,
almost certainly. But good?
offered, ‘character’ is anyway something of a misnomer. It is not a question of

whether the candidate is a good Boy Scout, or even someone you would seat next
to your daughter. ‘Character’, on the analysis offered above, is really much less
moralistic than that, and much more just a matter of ‘political dispositions’—
how he is likely to react in unanticipated political circumstances.
Trump’s voters thought that they knew the answer to that. We have our
doubts, given the scarcity of specifics in Trump’s election campaign promises,
the history of his firms’ bankruptcies and other broken promises. We have
similar doubts whether Brexit voters had any good grounds for any beliefs
whatsoever about what form Brexit might take or for trusting politicians to
negotiate the deal that they themselves would have preferred. But no matter.
What we are trying to do here is merely to explain why some voters might have
fallen for Trump and Brexit, and what it might mean for the CJT if that were
indeed the true explanation of their victories.
21.8 EVERYONE ON FACEB O OK AGREES WITH ME
There are many ways in which the truth might come under threat. Some of
them are politically innocent. Others are more politically charged.
First, let us consider some more innocuous versions of the story, based
purely on natural tendencies at work within the new media environment upon
which people increasingly depend for their news. Perhaps it was like that all
along in some places (Britain with its tabloid press, for example); perhaps it was
like that in other eras.134 But in mid-twentieth-century America, anyway, everyone
tended to get their news primarily from the same handful of broadcast and print
media, which by and large held to high standards of neutrality and impartiality;
their reports were authoritative, and generally taken to be such by the population
at large.135 With the rise of the Internet and especially of platforms, however,
that is decreasingly true.136
134 Newton and Brynin 2001. Francis Bacon commented similarly in 1620 about his generation’s
equivalent of ‘information bubbles’ in Novum Organum (Floridi 2016).
135 Thus, after Walter Cronkite’s 1968 post-Tet broadcast saying that he thought the
Vietnamese war would not be won, President Johnson said to his aides, ‘If I’ve lost Cronkite I’ve
lost middle America’ (Martin 2009). Cross-national studies continue to show that exposure to
public service broadcasting increases citizens’ knowledge of current affairs, compared to exposure
to commercial broadcasting (Soroka et al. 2013).
136 Berry and Sobieraj 2011. Indeed, as Silverman (2016) observes,
In the final three months of the US presidential campaign, the top-performing fake election
news stories on Facebook generated more engagement than the top stories from major
news outlets such as the New York Times, Washington Post, Huffington Post, NBC News,
and others. . . . During these critical months of the campaign, 20 top-performing false
election stories from hoax sites and hyperpartisan blogs generated 8,711,000 shares, reactions,
People now get much of their ‘news’ from more boutique sources, tailored
to their own particular interests and perspectives. To some extent they do so
deliberately. Nicholas Negroponte, and Cass Sunstein following him, entered a
prescient warning that future internet users would be able to construct their
very own personalized news feed, ‘Daily Me’, that told them only what they were
interested in and wanted to hear.137 Today, search engine and social network
algorithms (more of which below) do that for you. But people still deliberately
choose, in similar fashion, whom to include as their Facebook friends and
which Twitter feeds to follow.138
In part, it is merely a matter of people’s ‘likes’. You can obviously ‘like’ some-
thing (find it interesting or amusing) without believing for a moment that it is
true. But self-sorting based on ‘likes’ sometimes has an epistemic side to it as
well. People are not unreasonably inclined to give more credibility to reports
coming from people they deem to be trustworthy, understood as believing
other things that they themselves also believe to be true.139
Another driver of that phenomenon is purely commercial. The algorithms
underpinning Google and other search engines are designed to show people
web pages that are similar in relevant respects to those that they have previ-
ously viewed—in no small part in the hopes someone who has purchased
something from a previous website will be tempted to make similar pur-
chases from subsequent ones as well, or at least stay on similar pages to
see more adverts controlled by the engine or network. Commercially, that
makes perfectly good sense. And from the point of view of the customer—or
even those who are just using the search engine to find related material,
with no intention of buying anything—that feature of the search engine is
genuinely to be welcomed.
From an epistemic point of view, however, those search engine algorithms
are a disaster. They create information ‘bubbles’, in which a person perpetually
gets fed new information that reinforces the information he initially received,
and comments on Facebook. Within the same time period, the 20 best-performing election
stories from 19 major news websites generated a total of 7,367,000 shares, reactions, and
comments on Facebook.
According to its preamble, this is what motivated a bill introduced into the California state
legislature immediately after the 2016 US election to require schoolchildren be taught how to
recognize fake news (Dodd 2017).
137 Negroponte 1995. Sunstein 2001, pp. 3–23; 2017a. Jamieson and Cappella 2008. Lelkes,
Sood, and Iyengar 2017.
138 And, at least on some evidence (Bakshy et al. 2015), that reduces exposure to differing
political perspectives even more than the operation of algorithms alone—although still far from
eliminating it completely.
139 This is a variation on Hume’s (1777) argument ‘On Miracles’: if someone tells you he
just saw someone walking on water, do you upgrade your belief in miracles or downgrade your
estimation of that person’s credibility as a reporter of true facts?
however idiosyncratic and unrepresentative the original bit of information.140

People who once searched for information on conspiracy theories keep getting
fed more and more conspiracy theories, and so on. ‘The net result’, as the inventor
of the Internet Tim Berners-Lee observes, ‘is that these sites show us content they
think we’ll click on—meaning that misinformation, or fake news, which is sur-
prising, shocking, or designed to appeal to our biases can spread like wildfire.’141
The strength of those algorithms is well captured by this anecdote from
Internet activist Tom Steinberg, posted on Facebook immediately after the
Brexit referendum result was announced:
I am actively searching through Facebook for people celebrating the Brexit leave
victory, but the filter bubble is SO strong, and extends SO far into things like
Facebook’s custom search that I can’t find anyone who is happy *despite the fact
that over half the country is clearly jubilant today* and despite the fact that I’m
*actively* looking to hear what they are saying.142
There is another driver of that phenomenon which, while not exactly ‘innocent’,
is at least not politically motivated. People are rewarded more, both psychologic-
ally and financially, the more other people who click on their websites or share
their Internet postings. That incentivizes people to post fabulous, sensational,
incredible stories—whether true or not—purely as ‘clickbait’.143
Such clickbaiters are pure ‘bullshitters’, in Harry Frankfurt’s sense. Unlike
others (who will be the subject of Section 21.9), these clickbaiters do not delib-
erately post stories that they know to be false—certainly anyway they do not
deliberately post them because of their known falsehood. Instead, those click-
baiters simply do not care about the truth of their posts, one way or another.144
But in not caring, of course, they end up posting a good deal of information
that is patently false.145
140 Pariser (2011) coined the term ‘filter bubble’, but nowadays the term is rife. See, for example,
Ash (2016) and more generally O’Neil (2016). For a rich empirical analysis of how a ‘Breitbart-led
right-wing media ecosystem’ created an almost hermetically sealed bubble for Trump supporters
during the 2016 US presidential election, see Benkler et al. (2017).
141 Quoted in Solon 2017. For a sustained analysis of how homogeneous clusters of users (‘echo
chambers’) facilitate the spread of misinformation on the Internet, see Vicario et al. (2016).
142 Steinberg 2016.
143 Ohlheiser 2016; Tynan 2016 . Stories abound. One is of two unemployed restaurant work-
ers who signed on as writers to the alt-right website LibertyWritersNews with 300,000 Facebook
followers in the month before the 2016 election, and who ‘say they are making so much money
that they feel uncomfortable talking about it because they don’t want people to start asking for
loans’ (McCoy 2016). Another is of the small Macedonian town of Veles, home to ‘more than 150
domains’ dedicated to generating fake news for profit.
144 As Neetzan Zimmerman, a sometime Gawker specialist in viral stories, says, ‘Nowadays it
is not important if a story’s real. The only thing that really matters is whether people click on it. If
a person is not sharing a news story, it is, at its core, not news’ (quoted in Viner 2016).
145 One prominent hoaxer, Paul Horner, posted false news hoping it would get picked up by
Trump supporters, exposed and then make then look bad. Needless to say, that backfired. As he
explained (quoted in Dewey 2016):
Shortly after the 2016 US presidential election, Google and Facebook

announced they were taking steps to ban fake news sites and deprive them
of advertising revenue.146 If successful, such steps may help ameliorate this
particular part of the problem for future elections. Obviously, however, the
damage they did in the 2016 elections is already done. And laudatory though it
may be for Facebook to flag that some post is ‘Disputed by 3rd party factcheckers’,
‘the damage of a popular fake-news story is usually well done by the time it is
fact-checked and flagged’.147
In consequence of all these factors, people experience ‘alternative realities’
on the Internet. What one person reasonably believes, given the information
that he obtains there, can be radically different to what another person with a
different Internet experience might equally reasonably believe. As President
Obama once quipped, ‘If I watched Fox I wouldn’t vote for me!’148 In his fare-
well address, Obama bemoaned the creation of these ‘alternative realities’ in
the following terms:
[I]ncreasingly we become so secure in our bubbles that we start accepting only
information, whether it’s true or not, that fits our opinions, instead of basing our
opinions on the evidence that is out there. . . . In the course of a healthy debate, we
[rightly] prioritize different goals, and the different means of reaching them. But
without some common baseline of facts, without a willingness to admit new infor-
mation and concede that your opponent might be making a fair point, and that
science and reason matter, . . . we’re going to keep talking past each other.149
This is indeed an unfortunate outcome. But, as we have here seen, there are
some relatively innocent reasons that it might have occurred. There are also,
however, some much less innocent drivers, to which we now turn.
I just wanted to make fun of that insane belief, but it took off. They actually believed it.
I thought they’d fact-check it, and it’d make them look worse. . . . [T]hat’s how this always
works: someone posts something I write, then they find out it’s false, then they look like
idiots. But Trump supporters—they just keep running with it! They never fact-check
anything! Now he’s in the White House. Looking back, instead of hurting the campaign,
I think I helped it. And that feels [bad].
146 Isaac 2016. Wingfield et al. 2016. Naughton 2017. Persily 2017, pp. 72–5. Weedon, Nuland,
and Stamos 2017.
147 Persily 2017, p. 73. That is particularly likely because of the slow process by which Facebook
refers items for fact-checking (Jamieson and Solon 2016):
Facebook is working with five fact-checking organizations – ABC News, AP,
FactCheck.org, Politifact and Snopes – to launch the initiative. If enough of Facebook’s
users report a story as fake, the social network will pass it onto these third parties to
scrutinize. If a story is deemed to fail the fact check, it will be publicly flagged as ‘dis-
puted by 3rd party fact-checkers’ whenever it appears on the social network.
148 Obama 2016.
149 He went on to say that the ‘selective sorting of the facts . . . is self-defeating because, as
my mom used to tell me, reality has a way of catching up with you’; one can but hope it is so
(Obama 2017; see similarly Obama 2016).
21.9 EPISTEMIC MALEVOLENCE
There are various more politically charged ways in which the truth can come
under threat. One familiar way is through ‘epistemic populism’—political leaders
telling people to ignore reliable sources of information and to trust their own
instincts.150 Another familiar way is through ‘epistemic authoritarianism’—a
political leader telling people, after the fashion of Orwell’s Nineteen Eighty-Four,
that the truth just is whatever he or she says it is.151 Here we shall concentrate
on a third way, which we call ‘epistemic malevolence’.152 The malevolence here
in view is instead targeted at facts as such—it is the aim to prevent true facts
from emerging at all, or to prevent them from getting widespread currency if
they do.
On the first full day of his presidency, Trump went to the CIA and gave a
rambling address that was highly inappropriate in ever so many ways. Among
other things, he claimed in it that more people attended his inauguration than
any other—which was blatantly untrue, as was evident from photographs that
had already been published offering a side-by-side comparison of his inaugural
audience and Obama’s first.153 Later that day, his press secretary went down to
the White House pressroom and reiterated that lie in no uncertain terms.154
The next day, Trump’s former campaign manager and new White House
Counselor, Kellyanne Conway, was asked: ‘Why put [the Press Secretary] out
there for the very first time . . . to utter a provable falsehood’ about the size of the
crowd at the inauguration? ‘It’s a small thing. But the first time he confronts
150 That is what was involved when Trump told people to ignore the mainstream news media
and to trust whatever they find on the Internet (Borchers 2016; Swan 2016). That pattern appears
in a pre-Brexit referendum interview with Michael Gove: the interviewer challenged Gove to
defend his advocacy of Leaving the EU when so many economists, business and labour leaders,
and even the Chief Executive of the NHS ‘all say that you . . . are wrong’; Gove’s reply was that
‘I think the people of this country have had enough of experts’ (quoted in Islam 2016b). The British
Election Study’s (2016) post-referendum survey showed that Gove’s reading of the mood of the
electorate, or at least of Leave voters, was indeed correct: the probability of voting to Leave was
strongly associated with agreement to the proposition, ‘I’d rather put my trust in the wisdom of
ordinary people than the opinions of experts.’
151 Putin today (Kovalev 2017), and Stalin before him (Arendt 1967/1977), are perhaps the
clearest exemplars. But Trump’s insistently reasserting claims that have been fact-checked and
shown to be clearly false smacks of that. So too does the comment of his former campaign man-
ager and counselor designate, Kellyanne Conway, when replying to a question at the Harvard
post-election conference about whether Trump’s behaviour really is ‘presidential behaviour’: ‘He’s
the president elect so that’s presidential behavior’ (quoted in Sullivan 2016). But perhaps former
New York Times editor Bill Keller is right to suggest that the most chilling evidence is to be found
in Trump’s bare-faced lying in his January 2017 speech at the CIA: ‘He was spouting obvious
falsehoods to an audience for whom facts are matters of life and death. The implicit, and truly
dangerous, message to the intelligence community was “don’t bring me bad news; just tell me
we’re winning”’ (quoted in Farhi 2017b).
152 Baehr 2010. Cassam 2016. 153 Trump 2017c. Rucker et al. 2017.
154 Kessler 2017b.
the public it’s a falsehood?’ Conway offered this memorable reply: ‘You’re

saying it’s a falsehood. . . . [O]ur press secretary gave alternative facts to that.’
The incredulous interviewer rightly pressed her on that: ‘Wait a minute.
Alternative facts? . . . Four of the five facts he uttered were just not true. Look,
alternative facts are not facts. They are falsehoods.’155 And clearly they were:
the claims of the president and press secretary were contradicted by the photo-
graphs;156 they were contradicted by official Metrorail ridership statistics;157
and so on.
Why on earth would any president deliberately engage in such behaviour, in
his very first day in office?158 With Trump, who knows? It might be that a fragile
nouveau riche ego, already deeply suspecting it is somewhere it does not belong,
simply cannot bear the thought of being associated with anything demeaned
as ‘small’.159 Or maybe Trump really believes his own lies—maybe his grip on
reality truly is just that infirm.160 Or maybe he is using his preposterous tweets
as a smokescreen to distract from the many nefarious policies put in place
through executive orders signed on the same days.161 Or perhaps Trump is just
deploying the political equivalent of ‘an old sports strategy: foul so much in the
first 5 minutes of the game that the refs can’t call them all. From then on, [you’re
free to play] a more physical game.’162
Another far more nefarious explanation is also consistent with much that
Trump has done and said, however. That explanation certainly seems to fit the
intentions of many of Trump’s protégées and backers. Conspicuous among
them is Stephen Bannon, Trump’s campaign manager who became for a time
155 Sinderbrand 2017. Sean Spicer, the White House press secretary, made a similar
Freudian slip in a news conference, saying, ‘I think sometimes we can disagree with the facts’;
but it seems clear from context that what he really meant to say was that we can disagree about
the facts (Blake 2017c). It is far less clear from context that what Conway really meant to say
was ‘additional’ (rather than ‘alternative’) facts, as she subsequently rather disingenuously
claimed (Pengelly 2017).
156 Kessler 2017b.
157 The press secretary admitted as much in his first formal news conference—the first
one in which he actually took questions rather than merely having a rant—two days later
(Blake 2017c).
158 Cowen (2017) offers yet another speculation:
By requiring subordinates to speak untruths, a leader can undercut their independent stand-
ing, including their standing with the public, with the media and with other members of the
administration. That makes those individuals grow more dependent on the leader and less
likely to mount independent rebellions against the structure of command.
159 Amis 2016.
160 Rubin 2017. His first television interview post-inauguration, imploring the interviewer to
examine all his framed photos of his inauguration crowd, certainly sounded like Nixon at his most
needful petitioning Kissinger to join him on his knees in prayer (Johnson 2017). Freeman (2017)
comments similarly on Trump’s ‘neediness’, jocularly via a commentary on the length of his
neckties.
161 Balz 2017. Dionne 2017. 162 Sally Jenkins, quoted in Cillizza 2017.
Chief Strategist in his White House, who had previously been chief executive
of the alt-right ‘news’ site Breitbart. Also included among ‘Trump’s backers’
deploying this strategy are Russian officials and agents, if the US Director of
National Intelligence’s report is to be trusted.163
The strategy in question involves the intentional promulgation of false
stories, knowing them to be false, and doing so precisely because you know
them to be false.164 If people actually believe the false story in support of your
preferred position, so much the better.165 But the larger aim of promulgating
fake news is independent of people actually believing it. The aim is instead
simply to discredit, in the eyes of your followers, all sources of information,
true or false. As one commentator puts it, ‘It’s not an information war. It’s a war
on information.’166
This is a strategy that has been pursued by certain fragments of the American
right for some time, and with devastating effect on public trust in the main-
stream media.167 The strategy was taken to even further extremes by Breitbart
163 US Director of National Intelligence 2017.

164 For just one example, from literally thousands, consider the ‘fake news masterpiece’
concocted by Cameron Harris, since fired from his position as an aide to a Republican state
legislator in Maryland. During the autumn of 2016 when Trump was behind in the polls and
preparing his supporters for defeat by asserting the election was being rigged, Harris concocted
the story, ‘Tens of thousands of fraudulent Clinton votes found in Ohio warehouse’. Harris
(quoted in Shane 2017) was himself surprised by the success of the story that netted him $1,000
per hour he invested in it:
Given the severe distrust of the media among Trump supporters, anything that parroted
Trump’s talking points people would click. Trump was saying ‘rigged election, rigged
election’. People were predisposed to believe. . . . At first it kind of shocked me—the
response I was getting. How easily people would believe it. It was almost like a sociological
experiment.
165 Allcott and Gentzkow (2017) find that ‘of the known false news stories that appeared in
the three months before the election, those favoring Trump were shared a total of 30 million
times on Facebook, while those favoring Clinton were shared eight million times’, and that ‘the
average American saw and remembered 0.92 pro-Trump fake news stories and 0.23 pro-Clinton
fake news stories, with just over half of those who recalled seeing fake news stories believing
them’. They attempt to minimize the impact of that fake news, however, by saying, ‘for fake news
to have changed the outcome of the election, a single fake article would need to have had the
same persuasive effect as 36 television campaign ads’. But that may not be as implausible as it
sounds. People notoriously discount paid political advertisements. And, in any case, they see an
awful lot of them. The same source upon which Allcott and Gentzkow (2017) base their calculation
also reports that, on average, each person in the study saw 75 advertisements during a presidential
election campaign and ‘in some areas . . . , voting-aged adults see as many as 339 spots’ (Spenkuch
and Toniatti 2016, p. 11).
166 Peter Pomerantsev, quoted in Ignatius 2016.
167 Jamieson and Cappella 2008. Gallup have a series of polls asking Americans, ‘How much
trust and confidence do you have in the mass media – such as newspapers, TV and radio – when
it comes to reporting the news fully, accurately and fairly?’ The proportion of respondents saying
‘a great deal’ or ‘a fair amount’ ranged in the low-to-mid 50% range until 2004, at which point it
dropped to 44%; it has trended downwards since then, dipping particularly in election years,
standing at just 40% in 2015 (Riffkin 2015).
‘News’ under the leadership of Stephen Bannon.168 Trump powerfully associated

himself with that strategy, not only during the campaign but also as president,
when declaring—in the first official non-ceremonial fixture of his presidency,
an address to the CIA no less—that ‘I have a running war with the media. They
are among the most dishonest human beings on Earth.’169
No doubt in part this is an attempt at muzzling the press. In an interview
with the New York Times the very first week of the Trump Administration,
Bannon declared—telling the interviewer ‘I want you to quote this’—that the
media is ‘the opposition party’. He added, ‘the media should . . . keep its mouth
shut and just listen for a while’.170 In an interview the next day the president
himself endorsed Bannon’s sentiments.171
But the success of the larger strategy does not depend in any way upon the media
itself going silent. Nor does it depend upon succeeding in persuading people to get
their news from the Internet, which is much more of a hotbed of Trump-friendly
fake news, rather than the mainstream media (although Trump encourages
that, too172). The larger strategy is simply to instill widespread distrust in all sources
of information—including the evidence of one’s own eyes (as in the case of the
side-by-side photos of the crowds at Trump’s and Obama’s inaugurations).173
From the perspective of this strategy, that is the real point of generating false
news, as was so widely done by various agents throughout the 2016 US election
campaign and around the world, by Russian agents among many others.174 The
aim of these agents is not so much to persuade people to believe them (although
perhaps so much the better, from the point of view of the purveyors, if people
do) as to dull people’s sensitivity to truth in any form.175 It is a strategy that, in
168 As the former spokesperson for Breitbart, Kurt Bardella, said in interview, ‘There is no
question that Trump’s confrontational and combative tone towards the media is choreographed
by Bannon. It’s textbook Breitbart. If the facts aren’t on your side, attack the gatekeepers of the
facts. . . . From Team Trump’s perspective, . . . their objective will be to cast as much doubt as pos-
sible on traditional sources of information to ensure the environment is ripe for them to win in
2020’ (quoted in Farhi 2017b).
169 Trump (2017b), echoing rhetoric throughout his campaign (Baron 2016).
170 Grynbaum 2017b.
171 Wagner 2017. Trump followed up on that, tweeting: ‘Somebody with aptitude and convic-
tion should buy the FAKE NEWS and failing @nytimes and either run it correctly or let it fold
with dignity’ (Farhi 2017a).
172 Trump constantly told his rallies, ‘Forget the press, read the internet. . . . I . . . get a lot of hon-
esty over the internet. . . . Study over things. Don’t go for the mainstream media’ (Borchers 2016).
173 This particular variant on the strategy has come to be known as ‘gaslighting’, after the 1938
play and later movie of the same name (Gibson 2017).
174 Connolly et al. 2016. Reuters 2017.
175 According to a RAND Corporation analysis, that is how the current Russian propaganda
model works: ‘either through more direct persuasion and influence or by engaging in obfuscation,
confusion and the disruption or diminution of truthful reporting and messaging’ (Paul and
Matthews 2016, pp. 1–2). ‘They’re not trying to say that their version of events is the true one.
They’re saying: “Everybody’s lying! Nobody’s telling you the truth!” ’ (Richard Stengel, sometime
managing editor of Time magazine and US Under Secretary of State for Public Diplomacy and
Public Affairs, quoted in Ignatius 2016).
US politics, dates back at least to George W. Bush’s White House and its public
relations guru, Karl Rove.176
What James Fallows calls ‘the chaos-generating logic of Trump’s seemingly
illogical stream of nonstop lies big and small’ can be traced, more recently,
‘to reality TV, to Breitbart and Steve Bannon, and to Vladimir Putin’s advisor
Vladislav Surkov’.177 The latter’s strategy is particularly instructive:
[O]riginally from the avant-garde art world . . . , [w]hat Surkov has done is to import
ideas from conceptual art into the very heart of politics. His aim is to undermine
peoples’ perceptions of the world, so they never know what is really happening.
Surkov turned Russian politics into a bewildering, constantly changing piece of
theater. . . . [N]o one was sure what was real or fake. As one journalist put it: ‘It is
a strategy of power that keeps any opposition constantly confused.’ [He creates] a
ceaseless shape-shifting that is unstoppable because it is undefinable.178
That is of a cloth with Trump’s media strategy: the combination of empty
spectacle, empty words, discrediting everyone, crediting conspiracy theories
without any evidence, hogging attention and ‘gaslighting’ makes it hard for
people to know what, if anything, to believe to be true.179
The clearest expression of this attitude came in a post-election panel discus-
sion on NPR in which CNN commentator and ardent Trump advocate Scottie
Nell Hughes famously said: ‘facts, they’re not really facts. . . . There’s no such
thing . . . anymore [as] facts’.180 Although she subsequently claimed she had
misspoken, the Atlantic’s James Fallows rightly replied,
I think it actually is an intended result of this campaign and administration to
[make people] think, well, really there aren’t any facts, it’s all opinion. . . . I believe
that the job for the media and civil society now is essentially to say there are such
things as facts. So the line may be drawn here.181
176 Whom Fallows (2016a) assumes to be the unnamed ‘senior advisor to Bush’ who belittled
‘what we call the reality-based community’, saying, ‘That’s not the way the world really works
anymore’ (Suskind 2004). He continued:
We’re an empire now, and . . . we create our own reality. And while you’re studying that
reality – judiciously, as you will – we’ll act again, creating other new realities, which you can
study too, and that’s how things will sort out. We're history’s actors . . . and you, all of you, will
be left to just study what we do.
177 Fallows 2016a. 178 Curtis 2014.
179 Yuhas 2017. 180 Quoted in Rehm et al. (2016b).
181 Fallows in Rehm et al. (2016b). Later in the discussion Hughes backed off the ‘no facts’
claim to say that ‘any facts that they might be able to report nobody believes because [the reporter
has] interlaced his opinion’ in his reports of the facts. She gives an example: ‘look at reports . . . [in
outlets] like Daily Caller, Breitbart, Washington Times, . . . and you will say those are not facts. Well
guess what? It’s a two-sided mirror because they say the same about your reporting.’ And in a
subsequent follow-up she added (in Stelter 2016):
My comment was that if you were a Trump supporter, you believed the words his campaign
was saying were fact. If you were a Clinton supporter, you believed the words her campaign
were stating was fact. However, both sides did not believe nor acknowledge the other as fact.
Just like in a court of law where both sides honestly believe they are right. When a prosecutor
The larger strategy is to ‘disempower institutions that protect the truth’.182

President Trump took to Twitter to declare, ‘The FAKE NEWS media . . . is not
my enemy, it is the enemy of the American people’;183 and among those who
had voted for him, 88 per cent agreed.184 Retired General Michael Hayden,
one-time head of both the CIA and NSA, describes this as ‘a systematic effort to
invalidate and delegitimize all the institutions, governmental and nongovern-
mental, that create the factual basis for action . . . so they won’t push back against
arbitrary moves’.185
Part and parcel of that strategy is for its practitioners to appropriate the term
‘fake news’ themselves and to apply it indiscriminately to any reports that they
do not like, be they true or false.186 Thus, for example, when CNN reported that
President-Elect Trump had been briefed by the US intelligence community
that Russia had assembled a dossier of embarrassing material that might render
him vulnerable to blackmail, Trump responded by accusing CNN of prom-
ulgating ‘fake news’.187 While it may well be the case that the contents of the
dossier are not true (the CNN made no claim that they were), there was noth-
ing remotely ‘fake’ about the news report of the indisputable fact that Trump
had received just such a briefing. For Trump to say otherwise simply devalues
the language—which is part of this strategy. Once inaugurated, Trump doubled
down on his insistence that anything unflattering to him appearing in the
mainstream media was ‘fake news’.188
A final element of the strategy of undermining truth claims altogether is for
purveyors of lies to insist that others must accord equal epistemic respect to
their fabrications as to the genuine evidence. The mainstream media’s tradi-
tions of impartiality and ‘equal time’ serve us well epistemically when everyone
comes in he states his ‘facts’ of the case. The Defense Attorney does the same. It is up to the
jury to decide what is the truth. Of course I believe there are facts in this world; what I was
referencing, as I stated, was in regards to this campaign cycle. Facts to one side were seen as
opinion or untrue to the other.
182 Douglas 2017. And it works, at least according to the editor of one small-town Michigan
newspaper, who said, ‘You can give readers 50 facts that show that Trump is wrong, but when he
portrays us in the media industry as the bad guys, that seems to outweigh all those facts’ (Jeff
Payne quoted in Pilkington 2017).
183 Grynbaum 2017a. 184 Sargent 2017b. 185 Gersen 2017.
186 President Trump candidly tweeted, ‘Any negative polls are fake news’ (Marcus 2017). In
response to this sort of behaviour, the Washington Post (2017a) editorialized, ‘So overused and
misused is the phrase [“fake news”] – by those seeking to disparage things they simply dislike or
disagree with – that it loses real meaning.’ See similarly: Blake 2017b; Borchers 2017; Sullivan 2017b.
187 Trump 2017b. Nossel 2017. Wemple 2017a.
188 Sargent 2017a. Similarly in Sweden, a right-wing Facebook group, Mediakollen, emerged
pretending to be a fact checker but actually serving as ‘itself a tool of disinformation . . . , in effect,
a fake fact checker’ (Jackson 2017). Breitbart similarly posted partial truths and patent falsehoods in
ostensibly fact-checking the Guardian reporting about illegal immigration, which it disingenuously
described as ‘fake news’ (Carroll 2016).
honestly and honorably asserts only what they genuinely believe to be true. But
they serve us ill when people deliberately lie, asserting propositions they know
to be false for some strategic purpose unconnected to any quest for truth.189
There has been much debate within the mainstream media as to how best to
cover someone like Trump. First, many old-school editors insist that ‘more and
better of the same sort of journalism as always’—more and more fact-checking
and so on—would be the best response.190 But fact-checking of such trans-
parent falsehoods is a soul-destroying time-suck that diverts journalists from
investigations that might be of more consequence.191 In any case, there is a fair
bit of evidence that fact-checking pays decreasing dividends—once someone
has been caught in fifty whoppers, reports of a fifty-first (even if the substance
is such that, objectively, it really should be a very big deal) evokes very little
response from the public at large.192 Fact-checking might even backfire, insofar
as further reporting of the falsehood (if only to refute it) helps the falsehood
stick in people’s minds.193
Others advise, second, that we to ‘get out there with true facts first’ before
liars have a chance to spin their falsehoods.194 But liars are creative (who could
have imagined all the sorts of falsehoods Trump would come up with?), so it
seems impossibly hard to implement that strategy in such a way that would
forestall all successful political lies.
A third approach that has been mooted, but not seriously (or anyway sys-
tematically) attempted, would be for the press simply to boycott the White
House pressroom of a proven liar. But it would be hard to organize a successful
boycott among all the media actors in such a highly competitive environment.
And of course even if the mainstream media boycotted the White House press-
room, right-wing media like Fox News and Breitbart would remain.195
A fourth approach is to educate the public in how better to detect falsehoods.
Shortly after the 2016 election a bill was introduced into the California state legis-
lature along those lines, for example.196 The OECD’s director of education agrees
189 Patterson 2016. 190 Baron 2016. Voss 2016. Hiatt 2017. 191 Wemple 2017b.

192 Cook and Lewandowsky 2011; Nyhan and Reifler 2010; 2015; Harford 2017. Or worse:
Major Garrett of CBS News recounted how, during the 2016 presidential campaign,
Any fact checking I did . . . was prima facie evidence that I was biased. And that I was wrong.
So fact checking Trump was proof, not that he was wrong, but that he was right, and that
anyone who would raise a question about the underlying relationship between what he
said in the facts was biased. And therefore, [it can be] legitimately disregarded from the
beginning. So it wasn’t as if there was a conversation about this. It wasn’t as if facts were
litigated back and forth. The very raising of a question about the factual basis of a Trump
assertion was proof you were wrong and biased. And that is the atmosphere that I found
myself existing in as a reporter (Garrett in Rehm et al. 2016a).
193 Cook and Lewandowsky 2011, p. 2. 194 Paul and Matthews 2016, pp. 9–10.
195 Rosen, Wemple, and Downie 2017.
196 Bever 2017. Dodd 2017. The idea, if not the specific legislation, was endorsed by the CEO
of Apple, Tim Cook (Rawlinson 2017).
that ‘exposing fake news’ by helping students learn how to ‘distinguish . . . what

is true from what is not . . . is something that . . . schools can do something about’;
and Sweden has already instituted such a policy.197 In the US, the Washington
Post fact checker has provided an easy ‘guide for detecting fake news’.198
A fifth approach is to follow the lead of Germany and legislate to ‘compel large
outlets such as Facebook and Twitter to rapidly remove fake news that incites
hate, as well as other “criminal” content, or face fines as high as 50 million euros
($53 million)’.199
A sixth approach is simply to reset the default assumption of how to respond
to someone who has persistently been caught lying. Whereas we ordinarily
ought to assume that people are telling the truth unless we have evidence to
the contrary, once we have enough evidence that some particular person per-
sistently lies, we ought to assume that that person is lying unless evidence is
produced to the contrary.200
21.10 EPISTEMIC AGNOSTICISM
The success of the strategy of epistemic malevolence just discussed depends

crucially upon listeners mistaking epistemic saboteurs who are actually
strategically lying to them for genuine epistemic peers.
The right response, when confronted with conflicting reports on some
matter of fact that come from people whom you rightly regard as epistemic peers,
is often thought to be to ‘split the difference’ when you can or, when you cannot,
to ‘suspend judgement’ and treat the matter as an ‘open question’.201 When out
197 Andreas Schleicher, quoted in Siddique 2017. Priest and Birnbaum 2017.

198 Kessler 2016b. Sullivan 2017a.
199 Faiola and Kirchner 2017. See further Priest and Birnbaum 2017.
200 As one commentator says, ‘I don’t believe a word he says, and neither should you’ (Bernstein
2017; see similarly Fallows 2016a). Kellyanne Conway (quoted in Heim 2017) said in one interview,
We believe in a free and fair media, but with freedom comes responsibility. It would be great
for the media to be less presumptively negative and skeptical and more open and honest
about their past unfair and untoward coverage of [Trump]. . . . I was really astonished to see
respected print and electronic journalists outwardly admit during the campaign that Donald
Trump forces them to suspend the objective standards of journalism.
But on this analysis, through his own behaviour Trump has forfeited any claim (either moral or
epistemic) to be presumed to be a truth teller.
201 Such was the classical approach anyway, culminating perhaps with Sidgwick’s (1907, bk 3,
ch. 11, section 2, iv) Methods of Ethics:
[T]he denial by another of a proposition that I have affirmed has a tendency to impair my
confidence in its validity. . . . And it will be easily seen that the absence of such disagreement
must remain an indispensable negative condition of the certainty of our beliefs. For if I find
any of my judgements, intuitive or inferential, in direct conflict with a judgement of some
other mind, there must be error somewhere: and if I have no more reason to suspect error in
of a misplaced sense of fairness or impartiality, or respect for someone’s official

position, people treat reports from deliberate liars in that same way, they are
led—quite wrongly, from an epistemic point of view—to the same state of epi-
stemic agnosticism, treating as open questions matters of fact that are really
firmly settled.202
The tendency to take into account false claims from others, even against your
own better judgement, has various sources. Some are sociological and psycho-
logical.203 Some may even be neurophysiological. In fMRI studies, evidence has
been found to suggest that the neurophysiological mechanisms that ordinarily
inhibit lying weaken the more lies one tells.204 Extrapolating from those studies,
we might imagine that related neurophysiological mechanisms making us
resent being lied to weaken the more often we have been lied to.
Be all that as it may, there is clear evidence of deep scepticism among both
the US and UK electorates about any and all purported truth claims politicians
made during the 2016 campaigns. Nearly half of UK voters believed that both
sides were ‘mostly telling lies’ in the Brexit referendum campaign.205 Similarly
in the US election, a Washington Post correspondent’s analysis of why repeated
reports of Trump’s serial lying gained no traction among his supporters was
that those people were thinking, ‘So what if he doesn’t all the time tell the truth?
Politicians never do.’206
Right-wing activists have mounted a concerted effort, through pseudo-
scientific ‘shadow statistics’ websites, to discredit even official government
the other mind than in my own, reflective comparison between the two judgements neces-
sarily reduces me . . . to a state of neutrality.
Contemporary commentators take more varied views on the matter; for an overview see
Goldman and Blanchard (2015).
202 Relentlessly negative reporting, of the sort that has become increasingly common, can have
the same effect. As Patterson (2016) writes, ‘indiscriminate criticism has the effect of blurring
important distinctions. Were the allegations surrounding Clinton of the same order of magnitude
as those surrounding Trump? It’s a question that journalists made no serious effort to answer
during the 2016 campaign.’
203 See for example Asch (1955), Janis (1972), and the many studies that have followed on
from them.
204 Garrett et al. 2016. Engelmann and Fehr 2016.
205 Ipsos MORI 2016, p. 5. In a post-referendum comment, Minister for Brexit, David Davis,
dismissed the importance the ‘£350 million on the side of the bus’ lie promulgated by the Leave
campaign (discussed in Section 21.1.1), saying that the voters ‘dismissed those things [and] made
their judgment on other things’ (Stone 2016a).
206 Margaret Sullivan in Rehm et al. (2016b). Hochschild and Einstein (2015b, pp. 607–8)
quote one conservative commentator explaining Trump’s long-standing claim that Obama was
born in Kenya rather than the US in this way:
[W]hat Donald Trump is doing is questioning things and saying, ‘Why do we have to just
accept everything?’ To hold the birther view is to affiliate oneself with an attitude, not a
truth claim. . . . Your average Trump supporter may [simply] think that the proper attitude to
have toward America’s politicians is contempt’.
statistics. Judging from a poll released a month before the 2016 US election,
that strategy seems to have worked wonders: 44 per cent of Americans said
they distrust official US government economic data.207 The Washington Post
commentator reporting this story describes this as ‘part of his broader narra-
tive of numerical nihilism’—and it is hard to see how it could be described as
anything else.
Inducing people to take an agnostic attitude towards all factual claims can be
epistemically almost as damaging as instilling beliefs in the truth of false facts. It
liberates those in power to implement policies that could only be justified—if they
had to be justified, which in an environment of general epistemic agnosticism
they do not have to be at all—by arguments based on falsehoods.
21.11 CONCLUSION: EPISTEMIC DEMO CRACY

UNDER THREAT
What, then, is the real explanation for the Trump and Brexit victories? Probably
all of the above, in some measure. Our best guess (and we would claim no more
authority for it than that) is this. In the UK, differing priorities, values, and
interests (likely misperceived) were probably the dominant drivers. In the US,
‘fools led by knaves’ is probably a larger part of the story—with voters being
made more foolish by the malicious undermining of all standards of truth, and
knaves being more knavish for their deliberate role in so doing. But as we say,
there were almost certainly elements of all the explanations canvassed above at
work in both countries.
What are the implications for epistemic theories of democracy? That voters
might make mistakes has been part of that story all along. And it has also long
been recognized that even large groups of people might be mistaken when they
all follow too uncritically the same opinion leaders. What the events of 2016
have brought home with particular force is how much each of those standard
caveats must be amplified when voters are systematically subjected to deliberate
misinformation and efforts to undermine all bases of information.
The next step in elaborating the epistemic theory of democracy lies in find-
ing ways to overcome the deleterious effects of such deliberate lies in politics.
But that is not a challenge for epistemic theories alone. It is a challenge for
democratic theory of all forms.
207 Among those who reported themselves as likely to vote for Trump, ‘the share is 60%, with
nearly half saying they don’t trust government economic data “at all” ’ (Rampell 2016).
APPENDIX 1
Key to Notations
n number of voters
Basic CJT with Two Alternatives

pc homogeneous individual voter competence to vote for the correct
alternative, for both possible states of the world
Pn collective group competence: the probability that a majority will
vote for the correct alternative among two alternatives (or, in case
of a tie, will win in a random tiebreaker)
Heterogeneous Voter Competence

pc mean individual voter competence if voter competence is
heterogeneous, for both possible states of the world
pc1 , pc2 , . . . , pcn individual competence of voters v1, v2, . . . , vn
CJT Extension to More Than Two Alternatives

x1, x2, . . . , xk the k alternatives available to choose from
xc the correct alternative
〈 p1i , p2i , . . . , pki 〉 k-tuple of probabilities of believing that x1, x2, . . . , xk are correct, with
pci in this tuple being the probability of voting for xc (The index in
superscript is omitted if all voters have the same probabilities.)
pe the probability of an individual voting for the incorrect alternative e
PnPV the probability that a plurality of voters will vote for the correct
alternative
Idealized Voter Competence, Opinion Leaders, and Cues

pc* the probability of an individual voter voting for the correct
alternative if he or she were to decide without influence of the
specific common causal factor under discussion
π the probability of an individual following an opinion leader or cue
pOL the competence of an opinion leader to vote for the correct alternative
pK the probability that a cue will indicate the correct alternative
370 Appendix 1
Best Responder Corollary

pBR the probability that a voter will vote in the same way as the best
responder, for all possible decision situations
pBR the probability that an average voter will vote in the same way as the
best responder, for all possible decision situations
ω the probability that the decision situation is truth-conducive
Traditions and Cascades

w the weight a person gives to their own personal assessment
Different Values, Priorities, or Interests

V1, V2, . . . different values, priorities, or interests held or pursued by different
groups of voters
X1, X2, . . . the outcomes that are best from the perspective of the corresponding
value, priority, or interest
pcV the probability that a voter will vote for the alternative that is best
from perspective V
pcS and pcL the probability that a voter will vote for the alternative that is best
from their perspective, for a small group (S) or a large group (L)
nS and nL the size of a small (S) and a large group (L)
Factionalism
M the option in the interest of the Masses
E the option in the interest of the Elite
M number of Mass voters
E number of Elite voters
pcM individual competence of the Mass voters to vote for their true interest
pcE individual competence of the Elite voters to vote for their true interest
pgE, pgM competence of individual members of the Elites and the Masses in
selecting who shares the same interests as themselves
APPENDIX 2
Estimating Group Competence

by Monte Carlo Simulation
Table A2.1 specifies the sources of numerical data reported in each of the tables and
figures of the book that contain such data.
Table A2.1 Sources of numerical data in tables and figures.

Analytical results Based on a number of Monte Carlo simulations per data point/process
1,000 10,000 100,000
Figures
2.1 14.3 3.11 6.1
4.1 14.4 3.2 14.12
4.2 3.3
5.8 8.4
8.2 10.2
11.2 10.3
11.3 10.5
13.1 12.3
13.2 15.2
13.3 15.3
13.4
13.5
13.6
13.7
14.2
15.1
15.4
15.5
16.1
Tables
14.1 8.13
16.1 8.2
17.1 11.1
17.3 11.2
17.4 11.3
17.5 11.4
17.6 17.2
1 Lowest curve analytical. 2 CJT curve analytical. 3 k = 2 results analytical.

372 Appendix 2
As will be seen from Table A2.1, in many cases our results rely on Monte Carlo
simulation rather than analytic results. We resort to Monte Carlo simulations, where we
do, because an analytical solution is too complicated or computationally too demanding.
The idea of a Monte Carlo simulation is simple: to use a large number of (pseudo-)
random samples to estimate relevant properties of the process under consideration.4 In
this book we are typically interested in the probability of the correct alternative winning
a majority of votes. We know how individual votes come about, from what we know (or
in this book, typically stipulate) about how competent the voters are, whether and how
much they are influenced by other factors, etc. What we are trying to estimate is, given
those facts, how often the majority of voters will be correct.
To estimate that probability, we simulate a great many (typically 10,000 or more)
rounds of voting. That entails using a computer program to simulate in each round the
vote of each voter, making use of a random number generator set to reflect the various
probabilities stipulated for the case under investigation. After each round we record
whether the majority was correct, incorrect, or whether the vote was tied. After a large
number of rounds of such simulations, we then calculate the proportion of rounds
in which the correct alternative has won a majority of votes. This proportion is our
estimate of the probability of a majority voting for the correct alternative, under the
circumstances specified for that suite of simulations.5
In effect, our Monte Carlo simulations approximate the observation of a great many
actual elections. But a computational simulation has two major advantages: we can
calculate the results of many thousands simulated elections within a few seconds; and
we can set precisely the parameters we want to investigate.
Monte Carlo estimates are highly reliable if they are based on a very large number
of rounds. For the purposes of this book, in which qualitative numerical relations are
more important than numerical precision to many digits, 10,000 (or sometimes even
just 1,000) rounds will typically suffice.
4 See Mooney (1997) for an approachable introduction.

5 We typically also assume that ties are broken with a coin toss and adjust the ratio of ‘correct’
results accordingly.
References
Abercrombie, Nicholas; Stephen Hill; and Bryan S. Turner. 1980. The Dominant Ideology
Thesis. London: Allen & Unwin.
Achen, Christopher H. and Larry M. Bartels. 2016. Democracy for Realists: Why
Elections Do Not Produce Responsive Government. Princeton, NJ: Princeton
University Press.
Ackerman, Bruce. 1985. ‘Beyond Carolene Products’. Harvard Law Review, 98: 713–46.
Ackerman, Bruce A. and James S. Fishkin. 2004. Deliberation Day. New Haven, CT:
Yale University Press.
Adams, John. 1787. A Defence of the Constitutions of Government of the United States of
America. London: C. Dilly.
Agamben, Giorgio. 2005. State of Exception, trans. Kevin Attell. Chicago: University of
Chicago Press.
Allard-Tremblay, Yann. 2017. ‘Divide and Rule Better: On Subsidiarity, Legitimacy
and the Epistemic Aim of Political Decision-making’. Journal of Applied Philosophy,
34: 696-710.
Allcott, Hunt and Matthew Gentzkow. 2017. ‘Social Media and Fake News in the
2016 Election’. NBER Working Paper No. 23089. Available at <http://www.nber.org/
papers/w23089>.
Allen, Katie; Jill Treanor; and Simon Goodley. 2016. ‘Pound Slumps to 31-Year Low
Following Brexit Vote’. Guardian, 23 June; available at <https://www.theguardian.com/
business/2016/jun/23/british-pound-given-boost-by-projected-remain-win-in-eu-
referendum>.
Almandoz, Juan and András Tilcsik. 2016. ‘When Experts Become Liabilities: Domain
Experts on boArds and Organizational Failure’. Academy of Management Journal,
59: 1124–49.
Althaus, Scott L. 1998. ‘Information Effects in Collective Preferences’. American Political
Science Review, 92: 545–58.
Althaus, Scott L. 2003. Collective Preferences in Democratic Politics: Opinion Surveys
and the Will of the People. New York: Cambridge University Press.
Altman, Drew. 2017. ‘The Health Care Plan Trump Voters Really Want’. New York Times,
5 Jan.; available at <https://www.nytimes.com/2017/01/05/opinion/the-health-care-
plan-trump-voters-really-want.html>.
Ambrus, Atilla; Ben Greiner; and Parag A. Pathak. 2015. ‘How Individual Preferences
Are Aggregated in Groups: An Experimental Study’. Journal of Public Economics,
129: 1–13.
American Political Science Association (APSA). 1950. ‘Toward a More Responsible
Two-party System: A Report of the Committee on Political Parties’. American
Political Science Review 44 (no. 3, pt 2, Supplement: Sept).
Amin, Ash. 1996. ‘Beyond Associative Democracy’. New Political Economy, 1: 309–33.
Amis, Martin. 2016. ‘Don the Realtor: The Rise of Trump’. Harper’s Magazine, 13 Feb.;
available at <http://harpers.org/archive/2016/08/don-the-realtor/>.
374 References
Ancell, Aaron. 2017. ‘Democracy Isn’t that Smart (but We Can Make it Smarter): On
Landemore’s Democratic Reason’. Episteme, 14: 161–75.
Anderson, Elizabeth. 1991. ‘John Stuart Mill and Experiments in Living’. Ethics, 102: 4–26.
Anderson, Elizabeth. 2006. ‘The Epistemology of Democracy’. Episteme, 3: 8–22.
Anderson, Elizabeth. 2012. ‘Democracy, Public Policy, and Lay Assessments of Scientific
Testimony’. Episteme, 8: 144–64.
Arendt, Hannah. 1967/1977. ‘Truth and Politics’. New Yorker, 25 Feb., pp. 49–88. Reprinted
pp. 227–64 in Between Past and Future. New York: Penguin, 1977.
Aristotle. 1965. Poetics, trans R. Kassel. Vol. 2, pp. 2316–40 in The Complete Works of
Aristotle, ed. Jonathan Barnes. Princeton, NJ: Princeton University Press, 1984.
Aristotle. 350 bce/1997 The Politics. Pp. 9–208 in Aristotle: The Politics and the Constitution
of Athens, trans. and ed. Stephen Everson. Cambridge: Cambridge University Press;
originally written 350 bce.
Arneson, Richard. 2016. ‘Elitism’. Pp. 156–84 in Oxford Studies in Political Philosophy,
ed. David Sobel, Peter Vallentyne and Steven Wall. Oxford: Oxford University Press.
Arrow, Kenneth J. 1962. ‘The Economic Implications of Learning by Doing’. Review of
Economic Studies, 29: 155–73.
Arrow, Kenneth J. 1963. Social Choice and Individual Values. 2nd edn. New Haven,
CT: Yale University Press.
Arrow, Kenneth J.; Robert Forsythe; Michael Gorham; Robert Hahn; Robin Hanson;
John O. Ledyard; Saul Levmore; Robert Litan; Paul Milgrom; Forrest D. Nelson;
George R. Neumann; Marco Ottaviani; Thomas C. Schelling; Robert J. Shiller;
Vernon L. Smith; Erik Snowberg; Cass R. Sunstein, Paul C. Tetlock; Philip E. Tetlock;
Hal R. Varian; Justin Wolfers; and Eric Zitzewitz. 2008. ‘The Promise of Prediction
Markets’. Science, 320: 877–8.
Asch, Solomon E. 1955. ‘Opinions and Social Pressure’. Scientific American, 193
(5: Nov.): 31–5.
Ash, Timothy Garton. 2016. ‘Do You Live in a Trump Bubble or a Clinton Bubble?’
Guardian, 29 Sept.; available at <https://www.theguardian.com/commentisfree/2016/
sep/29/trump-clinton-media-left-right-democracy>.
Asthana, Anushka and Rowena Mason. 2016. ‘Gove: I Did Almost Everything to Avoid Tory
Leadership Bid’. Guardian, 1 July; available at <https://www.theguardian.com/politics/
2016/jul/01/michael-gove-launches-tory-leadership-bid-with-brexit-promise>.
Atlantic. 2013. ‘What Missile Gap?’ The Atlantic, 316, no. 8; available at <http://www.
theatlantic.com/magazine/archive/2013/08/what-missile-gap/309484/>.
Aumann, Robert J. 1976. ‘Agreeing to Disagree’. Annals of Statistics, 4: 1236–9.
Austen-Smith, David. 1992. ‘Strategic Models of Talk in Political Decision Making’.
International Political Science Review, 13: 45–58.
Austen-Smith, David. 2015. ‘Jon Elster’s Securities Against Misrule: Juries, Assemblies,
Elections: A Review Essay’. Journal of Economic Literature, 53: 65–78.
Austen-Smith, David and Jeffrey S. Banks. 1996. ‘Information Aggregation, Rationality,
and the Condorcet Jury Theorem’. American Political Science Review, 90: 34–45.
Austin, J. L. 1962. Sense and Sensibilia. Oxford: Clarendon Press.
Australia Senate Select Committee on the Scrafton Evidence. 2004. Report. Available at
<http://www.aph.gov.au/Parliamentary_Business/Committees/Senate/Former_
Committees/scrafton/report/index>.
References 375
Axelrod, Robert. 1970. Conflict of Interest. Chicago: Markham.

Bächtiger, André. 2011. ‘Contestatory Deliberation’. Paper presented to the ‘Epistemic
Democracy Conference’, Yale University, 22 October 2011; available at <http://www.
yale.edu/polisci/conferences/epistemic_democracy/aBachtiger.pd>.
Baehr, Jason. 2010. ‘Epistemic Malevolence’. Metaphilosophy, 41: 189–213.
Bagehot, Walter. 1867/2003. The English Constitution. London: Fontana; originally pub-
lished 1867.
Baker, Gerard. 2017. ‘Trump, “Lies” and Honest Journalism: Why Editors Should Be
Careful about Making Selective Moral Judgments about False Statements’. Wall Street
Journal, 4 Jan.; available at <http://www.wsj.com/articles/trump-lies-and-honest-
journalism-1483557700>.
Baker, Peter; Glenn Thrush and Maggie Haberman. 2017. ‘Rocky First Weekend for Trump
Troubles Even His Top Aides’. New York Times, 22 Jan.; available at <https://www.
nytimes.com/2017/01/22/us/politics/donald-trump-aides-rocky-first-weekend.html?h
p&action=click&pgtype=Homepage&clickSource=story-heading&module=first-
column-region&region=top-news&WT.nav=top-news&_r=0>.
Bakshy, Eytan; Solomon Messing; and Lada Adamic. 2015. ‘Exposure to Ideologically
Diverse News and Opinion on Facebook’. Science, 348: 1130–2.
Balz, Dan. 2017. ‘Amid Distractions He Created, Trump Moves Quickly to Change
the Country Dramatically’. Washington Post, 28 Jan.; available at <https://www.
washingtonpost.com/politics/amid-distractions-he-created-trump-moves-swiftly-to-
change-the-country-dramatically/2017/01/28/2f84f884-e4e3-11e6-ba11-63c4b4fb5a63_
story.html?utm_term=.b9c21aacbc3c>.
Banerjee, A. V. 1992. ‘A Simple Model of Herd Behavior’. Quarterly Journal of Economics,
107: 797–817.
Baquet, Dean. 2016. ‘New York Times’ Executive Editor on the New Terrain of Covering
Trump’. NPR Fresh Air, 8 Dec.; available at <http://www.npr.org/2016/12/08/504806512/
new-york-times-executive-editor-on-the-new-terrain-of-covering-trump>.
Barber, Benjamin. 1984. Strong Democracy: Participatory Politics for a New Age. Berkeley,
CA: University of California Press.
Barber, James David. 1972. The Presidential Character. Englewood Cliffs, NJ: Prentice-Hall.
Baron, Martin. 2016. ‘Washington Post Editor Marty Baron Has a Message to Journalists
in the Trump Era’. Vanity Fair, 30 Nov.; available at <https://www.vanityfair.com/
news/2016/11/washington-post-editor-marty-baron-message-to-journalists>.
Barrett, M. J. 2013. ‘Enabling Hybrid Space: Epistemological Diversity in Socio-
ecological Problem-solving’. Policy Sciences, 46: 179–97.
Barry, Brian. 1964. ‘The Public Interest’. Proceedings of the Aristotelian Society (Supplement),
38: 1–18.
Barry, Brian. 1965. Political Argument. London: Routledge & Kegan Paul.
Barstow, David. 2016. ‘Donald Trump’s Deals Rely on Being Creative with the Truth’.
New York Times, 16 July; available at <https://www.nytimes.com/2016/07/17/us/politics/
donald-trump-business.html>.
Bartels, Larry. 1996. ‘Uninformed Votes: Information Effects in Presidential Elections’.
American Journal of Political Science, 40: 194–230.
Bartholdi, J.; C. A. Tovey; and M. A. Trick. 1989. ‘Voting Schemes for which It Can Be
Difficult to Tell Who Won the Election’. Social Choice and Welfare, 6: 157–65.
376 References
Beatty, Jack. 2007. ‘Cognitive Dissonance’. The Atlantic, 310, no. 7. Available at <http://
www.theatlantic.com/magazine/archive/2007/07/cognitive-dissonance/306104/>.
Becker, Selwyn W. and Nicholas Baloff. 1969. ‘Organization Structure and Complex
Problem Solving’. Administrative Science Quarterly, 14: 260–71.
Beerbohm, Eric. 2012. In Our Name: The Ethics of Democracy. Princeton, NJ: Princeton
University Press.
Beerbohm, Eric. 2015. ‘Is Democratic Leadership Possible?’ American Political Science
Review, 109: 639–52.
Beiner, Ronald. 1983. Political Judgment. Chicago: University of Chicago Press.
Ben-Yashar, R. and J. Paroush. 2000. ‘A Non-asymptotic Condorcet Jury Theorem’.
Social Choice and Welfare, 17: 189–99.
Bendor, Jonathan and John G. Bullock. 2008. ‘Lethal Incompetence: Voters, Officials
and Systems’. Critical Review, 20: 1–23.
Benkler, Yochai; Robert Faris; Hal Roberts; and Ethan Zuckerman. 2017. ‘Breitbart-led
Right-wing Media Ecosystem Altered Broader Media Agenda’. Columbia Journalism
Review, 3 Mar.; available at <https://www.cjr.org/analysis/breitbart-media-trump-
harvard-study.php>.
Bentham, Jeremy. 1788/2002. ‘Considérations d’un Anglois sur la Composition des
États-Généraux y Compris Responses aux Questions Proposées aux Notables & c.’.
Pp. 63–146 in Bentham, Rights, Representation and Reform, ed. P. Schofield, C. Pease-
Watkin and C. Blamires. Oxford: Oxford University Press, 2002; originally published
1788.
Bentham, Jeremy. 1791/1999. Political Tactics. Oxford: Oxford University Press; originally
published 1791.
Bentham, Jeremy. 1823/1928. A Comment on the Commentaries, ed. Charles Warren
Everett. Oxford: Clarendon Press; originally published 1823.
Benz, Matthias and Alois Stutzer. 2004. ‘Are Voters Better Informed when They Have a
Larger Say in Politics? Evidence for the European Union and Switzerland’. Public
Choice, 119: 31–59.
Berelson, Bernard R.; Paul F. Lazarsfeld; and William N. McPhee. 1954. Voting: A Study
of Opinion Formation in a Presidential Campaign. Chicago: University of Chicago Press.
Berend, Daniel and Luba Sapir. 2005. ‘Monotonicity in Condorcet Jury Theorem’. Social
Choice and Welfare, 24: 83–92.
Bernstein, Jared. 2017. ‘You Can’t Believe what President Trump Says’. Washington Post,
23 Jan.; available at <https://www.washingtonpost.com/posteverything/wp/2017/01/
23/breaking-news-you-cant-believe-what-president-trump-says/?hpid=hp_hp-cards_
hp-card-posteverything%3Ahomepage%2Fcard&utm_term=.b461501d5687>.
Berry, Jeffrey M. and Sarah Sobieraj. 2011. ‘Understanding the Rise of Talk Radio’.
PS: Political Science and Politics, 44: 762–7.
Betz, Gregor. 2013. Debate Dynamics: How Controversy Improves Our Beliefs. Dordrecht:
Springer.
Bever, Lindsey. 2017. ‘If State Lawmakers Have Their Way, California Schoolchildren
May Be Taught How to Spot “Fake News” ’. Washington Post, 12 Jan.; available at
<https://www.washingtonpost.com/news/the-fix/wp/2017/01/12/if-state-lawmakers-
have-their-way-california-schoolchildren-may-be-taught-how-to-spot-fake-news/
?utm_term=.61266fc9c619>.
References 377
Bicchieri, Cristina and Hugo Mercier. 2013. ‘Self-serving Biases and Public Justification
in Trust Games’. Synthese, 190: 909–22.
Bikhchandani, Suhil; David Hirshleifer; and Ivo Welch. 1992. ‘A Theory of Fads,
Fashion, Custom and Cultural Change as Informational Cascades’. Journal of Political
Economy, 100: 992–1026.
Black, Duncan. 1958. The Theory of Committees and Elections. Cambridge: Cambridge
University Press.
Blackstone, William. 1765. Commentaries on the Laws of England. Oxford: Clarendon Press.
Blake, Aaron. 2016. ‘Do Republicans Have Buyer’s Remorse on Repealing Obamacare?’
Washington Post, 1 Dec.; available at <https://www.washingtonpost.com/news/the-
fix/wp/2016/12/01/do-republicans-have-buyers-remorse-on-repealing-obamacare/
?utm_term=.653a60e45125>.
Blake, Aaron. 2017a. ‘Kellyanne Conway’s Laughable “Look at What’s in His Heart” Defense
of Donald Trump’. Washington Post, 9 Jan.; available at <https://www.washingtonpost.
com/news/the-fix/wp/2017/01/09/kellyanne-conways-laughable-look-at-whats-
in-his-heart-defense-of-trump/?utm_term=.356b32c7e297>.
Blake, Aaron. 2017b. ‘President Trump’s Simplistic, Illogical Worldview, in One
Tweet’. Washington Post, 6 Feb.; available at <https://www.washingtonpost.com/
news/the-fix/wp/2017/02/06/president-trumps-simplistic-illogical-worldview-in-
one-tweet/?hpid=hp_hp-top-table-main_fix-fake-polls-1pm%3Ahomepage%
2Fstory&utm_term=.67ce356edd01>.
Blake, Aaron. 2017c. ‘Sean Spicer’s Defense of Himself and Explanation of Donald
Trump’s Sensitivity, Annotated’. Washington Post, 23 Jan.; available at <https://www.
washingtonpost.com/news/the-fix/wp/2017/01/23/sean-spicers-defense-of-himself-
and-explanation-of-donald-trumps-sensitivity-annotated/?utm_term=.
95931c679d93>.
Blake, Aaron. 2017d. ‘Trump Voters Don’t Have Buyer’s Remorse. But some Hillary Clinton
Voters Do’. Washington Post, 23 April; available at <https://www.washingtonpost.
com/news/the-fix/wp/2017/04/23/trump-voters-dont-have-buyers-remorse-but-
some-hillary-clinton-voters-do/?hpid=hp_hp-cards_hp-card-politics%3Ahomepage
%2Fcard&utm_term=.6e75a1871a59>.
Blumenthal, Sidney. 2016. ‘Abraham Lincoln Would Have Recognized the Con Man
Trump’. Newsweek, 14 Aug.; available at <http://europe.newsweek.com/abraham-lincoln-
would-have-recognised-conman-trump-489841?rm=eu>.
Bohman, James. 2006. ‘Deliberative Democracy and the Epistemic Benefits of Diversity’.
Episteme, 3: 175–91.
Bohman, James and William Rehg, eds, 1997. Deliberative Democracy: Essays on Reason
and Politics. Cambridge, MA: MIT Press.
Bok, Sisela. 1982. Secrecy: On the Ethics of Concealment and Revelation. New York:
Oxford University Press.
Boland, Philip J. 1989. ‘Majority Systems and the Condorcet Jury Theorem’. Journal of
the Royal Statistical Society. Series D (The Statistician), 38: 181–9.
Boland, Philip J.; Frank Proschan; and Y. L. Tong. 1989. ‘Modelling Dependence in
Simple and Indirect Majority Systems’. Journal of Applied Probability, 26: 81–8.
Booth, Robert; Alan Trais; and Amelia Gentleman. 2016. ‘Leave Donor Plans New Party
to Replace UKIP—Possibly without Farage in Charge’. Guardian, 29 June; available at
378 References
<https://www.theguardian.com/politics/2016/jun/29/leave-donor-plans-new-party-
to-replace-ukip-without-farage>.
Borchers, Callum. 2016. ‘4 Threats to the Media under President Trump’. Washington
Post, 10 Nov.; available at <https://www.washingtonpost.com/news/the-fix/wp/2016/
11/10/4-threats-to-the-media-under-president-trump/?utm_term=.30bfd3f8af9c>.
Borchers, Callum. 2017. ‘ “Fake News” Has Now Lost All Meaning’. Washington Post,
9 Feb.; available at <https://www.washingtonpost.com/news/the-fix/wp/2017/02/09/
fake-news-has-now-lost-all-meaning/?hpid=hp_hp-more-top-stories_fake-news-
9am%3Ahomepage%2Fstory&utm_term=.3240db16bdf2>.
Borda, Jean-Charles de. 1784/1994. ‘Mémoire sur les Élections au Scrutin par M. de
Borda [A paper on elections by ballot]’. Translated and reprinted at pp. 114–19 in
Condorcet: Foundations of Social Choice and Political Theory, ed. Iain McLean and
Fiona Hewitt. Aldershot, Hants.: Elgar, 1994; originally published 1784.
Borosage, Robert. 2016. ‘Revolution Incorporated: How Clinton Can Bring Sanders
Supporters into the Fold’. Washington Post, 25 Mar.; available at <https://www.
washingtonpost.com/opinions/revolution-incorporated-how-clinton-can-bring-sanders-
supporters-into-the-fold/2016/03/25/a6a980ec-f03d-11e5-a61f-e9c95c06edca_
story.html?hpid=hp_no-name_opinion-card-e%3Ahomepage%2Fstory>.
Boudreau, Cheryl. 2009. ‘Closing the Gap: When Do Cues Eliminate Differences Between
Sophisticated and Unsophisticated Citizens?’ Journal of Politics, 71: 964–76.
Bovens, Luc and Wlodek Rabinowicz. 2006. ‘Democratic Answers to Complex
Questions—An Epistemic Perspective’. Synthese, 150: 131–53.
Bovens, Mark and Paul ‘t Hart. 1995. Understanding Policy Fiascoes. New Brunswick,
NJ: Transaction Books.
Bovens, Mark and Anchrit Wille. 2010. ‘The Education Gap in Political Participation
and Its Political Consequences’. Acta Politica, 45: 393–422.
Bovens, Mark and Anchrit Wille. 2017. Diploma Democracy: The Rise of Political
Meritocracy. Oxford: Oxford University Press.
Bradley, Richard and Christopher Thompson. 2012. ‘A (Mainly Epistemic) Case for
Multiple-vote Majority Rule’. Episteme, 9: 63–79.
Brams, Steven J. 1978. The Presidential Election Game. New Haven, CT: Yale University
Press.
Brams, Steven J. and Peter C. Fishburn. 1978. ‘Approval Voting’. American Political
Science Review, 72 (3: Sept): 831–47.
Brandeis, Louis. 1932. Dissenting Opinion. New State Ice Co. v. Liebman. 285 US 262.
Brennan, Geoffrey. 2010. ‘The Division of Epistemic Labour’. Analyse & Kritik, 32: 231–46.
Brennan, Geoffrey; Lina Eriksson; Robert E. Goodin; and Nicholas Southwood. 2013.
Explaining Norms. Oxford: Oxford University Press.
Brennan, Geoffrey and Loren Lomasky. 1993. Democracy and Decision. Cambridge:
Cambridge University Press.
Brennan, Jason. 2011a. ‘Condorcet’s Jury Theorem and the Optimum Number of Voters’.
Politics, 31 (2): 55–62.
Brennan, Jason. 2011b. The Ethics of Voting. Princeton, NJ: Princeton University Press.
Brennan, Jason. 2011c. ‘The Right to a Competent Electorate’. Philosophical Quarterly,
61: 700–24.
Brennan, Jason. 2014. ‘How Smart Is Democracy? You Can’t Answer that Question A
Priori’. Critical Review, 26: 33–58.
References 379
Brennan, Jason. 2016. Against Democracy. Princeton, NJ: Princeton University Press.
Brennan, William J. 1958. Opinion of the US Supreme Court. Speiser v Randall. 357
U.S. 514–29.
Brewer, Scott. 1998. ‘Scientific Expert Testimony and Intellectual Due Process’. Yale Law
Journal, 107: 1535–682.
British Broadcasting Corporation (BBC). 2016. ‘EU Referendum Reality Check: Would
Brexit Mean Extra £350m a Week for NHS?’ <http://www.bbc.com/news/uk-politics-
eu-referendum-36040060>.
British Election Study. 2016. ‘Brexit Britain: British Election Study Insights from the
Post-EU Referendum Wave of the BES Internet Panel’. Available at <http://www.
britishelectionstudy.com/bes-resources/brexit-britain-british-election-study-insights-
from-the-post-eu-referendum-wave-of-the-bes-internet-panel/#.WHQKqdaYV_k>.
Brutus. 1787/2003. Letter III. Pp. 453–9 in Hamilton, Madison and Jay, The Federalist,
with Letters of ‘Brutus’, ed. Terence Ball. Cambridge: Cambridge University Press,
2003; originally published 1787.
Bryce, James. 1888. The American Commonwealth. Indianapolis, IN: Liberty Fund, 1995.
Available at <http://oll.libertyfund.org/titles/809>.
Buchanan, James M. and Gordon Tullock. 1962. The Calculus of Consent. Ann Arbor,
MI: University of Michigan Press.
Budge, Ian. 1996. The New Challenge of Direct Democracy. Oxford: Polity.
Budge, Ian; David Robertson; and Derek Hearl, eds. 1987. Ideology, Strategy and Party
Change. Cambridge: Cambridge University Press.
Bueno de Mesquita, Ethan and Matthew Stephenson. 2002. ‘Informative Precedent and
Intrajudicial Communication’. American Political Science Review, 96: 755–66.
Bullock, John G.; Alan S. Gerger; Seth J. Hill; and Gregory A. Huber. 2015. ‘Partisan Bias
in Factual Beliefs about Politics’. Quarterly Journal of Political Science, 10: 519–78.
Bump, Philip. 2016. ‘It’s Hard to Imagine a Much Worse Pitch Donald Trump Could
Have Made for the Black Vote’. Washington Post, 20 Aug.; available at <https://www.
washingtonpost.com/news/the-fix/wp/2016/08/20/its-hard-to-imagine-a-much-
worse-pitch-donald-trump-could-have-made-for-the-black-vote/?utm_term=
.a6956fc09e58>.
Burke, Edmund. 1774. ‘Speech to the Electors of Bristol at the Conclusion of the Poll’.
Available at <http://oll.libertyfund.org/index.php?option=com_staticxt&staticfile=
show.php?title%3D659&layout=html>.
Burns, James Macgregor. 1956. Roosevelt: The Lion and the Fox. New York: Harcourt,
Brace.
Caillaud, Bernard and Jean Tirole. 2002. ‘Parties as Political Intermediaries’. Quarterly
Journal of Economics, 117: 1453–89.
Calhoun, Craig. 2002. ‘The Class Consciousness of Frequent Travelers: Towards a Critique
of Actually Existing Cosmopolitanism’. Pp. 86–109 in Conceiving Cosmopolitanism, ed.
Steven Vertovec and Robin Cohen. Oxford: Oxford University Press.
Calhoun, John C. 1853/1992. ‘A Disquisition on Government’. Pp. 5–78 in John C.
Calhoun, Union & Liberty, ed. Ross M. Lence. Indianapolis, IN: Liberty Fund, 1992;
originally published 1853.
Caluwaerts, Didier and Kris Deschouwer. 2014. ‘Building Bridges across Political
Divides: Experiments on Deliberative Democracy in Deeply Divided Belgium’.
European Political Science Review, 6: 427–50.
380 References
Campbell, Angus; Philip E. Converse; Warren A. Miller; and Donald Stokes. 1960. The
American Voter. New York: Wiley.
Cannon, Lou. 1991. Ronald Reagan: The Role of a Lifetime. New York: Simon & Schuster.
Capehart, Jonathan. 2017. ‘Arthur Brooks: The Word Dignity Explains Why Trump is
President’. Washington Post, 24 Jan.; available at <https://www.washingtonpost.com/
blogs/post-partisan/wp/2017/01/24/arthur-brooks-the-word-dignity-explains-why-
trump-is-president/?hpid=hp_no-name_opinion-card-c%3Ahomepage%2Fstory&
utm_term=.ceca525c0b59>.
Caplan, Bryan. 2007. The Myth of the Rational Voter: Why Democracies Choose Bad
Policies. Princeton, NJ: Princeton University Press.
Carroll, Rory. 2016. ‘How Breitbart and the Conservative Right Opened a New Front in
the War over Fake News’. Guardian, 15 Dec.; available at <https://www.theguardian.
com/us-news/2016/dec/15/breitbart-fake-news-right-trump>.
Casarett, David. 2016. ‘The Science of Choosing Wisely—Overcoming the Therapeutic
Illusion’. New England Journal of Medicine, 374: 1203–5.
Cassam, Quassim. 2016. ‘Epistemic Postures’. Vice Epistemology. Mimeo., Department
of Philosophy, University of Warwick.
Central Intelligence Agency (CIA). 2013. Penetrating the Iron Curtain: Resolving the
Missile Gap with Technology. Washington, DC: Government Printing Office for
CIA. Available at <https://www.cia.gov/library/publications/cold-war/resolving-the-
missile-gap-with-technology>.
Chamley, C. P. 2004. Rational Herds: Economic Models of Social Learning. Cambridge:
Charles, Ron. 2017. ‘Why Orwell’s 1984 Matters So Much Now’. Washington Post, 25 Jan.;
available at <https://www.washingtonpost.com/entertainment/books/why-orwells-
1984-matters-so-much-now/2017/01/25/3cf81964-e313-11e6-a453-19ec4b3d09ba_
story.html?hpid=hp_hp-more-top-stories_bw-orwell552pm%3Ahomepage%2Fstor
y&utm_term=.ec8f6783d173>.
Chen, Kay-Yut and Charles R. Plott. 2002. ‘Information Aggregation Mechanisms:
Concept, Design and Implementation for a Sales Forecasting Problem’. Technical
Report 1131, California Institute of Technology. Available at: <http://www.hss.caltech.
edu/content/information-aggregation-mechanisms-concept-design-and-implementation-
sales-forecasting>.
Cheney, Kyle; Isaac Arnsdorf; Daniel Lippman; Daniel Strauss; and Brent Griffiths.
2016. ‘Donald Trump’s Week of Misrepresentations, Exaggerations and Half-truths’.
'Politico, 25 Sept.; available at <http://www.politico.com/magazine/story/2016/09/
2016-donald-trump-fact-check-week-214287>.
Chong, Dennis and James N. Druckman. 2007a. ‘Framing Theory’. Annual Review of
Political Science, 10: 103–26.
Chong, Dennis and James N. Druckman. 2007b. ‘A Theory of Framing and Opinion
Formation in Competitive Elite Environments’. Journal of Communication,
57: 99–118.
Christiano, Thomas. 1995. ‘Voting and Democracy’. Canadian Journal of Philosophy,
25: 395–414.
Christiano, Thomas. 2012. ‘Rational Deliberation among Experts and Citizens’.
Pp. 27–51 in John Parkinson and Jane Mansbridge, eds, Deliberative Systems.
References 381
Churchland, Paul. 1995. The Engine of Reason, the State of the Soul: A Philosophical
Journey into the Brain. Cambridge, MA: MIT Press.
Cillizza, Chris. 2016. ‘How the Heck Can Voters Think Donald Trump Is More Honest than
Hillary Clinton?’ Washington Post, 2 Nov.; available at <https://www.washingtonpost.
com/news/the-fix/wp/2016/11/02/donald-trump-hasnt-told-the-truth-repeatedly-in-
this-campaign-voters-still-think-he-is-more-honest-than-hillary-clinton/?utm_
term=.ac1c8206c351>.
Cillizza, Chris. 2017. ‘A Fascinating Theory from the World of Sports about Donald Trump’s
First 7 Days’. Washington Post, 30 Jan.; available at <https://www.washingtonpost.com/
news/the-fix/wp/2017/01/30/a-fascinating-theory-from-the-world-of-sports-about-
donald-trumps-governing-style/?hpid=hp_hp-cards_hp-card-politics%3Ahomepage
%2Fcard&utm_term=.eb6ba4db337f>.
Clark, Andy. 1997. Being There: Putting Brain, Body and World Together Again. Cambridge,
MA: MIT Press.
CNN. 2016. Exit polls. CNN Politics, 23 Nov.; available at <http://edition.cnn.com/election/
results/exit-polls>.
Coady, C.A.J. 1992. Testimony. Oxford: Clarendon Press.
Cohen, Joshua. 1986. ‘An Epistemic Conception of Democracy’. Ethics, 97: 26–38.
Cohen, Joshua. 2009. ‘Truth and Public Reason’. Philosophy & Public Affairs, 37: 2–42.
Cohen, Joshua. 2010. Rousseau: A Free Community of Equals. Oxford: Oxford University
Press.
Cohen, Joshua and Joel Rogers. 1995. Democracy and Secondary Associations. London:
Verso.
Colomer, Josep M. and Iain McLean. 1998. ‘Electing Popes: Approval Voting and
Qualified-Majority Rule’. Journal of Interdisciplinary History, 29: 1–22.
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1785. Essai sur
l’Application de l’Analyse à la Probabilité des Décisions Rendues à la Pluralité des Voix.
Paris: l’Imprimerie Royale, 1785. Fascimile edition: New York: Chelsea, 1972. Translated
and reprinted in part in: Condorcet 1785/1976; 1785/1989; 1785/1994.
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1785/1976. ‘Essay on the
Application of Mathematics to the Theory of Decision-making’. Pp. 33–70 in
Condorcet: Selected Writings, trans and ed. Keith Michael Baker. Indianapolis, IN:
Bobbs-Merrill, 1976; originally published 1785.
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1785/1989. ‘An Essay on
the Application of Probability Theory to Plurality Decision-making’. Pp. 69–118 in
The Political Theory of Condorcet, trans. and ed. Fiona Sommerlad and Iain McLean.
Social Studies Faculty Centre Working Paper 1/89. Oxford: Faculty of Social Sciences,
Oxford University, 1989; originally published 1785.
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1785/1994. ‘An Essay on
the Application of Probability Theory to Plurality Decision-making’. Pp. 120–38 in
Condorcet: Foundations of Social Choice and Political Theory, ed. Iain McLean and
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1787/1994. ‘Letters from
a Free Man of New Haven to a Citizen of Virginia on the Futility of Dividing the
Power among Several Bodies’. Pp. 292–334 in in Condorcet: Foundations of Social
Choice and Political Theory, trans. and ed. Iain McLean and Fiona Hewitt. Aldershot,
Hants.: Elgar, 1994; originally published 1787.
382 References
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1789/1994. ‘On the
Form of Elections’. Pp. 169–89 in Condorcet: Foundations of Social Choice and Political
Theory, trans. and ed. Iain McLean and Fiona Hewitt. Aldershot, Hants.: Elgar, 1994;
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1790/1994. ‘On Giving
Women the Right of Citizenship’. Pp. 335–40 in Condorcet: Foundations of Social
Choice and Political Theory, trans. and ed. Iain McLean and Fiona Hewitt. Aldershot:
Elgar, 1994; originally published 1790.
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1793/1994. ‘A Survey of
the Principles Underlying the Draft Constitution’. Pp. 190–227 in Condorcet:
Foundations of Social Choice and Political Theory, trans. and ed. Iain McLean and
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1794/1994. ‘Condorcet’s
Testament’. Pp. 290–1 in Condorcet: Foundations of Social Choice and Political
Theory, ed. Iain McLean and Fiona Hewitt. Aldershot, Hants.: Elgar; originally
written 1794.
Condorcet, Marie Jean Antoine Nicolas de Caritat, Marquis de. 1795/1796. Outlines of
an Historical View of the Progress of the Human Mind. Philadelphia, MN: M. Carey;
available at <http://oll.libertyfund.org/titles/1669>.
Cofessore, Nicholas and Karen Yourish. 2016. ‘$2 Billion Worth of Free Media for
Donald Trump’. New York Times, 16 Mar.; available at <https://www.nytimes.com/
2016/03/16/upshot/measuring-donald-trumps-mammoth-advantage-in-free-
media.html?_r=0>.
Congressional Budget Office (CBO). 2017. ‘How Repealing Portions of the Affordable
Care Act Would Affect Health Insurance Coverage and Premiums’. Available at
<https://www.cbo.gov/publication/52371>.
Connolly, Kate; Angelinque Chrisafis; Poppy McPherson; Stephanie Kirchgaessner;
Benjamin Haas, Dominic Phillips; Elle Hunt; and Michael Safi. 2016. ‘Fake News: An
Insidious Trend that’s Fast Becoming a Global Problem’. Guardian, 2 Dec.; available at
<https://www.theguardian.com/media/2016/dec/02/fake-news-facebook-us-election-
around-the-world>.
Converse, Philip E. 1964. ‘The Nature of Belief Systems in Mass Publics’. Pp. 206–61
in Ideology & Discontent, ed. David E. Apter. New York: Free Press. Reprinted in
Critical Review, 18 (2006): 1–74.
Converse, Philip E. 2000. ‘Assessing the Capacity of Mass Electorates’. Annual Reviews
of Political Science, 3: 331–53.
Converse, Philip E. and Gregory B. Markus. 1979. ‘Plus ça Change . . .: The New CPS
Election Study Panel’. American Political Science Review, 73: 32–49.
Cook, Charlie. 2016. ‘The Truthful Hyperbole of Donald Trump’. National Journal,
17 Nov.;availableat<https://www.nationaljournal.com/s/645222/truthful-hyperbole-
donald-trump>.
Cook, John and Stephan Lewandowsky. 2011. The Debunking Handbook. St. Lucia, QLD:
University of Queensland. Available at <http://sks.to/debunk>.
Cooke, Maeve. 2013. ‘Violating Neutrality? Religious Validity Claims and Democratic
Legitimacy’. Pp. 249–75 in Habermas and Religion, ed. Craig Calhoun, Eduardo
Mendieta and Jonathan VanAntwerpen. Cambridge: Polity.
References 383
Copp, David. 1993. ‘Could Political Truth Be a Hazard for Democracy?’ Pp. 101–17 in
The Idea of Democracy, ed. David Copp, Jean Hampton and John E. Roemer. Cambridge:
Cowen, Tyler. 2017. ‘Why Trump’s Staff Is Lying’. Bloomberg View, 23 Jan.; available at
<https://www.bloomberg.com/view/articles/2017-01-23/why-trump-s-staff-is-
lying>.
Cronin, Thomas E. 1989. Direct Democracy: The Politics of Initiative, Referendum and
Recall. Cambridge, MA: Harvard University Press.
Curtis, Adam. 2014. ‘Charlie Brooker’s 2014 Wipe: Oh Dearism II’. BBC, 30 Dec.;
available at <http://www.realclearpolitics.com/video/2014/12/31/bbcs_adam_curtis_
on_the_contradictory_vaudeville_of_post-modern_politics.html>.
Dahl, Robert A. 1967. Pluralist Democracy in the United States. Chicago: Rand McNally.
Dahl, Robert A. 1972. Polyarchy: Participation and Opposition. New Haven, CT: Yale
University Press.
Dahl, Robert A. 1982. Dilemmas of a Pluralist Democracy. New Haven, CT: Yale University
Press.
Dahl, Robert A. 1989. Democracy and Its Critics. New Haven, CT: Yale University Press.
Dahlberg, Stefan and Sören Holmberg. 2014 . ‘Democracy and Bureaucracy: How Their
Quality Matters for Popular Satisfaction’. West European Politics, 37: 515–37.
Dal Bó, Ernesto; Frederico Finan; Olle Folke; Torsten Persson; and Johanna Rickne.
2017. ‘Who Becomes a Politician?’ Quarterly Journal of Economics, 132: 1877–1914.
Dancey, Logan and Geoffrey Sheagley. 2013. ‘Heuristics Behaving Badly: Party Cues
and Voter Knowledge’. American Journal of Political Science, 57: 312–25.
Danziger, Shai; Jonathan Levav; and Liora Avnaim-Pesso. 2011. ‘Extraneous Factors in
Judicial Decisions’. Proceedings of the National Academy of Sciences, 108: 6889–92.
Daugherty, Andrew F. and Jennifer F. Reinganum. 1999. ‘Stampede to Judgment:
Persuasive Influence and Herding Behavior by Courts’. American Law & Economics
Review, 1: 158–89.
Dean, Alex. 2016. ‘Which UK Regions Receive the Most EU Funding? And Which of
Them Voted for Brexit?’ Prospect, 7 Sept.; available at <http://www.prospectmagazine.
co.uk/politics/which-uk-regions-receive-the-most-eu-funding>.
Decker, Cathleen. 2016. ‘Trump’s Twitter Falsehoods Fire Up His Base, Provoke
Opponents and Distract from Larger Issues’. Los Angeles Times, 15 Dec.; available at
<http://www.latimes.com/politics/la-na-pol-trump-truth-20161215-story.html>.
Delli Carpini, Michael X.; Fay Lomax Cook; and Lawrence R. Jacobs. 2004. ‘Public
Deliberation, Discursive Participation, and Citizen Engagement: A Review of the
Empirical Literature’. Annual Review of Political Science, 7: 315–44.
Delli Carpini, Michael X. and Scott Keeter. 1996. What Americans Know About Politics
and Why It Matters. New Haven, CT: Yale University Press.
Dewan, Torun and David P. Myatt. 2008. ‘The Qualities of Leadership: Direction,
Communication, and Obfuscation’. American Political Science Review, 102: 351–68.
Dewey, Caitlin. 2016. ‘Facebook Fake-news Writer: “I Think Donald Trump Is in the
White House because of Me” ’. Washington Post, 17 Nov.; available at <https://www.
washingtonpost.com/news/the-intersect/wp/2016/11/17/facebook-fake-news-writer-
i-think-donald-trump-is-in-the-white-house-because-of-me/?utm_term=
.8e3b9f26071b>.
384 References
Dhami, M. K. 2003. ‘Psychological Models of Professional Decision Making’. Psychological

Science, 14: 175–80.
Dickinson, Tim. 2011. ‘How Roger Ailes Built the Fox News Fear Factory’. Rolling Stone,
25 May; available at <http://www.rollingstone.com/politics/news/how-roger-ailes-
built-the-fox-news-fear-factory-20110525>.
Dickson, Eric S.; Catherine Hafer; and Dimitri Landa. 2008. ‘Cognition and Strategy:
A Deliberation Experiment’. Journal of Politics, 70: 974–89.
Dickson, Eric S.; Catherine Hafer; and Dimitri Landa. 2015. ‘Learning from Debate:
Institutions and Information’. Political Science Research and Methods, 3: 449–72.
Diehl, Michael and Wolfgang Stroebe. 1987. ‘Productivity Loss in Brainstorming
Groups: Toward the Solution of a Riddle’. Journal of Personality and Social Psychology,
53: 497–509.
Diehl, Michael and Wolfgang Stroebe. 1991. ‘Productivity Loss in Idea-generating Groups:
Tracking Down the Blocking Effect’. Journal of Personality and Social Psychology,
61: 392–403.
Dietrich, Franz. 2008. ‘The Premises of Condorcet’s Jury Theorem Are Not Simultaneously
Justified’. Episteme, 5 (1): 56–73.
Dietrich, Franz and Christian List. 2004. ‘A Model of Jury Decisions Where All Jurors
Have the Same Evidence’. Synthese, 142: 175–202.
Dietrich, Franz and Christian List. 2007. ‘Arrow’s Theorem in Judgment Aggregation’.
Dietrich, Franz and Kai Spiekermann. 2013a. ‘Epistemic Democracy with Defensible
Premises’. Economics & Philosophy, 29: 87–120.
Dietrich, Franz and Kai Spiekermann. 2013b. ‘Independent Opinions? On the Causal
Foundations of Belief Formation and Jury Theorems’. Mind, 122: 655–85.
Dilnot, Andrew. 2016a. Letter to The Rt Hon Norman Lamb MP. Available at <https://
www.statisticsauthority.gov.uk/wp-content/uploads/2016/04/Letter-from-Sir-
Andrew-Dilnot-to-Norman-Lamb-MP-210416.pdf>.
Dilnot, Andrew. 2016b. ‘UK Statistics Authority Statement on the Use of Official
Statistics on Contributions to the European Union’. Available at <https://www.
statisticsauthority.gov.uk/news/uk-statistics-authority-statement-on-the-use-of-
official-statistics-on-contributions-to-the-european-union/>.
Dionne, E. J. Jr 2017. ‘What’s the Method in Trump’s Madness?’ Washington Post,
25 Jan.; available at <https://www.washingtonpost.com/opinions/whats-the-method-
in-trumps-madness/2017/01/25/3fafae32-e33c-11e6-ba11-63c4b4fb5a63_story.
html?hpid=hp_no-name_opinion-card-c%3Ahomepage%2Fstory&utm_term=
.b01b7f271b04>.
Dodd, Bill. 2017. Senate Bill No. 135, California Legislature: 2017–2018 Regular Session.
Available at <https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=
201720180SB135>.
Donovan, Todd; Caroline J. Tolbert; and Daniel A. Smith. 2009. ‘Political Engagement,
Mobilization and Direct Democracy’. Public Opinion Quarterly, 73: 98–118.
Douglas, Lawrence. 2017. ‘Why Trump Wants to Disempower Institutions that Protect
the Truth’. Washington Post, 7 Feb.; available at <https://www.theguardian.com/
commentisfree/2017/feb/07/trump-disempower-institutions-protect-truth>.
Douglas, William O. 1949. ‘Stare Decisis’. Columbia Law Review, 49: 735–58.
Downs, Anthony. 1957. An Economic Theory of Democracy. New York: Harper.
References 385
Driver, Julia. 2013. ‘Moral Expertise: Judgment, Practice and Analysis’. Social Philosophy
and Policy, 30: 280–96.
Druckman, James N. 2004. ‘Political Preference Formation: Competition, Deliberation
and the (Ir)Relevance of Framing Effects’. American Political Science Review, 98: 671–86.
Druckman, James N. and Kjersten R. Nelson. 2003. ‘Framing and Deliberation: How
Citizens’ Conversations Limit Elite Influence’. American Journal of Political Science,
47: 729–45.
Dryzek, John S. 2000. Deliberative Democracy and Beyond. Oxford: Oxford University
Press.
Dryzek, John S. 2009. ‘Democratization as Deliberative Capacity Building’. Comparative
Political Studies, 42: 1379–402.
Dryzek, John S. and Simon J. Niemeyer. 2006. ‘Reconciling Pluralism and Consensus as
Political Ideals’. American Journal of Political Science, 50: 634–49.
Dunleavy, Patrick. 1986. ‘Explaining the Privatization Boom: Public Choice versus
Radical Approaches’. Public Administration, 64: 13–34.
Dunleavy, Patrick. 1991. Democracy, Bureaucracy and Public Choice. Hemel Hempstead:
Harverster Wheatsheaf.
Dunleavy, Patrick and Hugh Ward. 1981. ‘Exogenous Voter Preferences and Parties with
State Power’. British Journal of Political Science, 11: 351–80.
Dworkin, Gerald. 1988. ‘Is More Choice Better than Less?’ Pp. 62–84 in Dworkin, The
Theory and Practice of Autonomy. Cambridge: Cambridge University Press.
Dworkin, Ronald. 2009. ‘Looking for Cass Sunstein’. New York Review of Books,
30 April 2009. Available at <http://www.nybooks.com/articles/archives/2009/apr/30/
looking-for-cass-sunstein/>.
Ebeling, Martin. 2016. ‘Epistemic Political Egalitarianism and the Role of Parties’.
Political Theory, 44: 629–56.
Economist. 2016. ‘Brexit Regret’. Economist, 12 Oct.; available at <http://www.economist.
com/blogs/graphicdetail/2016/10/daily-chart-6>.
Edelman, Paul H. 2002. ‘On Legal Interpretations of the Condorcet Jury Theorem’.
Journal of Legal Studies, 31: 327–49.
Ehrenfreund, Max. 2016. ‘The Problem with Trump’s Question to Black Voters, “What
the Hell Do You Have to Lose” ’. Washington Post, 24 Aug.; available at <https://www.
washingtonpost.com/news/wonk/wp/2016/08/24/the-problem-with-trumps-question-
to-black-voters-what-the-hell-do-you-have-to-lose/?utm_term=.b5e4da776f21>.
Elster, Jon. 1983. Sour Grapes. Cambridge: Cambridge University Press.
Elster, Jon. 2000. ‘Arguing and Bargaining in Two Constituent Assemblies’. University of
Pennsylvania Journal of Constitutional Law, 2: 345–421.
Elster, Jon. 2013. Securities against Misrule: Juries, Assemblies, Elections. Cambridge:
Elster, Jon, ed. 2015. Secrecy and Publicity in Votes and Debates. Cambridge: Cambridge
University Press.
Engelmann, Jan B. and Ernst Fehr. 2016. ‘The Slippery Slope of Dishonesty’. Nature
Neuroscience, 19: 1543–4.
Eskridge, William N., Jr and John Ferejohn. 2001. ‘Structuring Lawmaking to Reduce
Cognitive Bias: A Critical View’. Cornell Law Review, 87: 616–47.
Estlund, David. 1989. ‘Democratic Theory and the Public Interest: Condorcet and
Rousseau Revisited’. American Political Science Review, 83: 1317–22.
386 References
Estlund, David. 1993. ‘Making Truth Safe for Democracy’. Pp. 71–100 in The Idea of
Democracy, ed. David Copp, Jean Hampton and John E. Roemer. New York:
Estlund, David. 1994. ‘Opinion Leaders, Independence and Condorcet’s Jury Theorem’.
Theory & Decision, 36: 131–62.
Estlund, David. 1997. ‘Beyond Fairness and Deliberation: The Epistemic Dimension of
Democratic Authority’. Pp. 173–204 in Deliberative Democracy, ed. James Bohman
and William Rehg. Cambridge, MA: MIT Press.
Estlund, David. 2005. ‘Review of Reflective Democracy by Robert E. Goodin’. Ethics,
115: 609–14.
Estlund, David. 2008. Democratic Authority: A Philosophical Framework. Princeton,
NJ: Princeton University Press.
Estlund, David. 2012. ‘Democratic Counts: Should Rulers be Numerous?’ Pp. 230–50 in
Collective Wisdom, ed. Hélène Landemore and Jon Elster. Cambridge: Cambridge
University Press.
Estlund, David and Hélène Landemore. 2018. ‘The Epistemic Value of Democratic
Deliberation’. In Oxford Handbook of Deliberative Democracy, ed. André Bächtiger,
John Dryzek, Jane J. Mansbridge and Mark Warren. Oxford: Oxford University Press.
Estlund, David; Jeremy Waldron; Bernard Grofman and Scott L. Owen. 1989.
‘Democratic Theory and the Public Interest: Condorcet and Rousseau Revisited’.
American Political Science Review, 83: 1317–40.
Fahrenthold, David A. and Robert O’Harrow, Jr 2016. ‘Trump: A True Story’. Washington
Post, 10 Aug.; available at <https://www.washingtonpost.com/graphics/politics/2016-
election/trump-lies/>.
Faiola, Anthony and Stephanie Kirchner. 2017. ‘How Do You Stop Fake News?
In Germany, with a Law’. Washington Post, 5 Ap.; available at <https://www.
washingtonpost.com/world/europe/how-do-you-stop-fake-news-in-germany-with-
a-law/2017/04/05/e6834ad6-1a08-11e7-bcc2-7d1a0973e7b2_story.html?hpid=hp_
hp-cards_hp-card-world%3Ahomepage%2Fcard&utm_term=.dcf3a54ab5bb>.
Fallows, James. 2016a. ‘How to Deal with the Lies of Donald Trump: Guidelines for the
Media’. The Atlantic, 28 Nov.; available at <https://www.theatlantic.com/notes/2016/
11/a-reflexive-liar-in-command-guidelines-for-the-media/508832/>.
Fallows, James. 2016b. ‘The Daily Trump: Filling a Time Capsule’. Available at <http://
www.theatlantic.com/notes/all/2016/05/the-daily-trump/484064/>.
Farhi, Paul. 2017a. ‘Trump Has a Love–Hate Relationship with the New York Times.
Now He Wishes it Would Die’. Washington Post, 29 Jan.; available at <http://www.
standard.net/National/2017/01/29/Trump-s-displeasure-with-the-NY-Times-grows>.
Farhi, Paul. 2017b. ‘Trump’s “War” with the Media (and the Facts) Forced Journalists
to Question Their Role’. Washington Post, 26 Jan.; available at <https://www.
washingtonpost.com/lifestyle/style/trumps-war-with-the-media-and-the-facts-
forces-journalists-to-question-their-role/2017/01/26/0b710f4c-e31b-11e6-ba11-
63c4b4fb5a63_story.html?tid=a_inl&utm_term=.0bfd1d2013b5>.
Farquharson, Robin. 1969. Theory of Voting. Oxford: Blackwell.
Farr, James. 2004. ‘The Science of Politics—as Civic Education—Then and Now’. PS,
37: 37–40.
Farrelly, Colin. 2012. ‘Virtue Epistemology and the “Epistemic Fitness” of Democracy’.
Political Studies Review, 10: 7–22.
References 387
Fearon, James. 1998. ‘Deliberation as Discussion’. Pp. 44–68 in Deliberative Democracy,

ed. Jon Elster. Cambridge: Cambridge University Press.
Feddersen, Timothy and Wolfgang Pesendorfer. 1998. ‘Convicting the Innocent: The
Inferiority of Unanimous Jury Verdicts under Strategic Voting’. American Political
Feld, Scott and Bernard Grofman. 1988. ‘Ideological Consistency as a Collective
Phenomenon’. American Political Science Review, 82: 773–88.
Feld, Scott L., and Bernard Grofman. 1992. ‘Who’s Afraid of the Big Bad Cycle? Evidence
from 36 Elections’. Journal of Theoretical Politics, 4: 231–7.
Feldman, Stanley; Leonie Huddy; and George E. Marcus. 2012. ‘Limits of Elite Influence
on Public Opinion’. Critical Review, 24: 489–504.
Felsenthal, Dan S.; Zeev Maoz; and Amnon Rapoport. 1993. ‘An Empirical Evaluation
of Six Voting Procedures: Do They Really Make Any Difference?’ British Journal of
Fenno, Richard F. Jr 1962. ‘The Appropriations Committee as a Political System’.
Fenno, Richard F. Jr 1973. Congressmen in Committees. Boston, MA: Little, Brown.
Ferejohn, John A. 1974. Pork Barrel Politics. Stanford, CA: Stanford University Press.
Ferejohn, John A. and James H. Kuklinski, eds. 1990. Information and Democratic
Processes. Urbana: University of Illinois Press.
Fey, Mark. 2003. ‘A Note on the Condorcet Jury Theorem with Supermajority Voting
Rules’. Social Choice and Welfare, 20: 27–32.
Financial Times. 2017. FTSE 100 Index. Available at <https://markets.ft.com/data/indices/
tearsheet/summary?s=FTSE:FSI>.
Fiorina, Morris P. and Samuel J. Abrams. 2007. ‘Political Polarization in the American
Public’. Annual Review of Political Science, 11: 556–60.
Fishkin, James S. 1995. The Voice of the People. New Haven, CT: Yale University Press.
Fishkin, James S. 2009. When the People Speak: Deliberative Democracy and Public
Consultation. New York: Oxford University Press.
Fishkin, James S.; Baogang He; Robert C Luskin; and Alice Siu. 2010. ‘Deliberative
Democracy in an Unlikely Place: Deliberative Polling in China’. British Journal of
FiveThirtyEight. 2016. FiveThirtyEight 2016 Election Forecast, President; updated Nov 8,
2016. Available at <https://projects.fivethirtyeight.com/2016-election-forecast/>.
Fletcher, George. 1968. ‘Two Kinds of Legal Rules’. Yale Law Journal, 78: 880–935.
Flood, Alison. 2016. ‘“Post-truth” Named Word of the Year by Oxford Dictionaries’.
Guardian, 15 Nov.; available at <https://www.theguardian.com/books/2016/nov/15/
post-truth-named-word-of-the-year-by-oxford-dictionaries>.
Flores, Reena. 2016. ‘Donald Trump Never Threw a Baby Out of a Rally’. CBS News,
6 Aug.; available at <http://www.cbsnews.com/news/report-donald-trump-never-
threw-a-baby-out-of-a-rally/>.
Floridi, Luciano. 2016. ‘Fake News and a 400-Year-Old Problem: We Need to Resolve the
“Post-truth” Crisis’. Guardian, 29 Nov.; available at <https://www.theguardian.com/
technology/2016/nov/29/fake-news-echo-chamber-ethics-infosphere-internet-digital>.
Flynn, D.J.; Brendan Nyhan; and Jason Reifler. 2017. ‘The Nature and Origins of
Misperceptions: Understanding False and Unsupported Beliefs about Politics’.
Advances in Political Psychology, 38 (Supplement S1): 127–50.
388 References
Forster, Katie. 2016. ‘EU Referendum: Thousands of Commuters Affected by Flooding’.

The Independent, 24 June; available at <http://www.independent.co.uk/news/uk/
home-news/eu-referendum-polling-stations-travel-time-extension-latest-weather-
disruptions-a7098466.html>.
Forsythe, Robert; Forrest Nelson; George R. Neumann; and Jack Wright. 1992. ‘Anatomy
of an Experimental Political Stock Market’. American Economic Review, 82: 1142–61.
Forsythe, Robert; Thomas A. Rietza; and Thomas W. Ross. 1999. ‘Wishes, Expectations
and Actions: A Survey on Price Formation in Election Stock Markets’. Journal of
Economic Behavior and Organization, 39: 83–110.
Frankfurt, Harry G. 1988. ‘On Bullshit’. Pp. 117–34 in Frankfurt, The Importance of What
We Care About. Cambridge: Cambridge University Press.
Frankfurt, Harry G. 2016. ‘Donald Trump is BS, Says Expert in BS’. Time, 12 May;
available at <http://time.com/4321036/donald-trump-bs/>.
Frayer, Lauren. 2016. ‘Boris Johnson: London’s Ex-Mayor Could Be Britain’s Next
Prime Minister’. NPR, 25 June; available at <http://www.npr.org/sections/parallels/
2016/06/25/483414071/boris-johnson-londons-ex-mayor-could-be-britains-next-
prime-minister>.
Frazer, Michael. 2014. ‘Including the Unaffected’. Journal of Political Philosophy,
22: 377–95.
Freedland, Jonathan. 2016. ‘Post-truth Politicians such as Donald Trump and Boris
Johnson Are No Joke’. Guardian, 14 May; available at <https://www.theguardian.com/
commentisfree/2016/may/13/boris-johnson-donald-trump-post-truth-politician>.
Freeman, Hadley. 2017. ‘Tie Length: A Worrying Guide to Trump’s State of Mind’.
Guardian, 30 Jan; available at <https://www.theguardian.com/fashion/2017/jan/30/
tie-length-a-worrying-guide-to-trumps-state-of-mind>.
Fricker, Miranda. 2007. Epistemic Injustice: Power and the Ethics of Knowing. Oxford:
Fricker, Miranda. 2013. ‘Epistemic Justice as a Condition of Political Freedom?’ Synthese,
190: 1317–32.
Fritz, Ben; Bryan Keefer; and Brenden Nyhan. 2004. All the President’s Spin: George W.
Bush, the Media and the Truth. New York: Simon & Schuster.
Fuerstein, Michael. 2008. ‘Epistemic Democracy and the Social Character of Knowledge’.
Episteme, 5: 74–93.
Fuerstein, Michael. 2013. ‘Epistemic Trust and Liberal Justification’. Journal of Political
Philosophy, 21: 179–99.
Fung, Archon. 2003. ‘Recipes for Public Spheres: Eight Institutional Choices and Their
Consequences’. Journal of Political Philosophy, 11: 338–67.
Gabbatt, Adam. 2017. ‘Even Rightwing Sites Call Out Trump Administration over
“Alternative Facts” ’. Guardian, 24 Jan.; available at <https://www.theguardian.com/
us-news/2017/jan/23/rightwing-sites-trump-administration-alternative-facts>.
Galbraith, John Kenneth. 1967/1972. The New Industrial State. 2nd edn. Harmondsworth,
Mddx.: Penguin; originally published 1967.
Galston, William A. 2001. ‘Political Knowledge, Political Engagement and Civic
Education’. Annual Review of Political Science, 4: 217–34.
Galston, William A. 2012. ‘Truth and Democracy: Theme and Variations’. Pp. 130–45 in
Truth and Democracy, ed. Jeremy Elkins and Andrew Norris. Philadelphia, PA:
University of Pennsylvania Press.
References 389
Galton, Francis. 1907a. ‘The Ballot-Box’. Nature, 75: 509–10.

Galton, Francis. 1907b. ‘One Vote, One Value’. Nature, 75: 414.
Galton, Francis. 1907c. ‘Vox Populi’. Nature, 75: 450–1.
Garrett, Neil; Stephanie C. Lazzaro; Dan Ariely; and Tali Sharot. 2016. ‘The Brain Adapts
to Dishonesty’. Nature Neuroscience, 19: 1727–32.
Gastil, John and James P. Dillard. 1999. ‘Increasing Political Sophistication through
Public Deliberation’. Political Communication, 16: 3–23.
Gastil, John and Peter Levine, eds. 2005. The Deliberative Democracy Handbook: Strategies
for Effective Civic Engagement in the Twenty-first Century. New York: Jossey-Bass.
Gastil, John; E. Pierre Deess; Philip J. Weiser; and Cindy Simmons. 2010. The Jury
and Democracy: How Jury Deliberation Promotes Civic Engagement and Political
Participation. Oxford: Oxford University Press.
Gastil, John; Robert Richards; and Katherine R. Knobloch. 2014. ‘Vicarious Deliberation:
How the Oregon Citizens Initiative Review Influenced Deliberation in Mass Elections’.
International Journal of Communication, 8: 62–89.
Gaus, Gerald. 1995. Justificatory Liberalism: An Essay on Epistemology and Political Theory.
Oxford: Clarendon Press.
Gaus, Gerald. 1997. ‘Does Democracy Reveal the Voice of the People? Four Takes on
Rousseau’. Australasian Journal of Philosophy, 75: 141–62.
Gaus, Gerald. 2016. The Tyranny of the Ideal: Justice in a Diverse Society. Princeton,
NJ: Princeton University Press.
Geertz, Clifford. 1964. ‘Ideology as a Cultural System’. Pp. 47–76 in Ideology and Discontent,
ed. David Apter. New York: Free Press.
Gehrlein, William. 1983. ‘Condorcet’s Paradox’. Theory & Decision, 15: 161–97.
Gerber, Elisabeth R. and Arthur Lupia. 1999. ‘Voter Competence in Direct Legislation
Elections’. Pp. 147–60 in Democracy and Citizen Competence, ed. Steven L. Elkin and
Karol E. Soltan. University Park, PA: Penn State Press.
Gersen, Jacob E. and Adrian Vermeule. 2007. ‘Chevron as a Voting Rule’. Yale Law
Journal, 116: 676–731.
Gersen, Michael. 2017. ‘Reality Will Get its Revenge on Donald Trump’. Washington
Post, 16 Feb.; available at <https://www.washingtonpost.com/opinions/reality-
will-get-its-revenge-on-donald-trump/2017/02/16/b24eaea6-f46d-11e6-8d72-
263470bf0401_story.html?hpid=hp_no-name_opinion-card-c%3Ahomepage%
2Fstory&utm_term=.28120022db7f>.
Gettier, Edmund L. 1963. ‘Is Justified True Belief Knowledge?’ Analysis, 23: 121–3.
Gibbard, Allan. 1973. ‘Manipulation of Voting Schemes: A General Result’. Econometrica,
41: 587–601.
Gibson, Caitlin. 2017. ‘What We Talk About when We Talk About Donald Trump and
“Gaslighting” ’. Washington Post, 27 Jan.; available at <https://www.washingtonpost.
com/lifestyle/style/what-we-talk-about-when-we-talk-about-donald-trump-and-
gaslighting/2017/01/27/b02e6de4-e330-11e6-ba11-63c4b4fb5a63_story.html?hpid=
hp_hp-more-top-stories_gaslight-3pm%3Ahomepage%2Fstory&utm_term=
.9e4a2aaec61b>.
Gigerenzer, Gerd. 2008. Rationality for Mortals: How People Cope with Uncertainty.
New York: Oxford University Press.
Gigerenzer, Gerd; Peter M. Todd; and the ABC Research Group. 1999. Simple Heuristics
that Make Us Smart. New York: Oxford University Press.
390 References
Gigone, Daniel and Reid Hastie. 1993. ‘The Common Knowledge Effect: Information
Sharing and Group Judgments’. Journal of Personality and Social Psychology,
65: 959–74.
Gillespie, Charles Coulston. 1972. ‘Probability and Politics: Laplace, Condorcet and
Turgot’. Proceedings of the American Philosophical Society, 116: 1–20.
Gillette, Felix. 2016. Tabloid’s Shocking LOVE AFFAIR with Trump Revealed! Bloomberg,
29 Sept.; available at <https://www.bloomberg.com/features/2016-trump-national-
enquirer/>.
Gingrich, Newt. 2016. ‘The Principles of Trumpism’. Heritage Foundation, 13 Dec.
Available at <https://www.c-span.org/video/?c4637292/newt-gingrich-heritage-
foundation>.
Gintis, Herbert. 1972a. ‘A Radical Analysis of Welfare Economics and Individual
Development’. Quarterly Journal of Economics 68: 572–99.
Gintis, Herbert. 1972b. ‘Consumer Behavior and the Concept of Sovereignty: Explanations
of Social Decay’. American Economic Review (Papers & Proceedings), 62 (2): 267–78.
Goren, Paul. 2013. On Voter Competence. Oxford: Oxford University Press.
Godfrey, Elaine. 2016. ‘Newt Gingrich Tries to Explain How Trump Will Remake
America’. Atlantic, 17 Dec.; available at <https://www.theatlantic.com/politics/
archive/2016/12/gingrich-trump-election/510983/>.
Goeree, Jacob K. and Leeat Yariv. 2011. ‘An Experimental Study of Collective
Deliberation’. Econometrica, 79: 893–921.
Goldman, Alvin I. 1999. Knowledge in a Social World. Oxford: Clarendon Press.
Goldman, Alvin I. 2001. ‘Experts: Which Ones Should You Trust?’ Philosophy and
Phenomenological Research, 63: 85–110.
Goldman, Alvin I. and Thomas Blanchard. 2015. ‘Social Epistemology’. Stanford
Encyclopedia of Philosophy. Available at <https://plato.stanford.edu/entries/
epistemology-social/>.
Goodin, Robert E. 1980. Manipulatory Politics. New Haven, CT: Yale University
Press.
Goodin, Robert E. 1983. ‘Voting through the Looking Glass’. American Political Science
Review, 77: 420–34.
Goodin, Robert E. 1985. ‘Erring on the Side of Kindness in Social Welfare Policy’. Policy
Sciences, 18: 141–56.
Goodin, Robert E. 1993. ‘Independence in Democratic Theory: A Virtue? A Necessity?
Both? Neither?’ Journal of Social Philosophy, 24: 50–7.
Goodin, Robert E. 1996. ‘Institutionalizing the Public Interest: The Defense of Deadlock
and Beyond’. American Political Science Review, 90: 331–43.
Goodin, Robert E. 2000. ‘Democratic Deliberation Within’. Philosophy & Public Affairs,
29: 79–107.
Goodin, Robert E. 2003. Reflective Democracy. Oxford: Oxford University Press.
Goodin, Robert E. 2007a. ‘Enfranchising All Affected Interests, and Its Alternatives’.
Philosophy & Public Affairs, 35: 40–68.
Goodin, Robert E. 2007b. ‘Liberal Multiculturalism: Protective and Polyglot’. Political
Theory, 34: 289–303.
Goodin, Robert E. 2008. Innovating Democracy. Oxford: Oxford University Press.
Goodin, Robert E. 2009. ‘Rationalizing Discursive Anomalies’. Theoria, 119: 1–13.
References 391
Goodin, Robert E. 2010. ‘Perverting the Course Of Politics’. British Journal of Political
Science, 40: 725–39.
Goodin, Robert E. 2012. On Settling. Princeton, NJ: Princeton University Press.
Goodin, Robert E. 2017. ‘The Epistemic Benefits of Deliberative Democracy’. Policy
Sciences, 50: 351–66.
Goodin, Robert E. and John S. Dryzek. 2006. ‘Deliberative Impacts: The Macro-political
Uptake of Mini-publics’. Politics & Society, 34: 219–44.
Goodin, Robert E. and Christian List. 2006. ‘Special Majorities Rationalized’. British
Journal of Political Science, 36: 213–42.
Goodin, Robert E. and K. W. S. Roberts. 1975. ‘The Ethical Voter’. American Political
Goodin, Robert E. and Michael Saward. 2005. ‘Dog Whistles and Democratic Mandates’.
Political Quarterly, 76: 471–6.
Goodin, Robert E. and Kai Spiekermann. 2012. ‘Epistemic Aspects of Representative
Government’. European Political Science Review, 4: 303–25.
Goodin, Robert E. and Kai Spiekermann. 2015. ‘Epistemic Solidarity as a Political
Strategy’. Episteme, 12: 439–57.
Goren, Paul. 2013. On Voter Competence. Oxford: Oxford University Press.
Gramsci, Antonio. 1971. Selections from the Prison Notebooks, ed. and trans. Q. Hoare
and G. Nowell Smith. London: Lawrence & Wishart.
Greider, William. 1988. ‘Is the Democrats’ New Harmony for Real?’ Rolling Stone,
534 (8 September); available at <http://www.rollingstone.com/politics/news/is-the-
democrats-new-harmony-for-real-19880908>.
Grice, Andrew. 2017. ‘Fake News Handed Brexiters the Referendum—and Now They
Have No Idea what They’re Doing’. Independent, 19 Jan.; available at <http://www.
independent.co.uk/voices/michael-gove-boris-johnson-brexit-eurosceptic-press-
theresa-may-a7533806.html>.
Grice, Paul. 1975. ‘Logic and Conversation’. Pp. 64–75 in The Logic of Grammar,
ed. Donald Davidson and Gilbert Harman. Encino, Calif: Dickenson Publishing Co.
Griffin, Andrew. 2016. ‘Brexit: Vote Leave Wipes NHS £350m Claim and Rest of Its Website
after EU Referendum’. Independent, 27 June; available at <http://www.independent.
co.uk/news/uk/home-news/brexit-vote-leave-wipes-nhs-350m-claim-and-rest-
of-its-website-after-eu-referendum-a7105546.html>.
Grim, Patrick; Daniel J. Singer; Steven Fisher; Aaron Bramson; William J. Berger;
Christopher Reade; Carissa Flocken; and Adam Sales. 2013. ‘Scientific Networks on
Data Landscapes: Question of Difficulty, Epistemic Success and Convergence’. Episteme,
10: 441–64.
Grofman, Bernard. 1975. ‘A Comment on “Democratic Theory: A Preliminary
Mathematical Model” ’. Public Choice, 21: 99–103.
Grofman, Bernard and Scott L. Feld. 1988. ‘Rosseau’s General Will: A Condorcetian
Perspective’. American Political Science Review, 82: 567–76.
Grofman, Bernard; Guillermo Owen; and Scott L. Feld. 1982. ‘Average Competence,
Individual Competence, and Group Judgmental Accuracy’. Psychology Report,
50: 683–8.
Grofman, Bernard; Guillermo Owen; and Scott L. Feld. 1983. ‘Thirteen Theorems in
Search of the Truth’. Theory & Decision, 15: 261–78.
392 References
Grundahl, Johs. 1995. ‘The Danish Consensus Conference Model’. In Public Participation
in Science: The Role of Consensus Conferences in Europe, ed. S. Joss and J. Durant.
London: Science Museum. <http://people.ucalgary.ca/~pubconf/Education/grundahl.
htm>.
Grynbaum, Michael M. 2017a. ‘Trump Calls the News Media the “Enemy of the American
People” ’. New York Times, 17 Feb.; available at <https://www.nytimes.com/2017/02/17/
business/trump-calls-the-news-media-the-enemy-of-the-people.html?_r=0>.
Grynbaum, Michael M. 2017b. ‘Trump Strategist Stephen Bannon Says Media Should
“Keep Its Mouth Shut” ’. New York Times, 26 Jan.; available at <https://www.nytimes.
com/2017/01/26/business/media/stephen-bannon-trump-news-media.html>.
Habermas, Jürgen. 1984. Reason and the Rationalization of Society, vol. 1 of The Theory
of Communicative Action, trans. Thomas McCarthy. Boston, MA: Beacon.
Habermas, Jürgen. 1990. ‘Discourse Ethics’. Pp. 43–115 in Moral Consciousness and
Communicative Action, trans. Shierry Weber Nicholsen and Christian Lenhardt.
Cambridge, MA: MIT Press.
Habermas, Jürgen. 1996. Between Facts and Norms, trans. William Rehg. Oxford:
Polity.
Habermas, Jürgen. 2006. ‘Political Communication in Media Society: Does Democracy
Still Enjoy an Epistemic Dimension? The Impact of Normative Theory on Empirical
Research’. Communication Theory, 16: 411–26.
Habermas, Jürgen. 2008. Between Nationalism & Religion, trans. C. Cronin. Cambridge:
Polity; originally published 2005.
Hafer, Catherine and Dimitri Landa. 2007. ‘Deliberation as Self-discovery and
Institutions for Political Speech’. Journal of Theoretical Politics, 19: 329–60.
Hale, Matthew. 1716/1971. The History of the Common Law of England, 2nd edn, ed. Charles
M. Gray. Chicago: University of Chicago Press, 1971; originally published 1716.
Halloway, Daniel. 2016. ‘Why TV News Couldn’t Quit Donald Trump’. Variety, 9 Nov.;
available at <http://variety.com/2016/tv/news/donald-trump-tv-news-1201914147/>.
Hamilton, Alexander. 1788/2003. Federalist no. 68. Pp. 331–4 in Hamilton, Madison ad
Jay, The Federalist, with Letters of ‘Brutus’, ed. Terence Ball. Cambridge: Cambridge
University Press, 2003; originally published 1788.
Hamilton, Alexander. 1788/2003. Federalist no. 73. Pp. 356–62 in Hamilton, Madison
ad Jay, The Federalist, with Letters of ‘Brutus’, ed. Terence Ball. Cambridge: Cambridge
Hamilton, Alexander; James Madison; and John Jay. 1787–1788/2003. The Federalist,
with Letters of ‘Brutus’, ed. Terence Ball. Cambridge: Cambridge University Press,
2003; originally published 1787–1788.
Hammond, Thomas H. and Brian D. Humes. 1993. ‘ “What This Campaign Is All About
Is . . .”: A Rational Choice Alternative to the Downsian Spatial Model of Elections’.
Pp. 141–60 in Information, Participation and Choice, ed. Bernard Grofman. Ann
Arbor. MI: University of Michigan Press.
Hanft, Adam. 2016. ‘140-Character Catharsis: The Secret Appeal of Trump’s Rust Belt
Strategy’. Huffington Post, 5 Dec.; available at <http://www.huffingtonpost.com/
adam-hanft/140-character-catharsis-t_b_13441064.html>.
Hanska, Jan. 2012. Reagan’s Mythical America: Storytelling as Political Leadership.
New York: Palgrave Macmillan.
References 393
Hanson, Robin. 2013. ‘Shall We Vote on Values, but Bet on Beliefs?’ Journal of Political
Hardin, Russell. 2006. ‘Ignorant Democracy’. Critical Review, 18: 179–95.
Hardin, Russell. 2009. How Do You Know? The Economics of Ordinary Knowledge.
Princeton, NJ: Princeton University Press.
Hardy, Bruce W. 2017. ‘Candidate Traits and Political Choice’. Pp. 437–51 in Oxford
Handbook of Political Communication, ed. Kate Kenski and Kathleen Hall Jamieson.
Oxford: Oxford University Press.
Harford, Tim. 2017. ‘The Problem with Facts’. Financial Times, 9 Mar.; available at
<https://www.ft.com/content/eef2e2f8-0383-11e7-ace0-1ce02ef0def9>.
Harlan, John Marshall. 1970. Concurring opinion. In re Winship. 397 U.S., 368–75.
Harsanyi, John C. 1969. ‘Rational-choice Models of Political Behavior vs. Functionalist
and Conformist Theories’. World Politics, 21: 513–38.
Harsanyi, John C. 1975. ‘Can the Maximin Principle Serve as a Basis for Morality?’
Harsanyi, John C. 1982. ‘Morality and the Theory of Rational Behaviour’. Pp. 39–62 in
Utilitarianism and Beyond, ed. Amartya Sen and Bernard Williams. Cambridge:
Hart, H. L. A. 1961. The Concept of Law. Oxford: Clarendon Press.
Hawthorne, James. 2014. ‘Voting in Search of the Public Good: The Probabalistic Logic
of Majority Judgements’. Unpublished ms., Department of Philosophy, University
of Oklahoma; available at <http://faculty-staff.ou.edu/H/James.A.Hawthorne-1/
Hawthorne--Jury-Theorems.pdf>.
Hayek, Friedrich A. 1945. ‘The Use of Knowledge in Society’. American Economic
Review, 35: 519–30.
Hayek, Friedrich A. 1958. ‘Freedom, Reason and Tradition’. Ethics, 6: 229–45.
Hayek, Friedrich A. 1960. The Constitution of Liberty. London: Routledge & Kegan Paul.
He, Baogang and Mark E. Warren. 2011. ‘Authoritarian Deliberation: The Deliberative
Turn in Chinese Political Development’. Perspectives on Politics, 9: 269–89.
Hedden, Brian R. 2017. ‘Should Juries Deliberate?’ Social Epistemology, 31: 368–86.
Heim, Joe. 2017. ‘ “They Never Saw this Coming”: A Q&A with Kellyanne Conway’.
Washington Post, 26 Jan.; available at <https://www.washingtonpost.com/lifestyle/
magazine/they-never-saw-this-coming-a-qanda-with-kellyanne-conway-on-
trumps-victory/2017/01/26/2bf64c10-da96-11e6-9a36-1d296534b31e_story.html?
hpid=hp_hp-top-table-main_conway-1130pm%3Ahomepage%2Fstory&utm_
term=.fd286ff9001a>.
Heller, Steven. 2014. ‘Bat Boy, Hillary Clinton’s Alien Baby, and a Tabloid’s Glorious
Legacy’. Atlantic, 16 Oct.; available at <http://www.theatlantic.com/entertainment/
archive/2014/10/the-ingenious-sensationalism-of-the-weekly-world-new/381525/>.
Hermann, Peter; Susan Svrluga; and Michael E. Miller. 2016. ‘Alleged Gunman Tells
Police He Wanted to Rescue Children at D.C. Pizza Shop after Hearing Fictional
Internet Accounts’. Washington Post, 5 Dec.; available at <https://www.washingtonpost.
com/local/public-safety/alleged-gunman-tells-police-he-wanted-to-rescue-children-
at-dc-pizza-shop-after-hearing-fictional-internet-accounts/2016/12/05/cb5ebabc-
bae8-11e6-ac85-094a21c44abc_story.html?hpid=hp_hp-top-table-main_no-
name%3Ahomepage%2Fstory&utm_term=.aa30c6bc523f>.
394 References
Hiatt, Fred. 2017. ‘Trump Considers the Media His Enemy. We Shouldn’t Treat Him
as Ours’. Washington Post, 29 Jan.; available at <https://www.washingtonpost.com/
opinions/trump-considers-the-media-his-enemy-we-shouldnt-treat-him-as-
ours/2017/01/29/43ac0702-e4b2-11e6-ba11-63c4b4fb5a63_story.html?hpid=hp_no-
name_opinion-card-b%3Ahomepage%2Fstory&utm_term=.477a3d0e250f>.
Hill, Lisa. 2016. ‘Voting Turnout, Equality, Liberty and Representation: Epistemic
versus Procedural Democracy’. Critical Review of International Social and Political
Hirst, Paul. 1994. Associative Democracy. Oxford: Polity.
Hirst, Paul Q., ed. 1989. The Pluralist Theory of the State: Selected Writings of G. D. H. Cole,
J. N. Figgis and H. J. Laski. London: Routledge.
Hobolt, Sarah and Thomas Leeper. 2017. ‘The British are Indifferent about Many
Aspects of Brexit, but Leave and Remain Voters are Divided on Several Issues’.
Available at: <http://blogs.lse.ac.uk/brexit/2017/08/13/the-british-are-indifferent-about-
many-aspects-of-brexit-but-leave-and-remain-voters-are-divided-on-several-key-
issues/>.
Hobsbawm, Eric and Terence Ranger, eds. 1992. The Invention of Tradition. Cambridge:
Hochschild, Arlie Russell. 2016. Strangers in Their Own Land: Anger and Mourning on
the American Right. New York: New Press.
Hochschild, Jennifer L. 2012. ‘Should the Mass Public Follow Elite Opinion? It
Depends . . .’ Critical Review, 24: 527–44.
Hochschild, Jennifer L. and Katherine Levine Einstein. 2015a. Do Facts Matter?
Information and Misinformation in American Politics. Norman, OK: University of
Oklahoma Press.
Hochschild, Jennifer L. and Katherine Levine Einstein. 2015b. ‘Do Facts Matter?
Information and Misinformation in American Politics’. Political Science Quarterly,
130: 585–624.
Hoeffding, W. 1956. ‘On the Distribution of the Number of Successes in Independent
Trials’. Annals of Mathematical Statistics, 27: 713–21.
Holmes, Oliver Wendall. 1897. ‘The Path of the Law’. Harvard Law Review, 10: 457–78.
Hong, Lu and Scott E. Page. 2004. ‘Groups of Diverse Problem Solvers Can Outperform
Groups of High-ability Problem Solvers’. Proceedings of the National Academy of
Sciences, 101: 16,385–9.
Hong, Lu and Scott E. Page. 2009. ‘Interpreted and Generated Signals’. Journal of Economic
Theory, 144: 2174–97.
Hong, Lu and Scott E. Page. 2012. ‘Some Microfoundations of Collective Wisdom’.
Pp. 56–71 in Collective Wisdom: Principles and Mechanisms, ed. Hélène Landemore
and Jon Elster. Cambridge: Cambridge University Press.
Huckfeldt, Robert and John Sprague. 1995. Citizens, Politics and Social Communication.
Cambridge: Cambridge University Press.
Huddy, Leonie. 2013. ‘From Group Identity to Political Cohesion and Commitment’.
Pp. 737–73 in Oxford Handbook of Political Psychology, ed. Leonie Huddy, David O.
Sears and Jack S. Levy. 2nd edn. Oxford: Oxford University Press.
Humberstone, I. L. 1992. ‘Direction of Fit’. Mind, 101: 59–83.
Hume, David. 1739. A Treatise of Human Nature. London: John Noon.
References 395
Hume, David. 1754/1760. ‘Idea of a Perfect Commonwealth’. Essay 16 in Essays, Literary,

Moral and Political. London: A. Millar; originally published 1754.
Hume, David. 1777. ‘Of Miracles’. Section 10 of Enquiry Concerning Human Understanding.
London: T. Cadell.
Hurd, Heidi. 1991. ‘Challenging Authority’. Yale Law Journal, 100: 1611–77.
Ignatius, David. 2016. ‘In Today’s World, the Truth is Losing’. Washington Post, 29 Nov.;
available at <https://www.washingtonpost.com/people/david-ignatius/?utm_term=.
ad642900d112>.
Iowa Electronic Markets (IEM). 2016. Data: 2016 US Presidential Election Markets.
Available at <https://tippie.biz.uiowa.edu/iem/markets/data_Pres16.html>.
Ipsos MORI. 2016. Political monitor, Topline results, 16 June 2016. Available at <https://
www.ipsos.com/sites/default/files/migrations/en-uk/files/Assets/Docs/Polls/pm-
16-june-2016-topline.pdf>.
Isaac, Mike. 2016. ‘Facebook Mounts Effort to Limit Tide of Fake News’. New York Times,
15 Dec.; available at <https://www.nytimes.com/2016/12/15/technology/facebook-
fake-news.html>.
Isenberg, Daniel J. 1986. ‘Group Polarization: A Critical Review and Meta-analysis’.
Journal of Personality and Social Psychology, 50: 1141–51.
Isenberg, Nancy. 2016. White Trash: The 400-year Untold History of Class in America.
New York: Viking.
Islam, Faisal. 2016a. ‘Can Wollaston Call £350m Brexit Claim a Lie?’ Sky News, 9 June;
available at <http://news.sky.com/story/can-wollaston-call-163350m-brexit-claim-a-
lie-10308489>.
Islam, Faisal. 2016b. ‘EU: In or Out?’ Faisal Islam interview with Michael Gove, 3 June
2016. Sky News, 3 June; available at <https://corporate.sky.com/media-centre/media-
packs/2016/eu-in-or-out-faisal-islam-interview-with-michael-gove,-30616-8pm>.
ITV. 2016. ‘UK Does Get Back Some of the £350m It Sends to EU, Boris Johnson Admits’.
ITV Report, 11 May; available at <http://www.itv.com/news/2016-05-11/uk-does-get-
back-some-of-350m-it-sends-to-eu-boris-johnson-admits/>.
Iyengar, Shanto. 1990. ‘Shortcuts to Political Knowledge: The Role of Selective Attention
and Accessibility’. In Information and Democratic Processes, ed. John A. Ferejohn and
James H. Kuklinski. Urbana, IL: University of Illinois Press.
Jackman, Simon and Paul M. Sniderman. 2002. ‘The Institutional Organization of
Choice Spaces: A Political Conception of Political Psychology’. Pp. 209–24 in Political
Psychology, ed. Kristen Monroe. Mahwah, NJ: Lawrence Erlbaum.
Jackson, Jasper. 2017. ‘In the Post-truth Era Sweden’s Far Right Fake Fact Checker was
Inevitable’. Guardian, 19 Jan.; available at <https://www.theguardian.com/media/2017/
jan/19/in-the-post-truth-era-swedens-far-right-fake-fact-checker-was-inevitable>.
Jacobs, Jennifer and Billy House. 2016. ‘Trump: Polls made Me Think “We’re Not Going
to Win” ’. Bloomberg News/Newsmax, 14 Dec.; available at <http://www.newsmax.
com/Politics/trump-polls-election-lose/2016/12/14/id/763840/>.
Jacobs, Lawrence R.; Fay Lomax Cook; and Michael X. Delli Carpini. 2009. Talking
Together: Public Deliberation and Political Participation in America. Chicago: University
of Chicago Press.
Jacobson, Gary C. 2010. ‘Perception, Memory and Partisan Polarization on the Iraq
War’. Political Science Quarterly, 125: 31–56.
396 References
James, Aaron. 2016. Assholes: A Theory of Donald Trump. New York: Doubleday.
Jamieson, Amber and Olivia Solon. 2016. ‘Facebook to Begin Flagging Fake News in
Response to Mounting Criticism’. Guardian, 16 Dec.; available at <https://www.
theguardian.com/technology/2016/dec/15/facebook-flag-fake-news-fact-check>.
Jamieson, Kathleen Hall and Joseph N. Cappella. 2008. Echo Chamber: Rush Limbaugh
and the Conservative Media Establishment. New York: Oxford University Press.
Janis, Irving L. 1972. Groupthink. Boston, MA: Houghton Mifflin.
Jefferson, Thomas. 1785. Notes on Virginia. New York: Harper & Row, 1964; originally
published 1785.
Jefferson, Thomas. 1816. Letter to Samuel Kercheval, June 12, 1816. Available at <http://
teachingamericanhistory.org/library/document/letter-to-samuel-kercheval/>.
Jeffrey, Anne. 2017. ‘Limited Epistocracy and Political Inclusion’. Episteme, forthcom-
ing, doi:10.1017/epi.2017.8.
Johnson, Jeena. 2017. ‘In His First Major TV interview as President, Trump is Obsessed
with His Popularity’. Washington Post, 26 Jan.; available at <https://www.washingtonpost.
com/news/post-politics/wp/2017/01/26/in-his-first-major-tv-interview-as-president-
trump-is-endlessly-obsessed-about-his-popularity/?hpid=hp_hp-top-table-main_
pp-trumptv-330am%3Ahomepage%2Fstory&utm_term=.893744c74691>.
Johnson, Jenna and Matt Zapotosky. 2017. ‘Trump Seeks ‘Major Investigation’ into
Unsupported Claims of Voter Fraud’. Washington Post, 25 Jan.; available at <https://
www.washingtonpost.com/news/post-politics/wp/2017/01/25/trump-seeks-major-
investigation-into-unsupported-claims-of-voter-fraud/?hpid=hp_hp-top-table-
main_investigation-825a%3Ahomepage%2Fstory&utm_term=.6cce33110c2a>.
Jones, Chris. 2017. ‘Alec Baldwin’s Quest to Conquer Trump’. Atlantic, 319 (no. 4: May): 46–58.
Jones, Susan. 2017. ‘Paul Ryan Views Trump as “A Man of Action, Ready to Get Working’.
CBS News, 11 Nov.; available at <http://www.cnsnews.com/news/article/susan-jones/
paul-ryan-views-trump-man-actionready-get-working>.
Jütten, Timo. 2017. ‘Dignity, Esteem and Social Contribution: A Recognition-theoretical
View’. Journal of Political Philosophy, 25: 259–80.
Kahn, Chris. 2016. ‘More Republicans Expect Clinton, Rather than Trump, to Win US
Election’. Reuters, 26 Oct.; available at <http://www.reuters.com/article/us-usa-election-
poll-idUSKCN12Q2WK>.
Kahneman, Daniel; Paul Slovic; and Amos Tversky, eds. 1982. Judgment Under Uncertainty:
Heuristics and Biases. Cambridge: Cambridge University Press.
Kaniovski, Serguei. 2010. ‘Aggregation of Correlated Votes and Condorcet’s Jury
Theorem’. Theory and Decision, 69: 453–68.
Kant, Immanuel. 1784. ‘What is Enlightenment?’ Pp. 132–9 in The Philosophy of Kant:
Immanuel Kant’s Moral and Political Writings, ed. Carl J. Friedrich. New York: Modern
Library, 1949.
Kant, Immanuel. 1785. Foundations of the Metaphysics of Morals, trans. L. W. Beck.
Chicago, IL: University of Chicago Press, 1949.
Karotkin, Drora and Jacob Paroush. 2003. ‘Optimum Committee Size: Quality-versus-
Quantity Dilemma’. Social Choice and Welfare, 20 (3): 429–41.
Kassam, Ashifa. 2017. ‘ “You Scare the Heck out of Me”: Twitter Feed Collects Trump
Voters’ Regrets’. Guardian, 5 Feb.; available at <https://www.theguardian.com/us-news/
2017/feb/05/trump-regrets-twitter-voters-election-ashamed>.
Katz, Elihu and Paul Lazarsfeld. 1955. Personal Influence. New York: Free Press.
References 397
Katz, Richard S. 1980. A Theory of Parties and Electoral Systems. Baltimore, MD: John
Hopkins University Press.
Keane, John. 1998. The Media and Democracy. Cambridge: Polity.
Keillor, Garrison. 2016a. ‘Trump Voters—It’s Not Me, It’s You’. Washington Post, 21 Nov.;
available at <https://www.washingtonpost.com/opinions/trump-voters--its-not-me-
its-you/2016/11/21/92d7c884-b005-11e6-be1c-8cec35b1ad25_story.html?utm_term=
.4e48c9282626>.
Keillor, Garrison. 2016b. ‘Trump Voters Will Not Like What Happens Next’. Washington
Post, 9 Nov.; available at <https://www.washingtonpost.com/opinions/trump-voters-
will-not-like-what-happens-next/2016/11/09/e346ffc2-a67f-11e6-8fc0-7be8f848c492_
story.html?utm_term=.8a225db695e3>.
Kelly, Andrew. 2016. ‘Why I Voted for Donald Trump’. Washington Post, 13 Nov.; avail-
able at <https://www.pressreader.com/usa/the-washington-post-sunday/20161113/
281784218671355>.
Kelly, Paul. 2009. The March of Patriots: The Struggle for Modern Australia. Melbourne:
Melbourne University Press.
Kelman, Mark. 2011. The Heuristics Debate. New York: Oxford University Press.
Kessler, Glenn. 2013. ‘About the Fact Checker’. Washington Post, 11 Sept.; available at
<https://www.washingtonpost.com/news/fact-checker/about-the-fact-checker/
?utm_term=.34e82f750bba>.
Kessler, Glenn. 2015. ‘Trump’s Absurd Claim that the “Real” Unemployment Rate is
42 Percent’. Washington Post, 18 Aug.; available at <https://www.washingtonpost.com/
news/fact-checker/wp/2015/08/21/trumps-absurd-claim-the-real-unemployment-
rate-is-42-percent/?tid=a_inl&utm_term=.bb95d53ec4f1>.
Kessler, Glenn. 2016a. ‘The Biggest Pinocchios of 2016’. Washington Post, 16 Dec.;
available at <https://www.washingtonpost.com/news/fact-checker/wp/2016/12/16/
the-biggest-pinocchios-of-2016/?hpid=hp_hp-more-top-stories_factchecker-
2016roundup-310am%3Ahomepage%2Fstory&utm_term=.4226cac14da4>.
Kessler, Glenn. 2016b. ‘The Fact Checker’s Guide for Detecting Fake News’. Washington
Post, 22 Nov.; available at <https://www.washingtonpost.com/news/fact-checker/wp/
2016/11/22/the-fact-checkers-guide-for-detecting-fake-news/?tid=a_inl&utm_
term=.10749890021d>.
Kessler, Glenn. 2017a. ‘Recidivism Watch: Trump’s Claim that Millions of People Voted
Illegally’. Washington Post, 24 Jan.; available at <https://www.washingtonpost.com/
news/fact-checker/wp/2017/01/24/recidivism-watch-trumps-claim-that-3-5-million-
people-voted-illegally-in-the-election/?utm_term=.e43104b3b61d>.
Kessler, Glenn. 2017b. ‘Spicer Earns Four Pinocchios for a Series of False Claims on
Inauguration Crowd Size’. Washington Post, 22 Jan.; available at <https://www.
washingtonpost.com/news/fact-checker/wp/2017/01/22/spicer-earns-four-pinocchios-
for-a-series-of-false-claims-on-inauguration-crowd-size/?hpid=hp_hp-top-table-
main_fc-spicer-1203pm%3Ahomepage%2Fstory&utm_term=.02ec8eae74c1>.
Kessler, Glenn and Michelle Ye Hee Lee. 2016. ‘Fact-checking Donald Trump’s Acceptance
Speech at the 2016 RNC’. Washington Post, 22 July; available at <https://www.wash-
ingtonpost.com/news/fact-checker/wp/2016/07/22/fact-checking-donald-trumps-
acceptance-speech-at-the-2016-rnc/?utm_term=.c32488ea2c17>.
Kessler, Glenn and Michelle Ye Hee Lee. 2017. ‘President Trump’s Cascade of False
Claims in Time’s Interview on His Falsehoods’. Washington Post, 23 Mar.; available at
398 References
<https://www.washingtonpost.com/news/fact-checker/wp/2017/03/23/president-
trumps-cascade-of-false-claims-in-times-interview-on-his-falsehoods/?hpid=hp_
hp-more-top-stories_fact-checker-940am%3Ahomepage%2Fstory&utm_term=
.b1cdc60e5421>.
Kessler, Glenn; Michelle Ye Hee Lee; Leslie Shapiro; and Aaron Williams. 2016. ‘The
2016 Election Fact Checker’. Washington Post, 3 Nov.; available at <https://www.
washingtonpost.com/graphics/politics/2016-election/fact-checker/?tid=a_inl>.
Keuschnigg, Marc and Christian Ganser. 2016. ‘Crowd Wisdom Relies on Agents’
Ability in Small Groups with a Voting Aggregation Rule’. Management Science,
63: 818–28.
Keynes, John Maynard. 2007. The General Theory of Employment, Interest, and Money.
Basingstoke: Palgrave Macmillan; originally published 1936.
Killick, Ana. 2017. ‘Do People Really Lack Knowledge about the Economy?’ Political
Quarterly, 88: 265–72.
Kim, Sung-youn; Charles S. Taber; and Milton Lodge. 2010. ‘A Computational Model of
the Citizen as Motivated Reasoner: Modeling the Dynamics of the 2000 Presidential
Election’. Political Behavior, 32: 1–28.
Kinder, Donald. 2006. ‘Belief Systems Today’. Critical Review, 18: 197–215.
King, Anthony and Ivor Crewe. 2013. The Blunders of Our Governments. London:
Oneworld.
Kirchheimer, Otto. 1966. ‘The Transformation of the Western European Party System’.
Pp. 5–28 in Political Parties and Political Development, ed. J. LaPalombara and
M. Weiner. Princeton, NJ: Princeton University Press.
Kirzinger, Ashley; Elise Sugarman; and Mollyann Brodie. 2016. Kaiser Health Tracking
Poll: November 2016. Available at <http://kff.org/health-reform/poll-finding/kaiser-
health-tracking-poll-november-2016/>.
Kitcher, Philip. 1990. ‘The Division of Cognitive Labor’. Journal of Philosophy, 87: 5–22.
Kizenko, Peter. 2016. ‘Uncensored: An Inside Look at a Donald Trump Rally’. The
Duran, 24 Oct.; available at <http://theduran.com/uncensored-an-inside-look-at-a-
donald-trump-rally/>.
Klingemann, Hans-Dieter; Richard I. Hofferbert; and Ian Budge. 1994. Parties, Policies
and Democracy. Boulder, CO: Westview.
Knight, Jack and Lee Epstein. 1996. ‘The Norm of Stare Decisis’. American Journal of
Knippenberg, Daan van and Michaéla C. Schippers. 2007. ‘Work Group Diversity’.
Annual Review of Psychology, 58: 515–41.
Kouser, J. Morgan. 1984. ‘Suffrage’. Vol. 3, pp. 1236–58, in Encyclopedia of American
Political History, ed. J. P. Greene. New York: Scribner.
Kovalev, Alexey. 2017. ‘I’ve Reported on Putin – Here Are My Tips for Journalists
Dealing with Trump’. Guardian, 23 Jan.; available at <https://www.theguardian.com/
commentisfree/2017/jan/23/reported-putin-journalists-trump-media>.
Kremer, Ilan; Yishay Mansour; and Motty Perry. 2014. ‘Implementing the “Wisdom of
the Crowd” ’. Journal of Political Economy, 122: 988–1012.
Kruglanski, A. W. and D. M. Webster. 1996. ‘Motivated Closing of the Mind: Its Cognitive
and Social Effects’. Psychological Review, 103: 263–83.
Kuehn, Daniel. 2017. ‘Diversity, Ability and Democracy: A Note on Thompson’s
Challenge to Hong and Page’. Critical Review, 29: 72–87.
References 399
Kuklinski, James H. and Norman L. Hurley. 1994. ‘On Hearing and Interpreting Political
Messages: A Cautionary Tale of Citizen Cue-taking’. Journal of Politics, 56: 729–51.
Kuklinski, James H.; Paul J. Quirk; Jennifer Jerit; David Schwieder; and Robert F. Rich.
2000. ‘Misinformation and the Currency of Democratic Citizenship’. Journal of Politics,
62: 790–816.
Kunda, Ziva. 1990. ‘The Case for Motivated Reasoning’. Psychological Bulletin, 108: 480–98.
Kuran, Timur. 1995. Private Truths, Public Lies. Cambridge, MA: Harvard University Press.
Ladd, Everett Carl. 1998. ‘Nixon and Watergate Revisited’. The Public Perspective,
9 (3: April/May): 25–32.
Ladha, Krishna K. 1992. ‘The Condorcet Jury Theorem, Free Speech and Correlated
Votes’. American Journal of Political Science, 36: 617–34.
Ladha, Krishna K. 1993. ‘Condorcet’s Jury Theorem in Light of de Finetti’s Theorem’.
Ladha, Krishna K. 1995. ‘Information Pooling through Majority-rule Voting: Condorcet’s
Jury Theorem with Correlated Votes’. Journal of Economic Behavior & Organization,
26: 353–72.
Lagerspetz, Eerik. 2002. ‘The Legitimacy of Majority Rule’. Associations, 6 (2): 261–91.
Lagerspetz, Eerik. 2010. ‘Wisdom and Numbers’. Social Science Information, 49: 29–59.
Landa, Dimitri and Scott A. Tyson. 2017. ‘Coercive Leadership’. American Journal of
Landemore, Hélène. 2012. ‘Democratic Reason: The Mechanisms of Collective
Intelligence in Politics’. Pp. 251–89 in Collective Wisdom: Principles and Mechanisms,
ed. Hélène Landemore and Jon Elster. Cambridge: Cambridge University Press.
Landemore, Hélène. 2013a. ‘Deliberation, Cognitive Diversity and Democratic
Inclusiveness: An Epistemic Argument for the Random Selection of Representatives’.
Synthese, 190: 1209–31.
Landemore, Hélène. 2013b. Democratic Reason: Politics, Collective Intelligence, and the
Rule of the Many. Princeton, NJ: Princeton University Press.
Landemore, Hélène. 2014. ‘Yes, We Can (Make It Up on Volume): Answers to Critics’.
Critical Review, 26: 184–237.
Landemore, Hélène. 2015. ‘Inclusive Constitution-making: The Icelandic Experiment’.
Journal of Political Philosophy, 23: 166–91.
Landemore, Hélène and Scott E. Page. 2015. ‘Deliberation and Disagreement: Problem
Solving, Prediction, and Positive Dissensus’. Politics, Philosophy & Economics,
14: 229–54.
Lane, Melissa. 1999. ‘States of Nature, Epistemic and Political’. Proceedings of the
Aristotelian Society, 99: 211–24.
Lane, Melissa. 2014. ‘When the Experts are Uncertain: Scientific Knowledge and the
Ethics of Democratic Judgment’. Episteme, 11: 97–118.
Langton, Kenneth and M. Kent Jennings. 1968. ‘Political Socialization and the High
School Civics Curriculum in the United States’. American Political Science Review,
62: 852–67.
Larson, James R. 2010. In Search of Synergy in Small Group Performance. New York:
Psychology Press.
Latané, Bibb; Kipling Williams; and Stephen Harkins. 1979. ‘Many Hands Make Light
the Work: The Causes and Consequences of Social Loafing’. Journal of Personality
and Social Psychology, 37: 822–32.
400 References
Latour, Bruno and Steve Woolgar. 1979. Laboratory Life. Beverley Hills, CA: SAGE.
Lau, Richard R. and David P. Redlawsk. 1997. ‘Voting Correctly’. American Political
Lau, Richard R. and David P. Redlawsk. 2001. ‘Advantages and Disadvantages of
Cognitive Heuristics in Political Decision Making’. American Journal of Political
Science, 45: 951–71.
Lau, Richard R. and David P. Redlawsk. 2006. How Voters Decide: Information Processing
During Election Campaigns. Cambridge: Cambridge University Press.
Lau, Richard R.; Parina Patel; Dalia F. Fahmy; and Robert R. Kaufman. 2014. ‘Correct
Voting Across Thirty-three Democracies: A Preliminary Analysis’. British Journal of
Lau, Richard R.; Lee Sigelman; Caroline Heldman; and Paul Babbit. 1999. ‘The Effects
of Negative Political Advertisements: A Meta-analytic Assessment’. American Political
Laughlin, Patrick R. and Alan Ellis. 1986. ‘Demonstrability and Social Combination
Processes on Mathematical Intellective Tasks’. Journal of Experimental Social Psychology,
22: 177–89.
Layman, Geoffrey C.; Thomas M. Carsey; and Juliana Menasce Horowitz. 2006. ‘Party
Polarization in American Politics: Characteristics, Causes, and Consequences’. Annual
Review of Political Science, 9: 83–110.
Lazarsfeld, Paul F.; Bernard Berelson; and Hazel Gaudet. 1944. The People’s Choice: How
the Voter Makes Up His Mind in a Presidential Campaign. New York: Columbia
University Press.
Lee, M.J.; Sara Murray; Jeremy Diamond; Noah Gray; and Tal Kopan. 2016. ‘Why I’m
Voting for Trump: CNN Talks to More than 150 people in 31 Cities to Explore What’s
Driving the Trump Phenomenon’. CNN, 28 Jan.; available at <http://edition.cnn.
com/2016/01/27/politics/donald-trump-voters-2016-election/>.
Lee, Michelle Ye Hee. 2017. ‘Recidivism Watch: Spicer Uses Repeatedly Debunked
Citations for Trump’s Voter Fraud Claims’. Washington Post, 24 Jan.; available at
<https://www.washingtonpost.com/news/fact-checker/wp/2017/01/24/recidivism-
watch-spicer-uses-repeatedly-debunked-citations-for-trumps-voter-fraud-claims/
?utm_term=.7cd16d4f44c3>.
Lee, Michelle Ye Hee; Glenn Kessler; and Leslie Shapiro. 2017. ‘100 Days of Trump
Claims’. Washington Post, 21 Feb.; available at <https://www.washingtonpost.com/
graphics/politics/trump-claims/?utm_term=.8558219038d6>.
Lehdonvirta, Vili and Jonathan Bright. 2015. ‘Crowdsourcing for Public Policy and
Government’. Policy & Internet, 7: 263–7.
Lelkes, Yphtach; Gaurav Sood; and Shanto Iyengar. 2017. ‘The Hostile Audience: The
Effect of Access to Broadband Internet on Partisan Affect’. American Journal of
Lever, Annabelle. 2016. ‘Democracy, Epistemology and the Problem of All-white Juries’.
Journal of Applied Philosophy, forthcoming, doi: 10.1111/japp.12203.
Levin, Jonathan and Barry Nalebuff. 1995. ‘An Introduction to Vote-counting Schemes’.
Journal of Economic Perspectives, 9 (1: Winter): 3–26.
Lewallen, Jonathan. 2016. ‘Legislative Error and the “Politics of Haste” ’. PS, 49: 239–43.
Lichfield, John. 2015. ‘Man Confesses to Being Marseille’s Serial Rapist 10 Months
after DNA Tests and ID Parades Failed to Tell Him Apart from His Twin Brother’.
References 401
The Independent, 9 Sept. 2015. Available at <http://www.independent.co.uk/news/

world/europe/man-confesses-to-being-marseilles-serial-rapist-10-months-after-dna-tests-
and-id-parades-failed-to-tell-him-apart-from-his-twin-brother-10491970.html>.
Lindblom, Charles E. 1977. Politics and Markets. New York: Basic.
Lindsay, A. D. 1929. The Essentials of Democracy. London: Oxford University Press.
Lippert-Rasmussen, Kasper. 2012. ‘Estlund on Epistocracy: A Critique’. Res Publica,
18: 241–58.
List, Christian. 2004. ‘On the Significance of the Absolute Margin’. British Journal for
the Philosophy of Science, 55: 521–44.
List, Christian. 2005a. ‘Group Knowledge and Group Rationality: A Judgment
Aggregation Perspective’. Episteme, 2: 25–38.
List, Christian. 2005b. ‘The Probability of Inconsistencies in Complex Collective
Decisions’. Social Choice and Welfare, 24: 3–31.
List, Christian. 2006. ‘The Discursive Dilemma and Public Reason’. Ethics, 116: 362–402.
List, Christian. 2011. ‘Group Communication and the Transformation of Judgments:
An Impossibility Result’. Journal of Political Philosophy, 19: 1–27.
List, Christian; Christian Elsholtz; and Thomas D. Seeley. 2009. ‘Independence and
Interdependence in Collective Decision Making: An Agent-based Model of Nest-site
Choice by Honeybee Swarms’. Philosophical Transactions of the Royal Society B, 364:
755–62.
List, Christian and Robert E. Goodin. 2001. ‘Epistemic Democracy: Generalizing the
Condorcet Jury Theorem’. Journal of Political Philosophy, 9: 276–306.
List, Christian; Robert C. Luskin; James S. Fishkin; and Iain McLean. 2013. ‘Deliberation,
Single-peakedness, and the Possibility of Meaningful Democracy: Evidence from
Deliberative Polls’. Journal of Politics, 75: 80–95.
List, Christian and Philip Pettit. 2004. ‘An Epistemic Free-riding Problem?’ Pp. 128–58 in
Karl Popper: Critical Appraisals, ed. P. Catton and G. Macdonald. Abington: Routledge.
List, Christian and Kai Spiekermann. 2016. ‘Core Voter Knowledge and the Condorcet
Jury Theorem’. Pp. 219–32 in Goldman and His Critics, ed. Brian P. McLaughlin and
Hilary Kornblith. Chichester: Wiley Blackwell.
List, Christian and Adrian Vermeule. 2014. ‘Independence and Interdependence:
Lessons from the Hive’. Rationality & Society, 26: 170–207.
Locke, John. 1669. ‘The Fundamental Constitutions of Carolina’. Pp. 210–32 in Political
Writings of John Locke, ed. David Wootton. New York: Penguin/Mentor, 1983.
Loofbourow, Lili. 2016. ‘Welcome to Trumplandia, Where Feelings Trump Facts’.
The Week, 20 Oct.; available at <http://theweek.com/articles/656455/welcome-
trumplandia-where-feelings-trump-facts>.
Lozada, Carlos. 2015. ‘How Donald Trump Plays the Press, in His Own Words’.
Washington Post, 17 June; available at <https://www.washingtonpost.com/news/book-
party/wp/2015/06/17/how-donald-trump-plays-the-press-in-his-own-words/
?utm_term=.262e91ba30db>.
Lu, Rachel. 2016. ‘You Can Make Your Protest Vote Count. Here’s How’. The Federalist,
31 Oct.; available at <http://thefederalist.com/2016/10/31/make-your-protest-vote-
count/>.
Luck, Matthew. 2016. Polling Lessons from Brexit: The Reason the UK Voted to Leave the
European Union and Others Could Follow. Brussels: Association for Direct Democracy
in Europe.
402 References
Lukes, Steven. 1974/2005. Power: A Radical View. 2nd edn. London: Palgrave Macmillan;
Lukes, Steven and Nadia Urbinati. 2012. ‘Editors’ Introduction’. Pp. xv–xlv in Condorcet:
Political Writings, ed. Lukes and Urbinati. Cambridge: Cambridge University Press.
Lundin, Martin and PerOla Öberg. 2014. ‘Expert Knowledge Use and Deliberation in
Local Policy Making’. Policy Sciences, 47: 25–49.
Lupia, Arthur. 1994a. ‘The Effect of Information on Voting Behavior and Electoral
Outcomes: An Experimental Study of Direct Legislation’. Public Choice, 78: 65–86.
Lupia, Arthur. 1994b. ‘Shortcuts versus Encyclopedias: Information and Voting
Behavior in California Insurance Reform Elections’. American Political Science
Review, 88: 63–76.
Lupia, Arthur. 2002. ‘Deliberation Disconnected: What It Takes to Improve Civic
Competence’. Law & Contemporary Problems, 65: 133–50.
Lupia, Arthur. 2006. ‘How Elitism Undermines the Study of Voter Competence’. Critical
Review, 18: 217–32.
Lupia, Arthur. 2015. Uninformed: Why People Know so Little about Politics and What We
Can Do about It. Oxford: Oxford University Press.
Lupia, Arthur and John G. Matsusaka. 2004. ‘Direct Democracy: New Approaches to
Old Questions’. Annual Review of Political Science, 7: 463–82.
Lupia, Arthur and Mathew D. McCubbins. 1998. The Democratic Dilemma: Can Citizens
Learn What They Need to Know? Cambridge: Cambridge University Press.
Luskin, Robert; Ian O’Flynn; James Fishkin; and David Russell. 2014. ‘Deliberating
Across Deep Divides’. Political Studies, 62: 116–35.
Luskin, Robert C.; James S. Fishkin; and Roger Jowell. 2002. ‘Considered Opinions:
Deliberative Polling in Britain’. British Journal of Political Science, 32: 455–88.
Maass, Arthur. 1983. Congress and the Common Good. New York: Basic.
McAdams, Dan P. 2016. ‘The Mind of Donald Trump’. Atlantic, June; available at <http://
www.theatlantic.com/magazine/archive/2016/06/the-mind-of-donald-trump/
480771/>.
McAllister, Ian and Dudley Studlar. 1989. ‘Popular Versus Elite Views of Privatization:
The Case of Britain’. Journal of Public Policy, 9: 157–78.
McCarthy, Tom. 2017. ‘Can Trump Really Make America Great Again?’ Guardian,
18 Jan.; available at <https://www.theguardian.com/us-news/2017/jan/17/donald-
trump-america-great-again-northampton-county-pennsylvania>.
McCoy, Terrence. 2016. ‘For the “New Yellow Journalists,” Opportunity Comes in Clicks
and Bucks’. 20 Nov.; available at <https://www.washingtonpost.com/national/for-
the-new-yellow-journalists-opportunity-comes-in-clicks-and-bucks/2016/11/20/
d5 8d03 6c-adbf-11e6-8b45-f8e493f0 6fcd_stor y.ht m l?t id=a_in l&ut m_
term=.7d905ff067c7>.
McCubbins, Matthew D. and Daniel B. Rodriguez. 2006. ‘When Does Deliberating
Improve Decisionmaking?’ Journal of Contemporary Legal Issues, 15: 9–50.
Macedo, Stephen et al. 2005. Democracy at Risk: How Political Choices Undermine Citizen
Participation, and What We Can Do About It. Washington, DC: Brookings Institution.
Mackie, Gerry. 2003. Democracy Defended. Cambridge: Cambridge University Press.
Mackie, Gerry. 2012. ‘Rational Ignorance and Beyond’. Pp. 290–318 in Collective Wisdom:
Principles and Mechanisms, ed. Hélène Landemore and Jon Elster. Cambridge:
References 403
Mackie, J. L. 1977. Ethics: Inventing Right and Wrong. Harmondsworth, Mddx.: Penguin.
McLean, Iain and Fiona Hewitt, eds. 1994. ‘Introduction’. Pp. 3–90 in Condorcet:
Foundations of Social Choice and Political Theory. Aldershot, Hants.: Elgar.
McLean, Iain and Arnold B. Urken. 1992. ‘Did Jefferson or Madison Understand
Condorcet’s Theory of Social Choice?’ Public Choice, 73: 445–57.
Macpherson, C.B. 1973. Democratic Theory. Oxford: Oxford University Press.
Madison, James. 1787/1911. Speech in the Federal Convention, Tuesday, June 26. Vol. 1,
pp. 421–2 in The Records of the Federal Convention of 1787, ed. Max. Farrand. New
Haven: Yale University Press, 1911. Available at <http://oll.libertyfund.org/titles/1057>.
Madison, James. 1787/2003. Federalist No. 10. Pp. 40–6 in Hamilton, Madison and Jay,
The Federalist, with Letters of ‘Brutus’, ed. Terence Ball. Cambridge: Cambridge
Madison, James. 1788/2003. Federalist No. 57. Pp. 277–82 in Hamilton, Madison and
Madison, James. 1788/2003. Federalist No. 63. Pp. 305–12 in Hamilton, Madison and
Major, Kirsty. 2016. ‘Why Are We so Surprised that Boris Johnson Lied when He’s
Been Sacked for Lying Twice Before?’ Independent, 28 June; available at <http://www.
independent.co.uk/voices/why-are-we-so-surprised-that-boris-johnson-lied-
when-he-s-been-sacked-for-lying-twice-before-a7105976.html>.
Malthus, T. R. 1798/1992. An Essay on the Principle of Population, ed. Donald Winch.
Cambridge: Cambridge University Press, 1992; originally published 1798.
Manin, Bernard. 1997. Principles of Representative Government. Cambridge: Cambridge
University Press.
March, James G. 1991. ‘Exploration and Exploitation in Organizational Learning’.
Organization Science, 2: 71–87.
Marcus, George E. 2008. ‘Different Situations, Different Responses: Threat, Partisanship,
Risk and Deliberation’. Critical Review, 20: 75–89.
Marcus, Ruth. 2016. ‘Welcome to the Post-truth Presidency’. Washington Post, 2 Dec.;
availableat<https://www.washingtonpost.com/opinions/welcome-to-the-post-truth-
presidency/2016/12/02/baaf630a-b8cd-11e6-b994-f45a208f7a73_story.html?utm_
term=.2d96cf1e091f>.
Marcus, Ruth. 2017. ‘Forget the Post-truth Presidency. Welcome to the Pre-truth
Presidency’. Washington Post, 23 Mar.; available at <https://www.washingtonpost.
com/opinions/welcome-to-the-pre-truth-presidency/2017/03/23/b35856ca-1007-
11e7-9b0d-d27c98455440_story.html?utm_term=.810ba1ae09c2>.
404 References
Marti, Jose Luis. 2006. ‘The Epistemic Conception of Deliberative Democracy Defended’.
In Deliberative Democracy and Its Discontents, ed. S. Besson and J.L. Marti. Aldershot:
Ashgate.
Martin, Douglas. 2009. ‘Walter Cronkite, 92, Dies; Trusted Voice of TV News’. New York
Times, 17 July; available at <http://www.nytimes.com/2009/07/18/us/18cronkite.
html?pagewanted=all&_r=>.
Martin, Judith. 1999. Miss Manners: A Citizen’s Guide to Civility. New York: Random
House.
Martin, Rex. 1993. A System of Rights. Oxford: Clarendon Press.
Marx, Karl and Friedrich Engels. 1848. ‘Manifesto of the Communist Party’. Pp. 335–62
in The Marx–Engels Reader, ed. Robert C. Tucker. New York: Norton, 1972.
Maskin, Eric and Amartya Sen. 2016. ‘How Majority Rule Might Have Stopped Donald
Trump’. New York Times, 28 April; available at <http://www.nytimes.com/2016/05/01/
opinion/sunday/how-majority-rule-might-have-stopped-donald-trump.html>.
Maskin, Eric and Amartya Sen. 2017. ‘The Rules of the Game: A New Electoral System’.
New York Review of Books, 19 Jan.; available at <http://www.nybooks.com/articles/
2017/01/19/rules-of-the-game-new-electoral-system/>.
Masters, Nicholas A. 1961. ‘Committee Assignments’. American Political Science Review,
55: 345–57.
Masur, Louis P. 2012. ‘Lincoln Tells a Story’. New York Times, 27 Jan.; available at <https://
opinionator.blogs.nytimes.com/2012/01/27/lincoln-tells-a-story/?_r=0>.
Matthews, Donald R. 1960. US Senators and Their World. Chapel Hill, NC: University of
North Carolina Press.
Maurin, Vincent and Jean-Pierre Vidal. 2012. ‘Monetary Policy Deliberations, Committee
Size and Voting Rules’. Working Paper Series No. 1434 (May 2012). Frankfurt: European
Central Bank. Available at <https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp1434.
pdf?4e81f2ff31fa53b0941ac3788b5d932a>.
Mayer, Jane. 2016. ‘Donald Trump’s Ghostwriter Tells All’. New Yorker, 25 July; available
at <http://www.newyorker.com/magazine/2016/07/25/donald-trumps-ghostwriter-
tells-all>.
Mayo, Henry B. 1960. An Introduction to Democratic Theory. New York: Oxford
University Press.
Mele, Alfred R. 1993. ‘Motivated Belief ’. Behavior and Philosophy, 21 (2): 19–27.
Mencken, H. L. 1922. ‘Vox Populi’. Reprinted pp. 121–30 in H. L. Mencken’s Smart Set
Criticism, ed. William H. Nolte. Ithaca, NY: Cornell University Press, 1968.
Mencken, H. L. 1926. ‘Notes on Journalism’. Chicago Tribune, 19 September.
Mendelberg, Tali. 2001. The Race Card: Campaign Strategy, Implicit Messages and the
Norm of Equality. Princeton, NJ: Princeton University Press.
Mendelberg, Tali. 2008. ‘Racial Priming Revived’. Perspectives on Politics, 6: 109–23.
Mendelberg, Tali and Adam Berinsky. 2005. ‘The Indirect Effects of Discredited
Stereotypes’. American Journal of Political Science, 49: 846–65.
Mendelberg, Tali and John Oleske. 2000. ‘Race and Public Deliberation’. Political
Communication, 17: 169–91.
Mercier, Hugo. 2011. ‘When Experts Argue: Explaining the Best and the Worst of
Reasoning’. Argumentation, 25: 313–27.
Mercier, Hugo and Hélène Landemore. 2012. ‘Reasoning is for Arguing: Understanding
the Successes and Failures of Deliberation’. Political Psychology, 33: 243–58.
References 405
Mercier, Hugo and Dan Sperber. 2011. ‘Why Do Humans Reason? Arguments for an
Argumentative Theory’. Behavioral and Brain Sciences, 34: 57–74.
Merriman, Justin. 2016. ‘Taking Trump Seriously, Not Literally’. Atlantic, 319, no. 9;
available at <http://www.theatlantic.com/politics/archive/2016/09/trump-makes-his-
case-in-pittsburgh/501335/>.
Milbank, Dana. 2017. ‘Shoker! Rediculous Chocker TRUMP Attaks and Dishoners
English with Ever-dummer Spellings’. Washington Post, 7 Feb.; available at <https://
www.washingtonpost.com/opinions/shoker-rediculous-chocker-trump-attaks-
and-dishoners-english-with-ever-dummer-spellings/2017/02/07/9556faf4-ed58-
11e6-9662-6eedf1627882_story.html?utm_term=.bb888fef6c23>.
Mill, John Stuart. 1859/1977. On Liberty. Vol. XVIII, pp. 212–310 in The Collected Works
of John Stuart Mill, ed. John M. Robson. Toronto: University of Toronto Press. <http://
oll.libertyfund.org/titles/233>; originally published 1859.
Mill, John Stuart. 1861/1977. Considerations on Representative Government. Vol. XIX,
Pt. II in The Collected Works of John Stuart Mill, ed. John M. Robson. Toronto: University
of Toronto Press. <http://oll.libertyfund.org/titles/234>; originally published 1861.
Mill, John Stuart. 1872/1974. A System of Logic, 8th edn. Vols VII and VIII in The
Collected Works of John Stuart Mill, ed. John M. Robson. Toronto: University
of Toronto Press. <http://oll.libertyfund.org/titles/mill-the-collected-works-of-
john-stuart-mill-volume-vii-a-system-of-logic-part-i>; originally published 1843,
8th edn 1872.
Miller, David. 1992. ‘Deliberative Democracy and Social Choice’. Political Studies,
40 (5: Special Issue): 54–67. Reprinted Pp. 54–67 in Prospects for Democracy, ed.
David Held. Oxford: Blackwell.
Miller, Nicholas R. 1986. ‘Information, Electorates and Democracy: Some Extensions
and Interpretations of the Condorcet Jury Theorem’. Pp. 175–94 in Information
Pooling and Group Decision Making, ed. Bernard Grofman and Guillermo Owen.
Greenwich, CT: JAI Press.
Miller, Warren E. and Donald E. Stokes. 1963. ‘Constituency Influence in Congress’.
Misak, Cheryl. 2000. Truth, Politics, Morality. New York: Routledge.
Mohr, Lawrence B. 1969. ‘Determinants of Innovation in Organizations’. American
Political Science Review, 63: 111–26.
Montesquieu, Charles-Louis de Secondat, Baron de. 1721/1973. Persian Letters, trans.
and ed. C. J. Betts. London: Penguin, 1973; originally published 1721.
Mooney, Christopher Z. 1997. ‘Monte Carlo Simulation’. Quantitative Applications in
the Social Sciences, 116. Thousand Oaks, CA: SAGE.
Moore, Alfred. 2017. Critical Elitism: Deliberation, Democracy and the Problem of
Expertise. Cambridge: Cambridge University Press.
Moore, Michael. 1996. ‘The Dead Hand of Constitutional Tradition’. Harvard Journal of
Law & Public Policy, 19: 263–73.
Moyer, Justin Wm.; Jenny Starrs; and Sarah Larimer. 2016. ‘Trump Supporter Charged
after Sucker-punching Protester at North Carolina Rally’. Washington Post, 11 Mar.;
available at <https://www.washingtonpost.com/news/morning-mix/wp/2016/03/10/
trump-protester-sucker-punched-at-north-carolina-rally-videos-show/?utm_term=
.6dd4483767ec>.
Muirhead, Russell. 2014. ‘The Politics of Getting It Right’. Critical Review, 26: 115–28.
406 References
Mutz, Diana C. 2006. Hearing the Other Side: Deliberative versus Participatory Democracy.
Cambridge: Cambridge University Press.
Myers, C. Daniel and Tali Mendelberg. 2013. ‘Political Deliberation’. Pp. 699–734 in
Oxford Handbook of Political Psychology, ed. Leonie Huddy, David O. Sears and Jack
S. Levy. 2nd edn. Oxford: Oxford University Press.
National Institutes of Health (NIH). 2013. NIH Consensus Development Program.
Available at: <http://consensus.nih.gov/>.
Naughton, John. 2017. ‘Is Technology Smart Enough to Fix the Fake News Frenzy?’
Guardian, 4 Feb.; available at <https://www.theguardian.com/commentisfree/2017/
feb/04/technology-fake-news-frenzy-google-facebook-donald-trump>.
Naurin, Daniel. 2007. Deliberation Behind Closed Doors: Transparency and Lobbying in
the European Union. New York: Columbia University Press.
NBC Meet the Press. 2017. Meet the Press 1/1/17. Available at <http://www.nbcnews.
com/meet-the-press/meet-press-1-1-17-n702006>.
Negroponte, Nicholas. 1995. Being Digital. New York: Knopf.
Nemeth, Charlan J.; Bernard Personnaz; Marie Personnaz; and Jack A. Goncalo. 2004.
‘The Liberating Role of Conflict in Group Creativity: A Study in Two Countries’.
European Journal of Social Psychology, 34: 365–74.
Newton, Kenneth and Malcolm Brynin. 2001. ‘The National Press and Party Voting in
Britain’. Political Studies, 49: 265–85.
Nie, Norman H. and D. Sunshine Hillygus. 2001. ‘Education and Democratic Citizenship’.
Pp. 30–56 in Making Good Citizens: Education and Civil Society, ed. Diane Ravitch
and Joseph P. Viteritti. New Haven, CT: Yale University Press.
Nielsen, Annika P.; Janus Hansen; Barbara Skorupinski; Hans-Werner Ingensiep; Heike
Baranzke; Jesper Lassen; and Peter Sandoe. 2006. Consensus Conference Manual.
The Hague: LEI. <http://estframe.net/ethical_bio_ta_tools_project/content_2/text_
2c81d261-b7a8-43e8-8f1e-d724b43e2ba3/1346076808107/et4_manual_cc_
binnenwerk_40p.pdf>.
Niemeyer, Simon. 2011. ‘The Emancipatory Effect of Deliberation: Empirical Lessons
from Mini-publics’. Politics & Society, 39: 103–40.
Niemi Richard G. and Jane Junn. 1998. Civic Education: What Makes Students Learn.
New Haven, CT: Yale University Press.
Niemi, Richard G. and Herbert F. Weisberg. 1968. ‘A Mathematical Solution for the
Probability of the Paradox of Voting’. Behavioral Science, 13: 317–23.
Niemi, Richard G. and Herbert F. Weisberg, eds. 1972. Probability Models of Collective
Decision Making. Columbus, OH: Merrill.
Nisbett, Richard E. and Lee Ross. 1977. ‘Telling More Than We Know: Verbal Reports
on Mental Processes’. Psychological Review, 84: 231–59.
Nitzan, Shmuel and Jacob Paroush. 1982. ‘Optimal Decision Rules in Uncertain
Dichotomous Choice Situation’. International Economic Review, 23: 289–97.
Nossel, Suzanne. 2017. ‘Donald Trump Treats the Press Like an Authoritarian Thug’.
Guardian, 12 Jan.; available at <https://www.theguardian.com/commentisfree/2017/
jan/12/donald-trump-press-conference-cnn-journalist-acosta>.
Nossiter, T.J. 1975. Influence, Opinion and Political Idioms in Reformed England: Case
Studies from the North-east 1832–74. Brighton: Harvester Press.
Noveck, Beth Simone. 2015. Smart Citizens, Smarter State: The Technologies of Expertise
and the Future of Governing. Cambridge, MA: Harvard University Press.
References 407
Novitz, David. 1980. ‘Of Fact and Fancy’. American Philosophical Quarterly, 17: 143–9.
Nozick, Robert. 1981. Philosophical Explanations. Cambridge, MA: Harvard University
Press.
Nurmi, Hannu. 2002. Voting Procedures under Uncertainty. Berlin: Springer.
Nyhan, Brendan and Jason Reifler. 2010. ‘When Corrections Fail: The Persistence of
Political Misperceptions’. Political Behavior, 32: 303–30.
Nyhan, Brendan and Jason Reifler. 2015. ‘The Effect of fact-checking on Elites: A Field
Experiment on US State Legislators’. American Journal of Political Science, 59: 628–40.
O’Connor, Sandra Day; Anthony Kennedy; and David Souter. 1992. Opinion of the US
Supreme Court. Planned Parenthood v Casey, 505 U.S. 833.
O’Harrow, Robert Jr and Shawn Boburg. 2016. ‘The Man Who Showed Donald Trump
How to Exploit Power and Instill Fear’. Washington Post, 17 June; available at <https://
www.washingtonpost.com/investigations/former-mccarthy-aide-showed-trump-
how-to-exploit-power-and-draw-attention/2016/06/16/e9f44f20-2bf3-11e6-9b37-
42985f6a265c_story.html?utm_term=.d3ae78e4d781>.
O’Neil, Cathy. 2016. Weapons of Math Destruction: How Big Data Increases Inequality
and Threatens Democracy. New York: Random House.
Obama, Barack. 2006. The Audacity of Hope. New York: Crown Publishers.
Obama, Barack. 2009. ‘Presidential Memorandum on Transparency and Open
Government’. January 21. Available at <https://obamawhitehouse.archives.gov/the-
press-office/transparency-and-open-government>.
Obama, Barack. 2016. ‘Obama Reckons with a Trump Presidency’. New Yorker, 28 Nov.;
available at <http://www.newyorker.com/magazine/2016/11/28/obama-reckons-with-
a-trump-presidency>.
Obama, Barack. 2017. ‘President Obama’s Farewell Address: Full Text and Video’.
New York Times, 10 Jan.; available at <https://www.nytimes.com/2017/01/10/us/politics/
obama-farewell-address-speech.html>.
Ober, Josiah. 2012. ‘Democracy’s Dignity’. American Political Science Review, 106: 827–46.
Ohlheiser, Abby. 2016. ‘This is How Facebook’s Fake-news Writers Make Money’.
Washington Post, 18 Nov.; available at <https://www.washingtonpost.com/news/the-
intersect/wp/2016/11/18/this-is-how-the-internets-fake-news-writers-make-
money/?tid=a_inl&utm_term=.835b9f726f29>.
Oliver, Eric and Thomas Wood. 2016. ‘A New Poll Shows 52% of Republicans Actually
Think Trump Won the Popular Vote’. Monkey Cage, 18 Dec.; available at <https://
www.washingtonpost.com/news/monkey-cage/wp/2016/12/18/a-new-poll-shows-
an-astonishing-52-of-republicans-think-trump-won-the-popular-vote/?hpid=hp_hp-
top-table-main_mc-goppoll-1114am%3Ahomepage%2Fstory&utm_term=
.6ead01f6dce8>.
Orbell, John M.; von de Kragt, Alphons J. C.; and Dawes, Robyn M. 1988. ‘Explaining
Discussion-induced Cooperation’. Journal of Personality and Social Psychology
54: 811–19.
Orwell, George. 1949. Nineteen Eighty-Four. London: Secker & Warburg.
Osborn, Alex F. 1948. Your Creative Power. New York: Scribner.
Ottaviani, Marco and Peter Sørensen. 2001. ‘Information Aggregation in Debate: Who
Should Speak First?’ Journal of Public Economics, 81: 393–421.
Owen, Guillermo and Bernard Grofman. 1984. ‘To Vote or Not to Vote: The Paradox of
Nonvoting’. Public Choice, 42: 311–25.
408 References
Owen, Guillermo; Bernard Grofman; and Scott L. Feld. 1989. ‘Proving a Distribution-
free Generalization of the Condorcet Jury Theorem’. Mathematical Social Sciences,
17: 1–16.
Oxford Dictionaries. 2016. ‘Word of the Year 2016 is . . ’. Available at <https://en.
oxforddictionaries.com/word-of-the-year/word-of-the-year-2016>.
Packard, Vance. 1957/1981. The Hidden Persuaders. 2nd edn. Harmondsworth, Mddx.:
Penguin; originally published 1957.
Page, Benjamin I., and Robert Y. Shapiro. 1992. The Rational Public. Chicago: University
of Chicago Press.
Page, Scott E. 2007. The Difference: How the Power of Diversity Creates Better Groups,
Firms, Schools, and Societies. Princeton, NJ: Princeton University Press.
Page, Scott E. 2015. ‘Diversity Trumps Ability and the Proper Use of Mathematics’.
Notices of the American Mathematical Society, 62: 9–10.
Parekh, Bhikhu. 2000. The Future of Multi-Ethnic Britain: Report of the Commission on
the Future of Multi-Ethnic Britain. London: Profile Books.
Pariser, Eli. 2011. The Filter Bubble: What the Internet Is Hiding from You. New York:
Penguin.
Parker, Ashley and Sean Sullivan. 2017. ‘Reality Check: Many of Trump’s Early Vows
Will Probably Never Happen’. Washington Post, 26 Jan.; available at <https://www.
washingtonpost.com/politics/reality-check-many-of-trumps-early-vows-will-never-
actually-happen/2017/01/26/2d56bfda-e3e3-11e6-879b-356663383f1b_story.
html?hpid=hp_hp-top-table-main_trumpreality-7a%3Ahomepage%2Fstory&utm_
term=.d8b55d05bc99>.
Paroush, Jacob. 1998. ‘Stay Away from Fair Coins: A Condorcet Jury Theorem’. Social
Choice and Welfare, 15: 15–20.
Pateman, Carole. 1970. Participation and Political Theory. Cambridge: Cambridge
University Press.
Pateman, Carole. 2012. ‘Participatory Democracy Revisited’. Perspectives on Politics,
10: 7–19.
Patterson, Thomas E. 2016. ‘News Coverage of the 2016 General Election: How the Press
Failed the Voters’. Available at <https://shorensteincenter.org/news-coverage-
2016-general-election/?platform=hootsuite>.
Paul, Christopher and Miriam Matthews. 2016. The Russian ‘Firehose of Falsehood’
Propaganda Model: Why It Might Work and Options to Counter It. Perspective
PE-198-OSD, RAND National Defense Research Institute. Available at <http://www.
rand.org/pubs/perspectives/PE198.html>.
Pengelly, Martin. 2017. ‘Kellyanne Conway: “Alternative Facts” Was My Oscars La La
Land Blunder’. Guardian, 3 Mar.; available at <https://www.theguardian.com/
us-news/2017/mar/03/kellyanne-conway-alternative-facts-mistake-oscars>.
Pennock, J. Roland. 1979. Democratic Political Theory. Princeton, NJ: Princeton University
Press.
Perry-Coste, F. H. 1907. ‘The Ballot-Box’. Nature, 75: 509.
Persily, Nathaniel. 2017. ‘Can Democracy Survive the Internet?’ Journal of Democracy,
28 (2): 63–76.
Peter, Fabienne. 2016. ‘The Epistemic Circumstances of Democracy’. Ch. 7 in The
Epistemic Life of Groups: Essays in the Epistemology of Collectives, ed. Michael S.
Brady and Miranda Fricker. Oxford: Oxford University Press.
References 409
Philonenko, Alexis. 1984. Jean-Jacques Rousseau et la Pensée Malheur. Paris: Vrin.

Pigozzi, G. 2006. ‘Belief Merging and the Discursive Dilemma: An Argument-based
Account to Paradoxes of Judgment Aggregation’. Synthese, 152: 285–98.
Piketty, Thomas. 1999. ‘The Information-aggregation Approach to Political Institutions’.
European Economic Review, 43: 791–800.
Pilkington, Ed. 2016. ‘Inside a Donald Trump Rally: Good People in a Feedback Loop
of Paranoia and Hate’. Guardian, 30 Oct.; available at <https://www.theguardian.
com/us-news/2016/oct/30/donald-trump-voters-rally-election-crowd>.
Pilkington, Ed. 2017. ‘Macomb v Media: Voters Who Read Little News Think Trump
Had a Great First Week’. Guardian, 27 Jan.; available at <https://www.theguardian.
com/us-news/2017/jan/27/donald-trump-michigan-voters-media>.
Pinter, Harold. 2005/2012. Nobel Lecture: ‘Art, Truth and Politics’. Available at: <http://
www.nobelprize.org/nobel_prizes/literature/laureates/2005/pinter-lecture-e.html>.
Reprinted in part at pp. 9–15 in Truth and Democracy, ed. Jeremy Elkins and Andrew
Norris. Philadelphia: University of Pennsylvania Press, 2012.
Pitkin, Hanna F. 1967. The Concept of Representation. Berkeley, CA: University of
California Press.
Pocock, J. G. A. 1957. The Ancient Constitution and the Feudal Law. Cambridge: Cambridge
University Press.
Polsby, Nelson W. 1968. ‘The Institutionalization of the US House of Representatives’.
Poole, Keith T. and Howard Rosenthal. 2000. Congress: A Political-Economic History of
Roll Call Voting. New York: Oxford University Press.
Popkin, Samuel L. 1991. The Reasoning Voter. Chicago: University of Chicago Press.
Popkin, Samuel L. 1993. ‘Information Shortcuts and the Reasoning Voter’. Pp. 17–36 in
Information, Participation and Choice, ed. Bernard Grofman. Ann Arbor, MI:
University of Michigan Press.
Popkin, Samuel L. 2006. ‘The Factual Basis of Belief Systems: A Reassessment’. Critical
Review, 18: 233–54.
Popkin, Samuel L. and Michael A. Dimock. 1999. ‘Political Knowledge and Citizen
Competence’. Pp. 117–46 in Citizen Competence and Democratic Institutions, ed.
Steven K. Elkin and Karol E. Soltan. University Park, PA: Penn State University Press.
Popkin, Samuel; John W. Gorman, Charles Phillips and Jeffrey A. Smith. 1976. ‘What
Have You Done for Me Lately?’ American Political Science Review, 70: 779–805.
Posner, Richard A. 2003. Law, Pragmatism and Democracy. Cambridge, MA: Harvard
University Press.
Poynter. 2017. International Fact-checking Network Fact-checkers’ Code of Principles.
Available at <http://www.poynter.org/fact-checkers-code-of-principles/>.
Price, David E. 1981. ‘Congressional Committees in the Policy Process’. Pp. 156–85 in
Congress Reconsidered, 2nd edn, ed. Lawrence C. Dodd and Bruce I. Oppenheimer.
Washington, DC: Congressional Quarterly Press.
Priest, Dana and Michael Birnbaum. 2017. ‘Europe Has Been Working to Expose
Russian Meddling for Years’. Washington Post, 25 June; available at <https://www.
washingtonpost.com/world/europe/europe-has-been-working-to-expose-russian-
meddling-for-years/2017/06/25/e42dcece-4a09-11e7-9669-250d0b15f83b_story.
html?hpid=hp_hp-top-table-main_europefake-6pm%3Ahomepage%2Fstory&utm_
term=.468141355470>.
410 References
Prior, Markus; Gaurav Sood; and Kabir Khanna. 2015. ‘You Cannot Be Serious: The
Impact of Accuracy Incentives on Partisan Bias in Reports of Economic Perceptions’.
Quarterly Journal of Political Science, 10: 489–518.
Prose, Francine. 2016. ‘Truth Is Evaporating Before Our Eyes’. Guardian, 19 Dec.;
available at <https://www.theguardian.com/commentisfree/2016/dec/19/truth-is-
evaporating-before-our-eyes>.
Przeworski, Adam. 2015. ‘Suffrage and Voting Secrecy in General Elections’. Pp. 97–107
in Secrecy and Publicity in Votes and Debates, ed. Jon Elster. Cambridge: Cambridge
University Press.
Queenan, Joe. 2017. ‘Mel Brooks: “Donald Trump Doesn’t Scare Me. He’s a Song-
and-Dance Man” ’. Guardian, 10 Feb.; available at <https://www.theguardian.com/
film/2017/feb/10/mel-brooks-donald-trump-doesnt-scare-me-hes-a-song-and-
dance-man>.
Quirk, Paul J. 2014. ‘Making It Up on Volume: Are Larger Groups Really Smarter?’
Critical Review, 26: 129–50.
Rabin-Havt, Ari. 2016. ‘Trump’s Outrageous Lies Come Straight from Big Businesses’
Playbook’. Washington Post, 16 Dec.; available at <https://www.washingtonpost.com/
posteverything/wp/2016/12/16/trumps-outrageous-lies-come-straight-from-big-
businessess-playbook/?hpid=hp_hp-cards_hp-card-posteverything%3Ahomepage%
2Fcard&utm_term=.1ac288a5445a>.
Rae, Douglas W. 1975. ‘The Limits of Consensual Decision’. American Political Science
Review, 69: 1270–94.
Rampell, Catherine. 2016. ‘When the Facts Don’t Matter, How Can Democracy
Survive?’ Washington Post, 17 Oct.; available at <https://www.washingtonpost.com/
opinions/when-the-facts-dont-matter-how-can-democracy-survive/2016/10/17/56
0ff302-94a5-11e6-9b7c-57290af48a49_story.html?tid=a_inl&utm_term=.
729435a50d7e>.
Rawlinson, Kevin. 2017. ‘Fake News is “Killing People’s Minds” Says Apple Boss Tim
Cook’. Guardian, 11 Feb.; available at <https://www.theguardian.com/technology/2017/
feb/11/fake-news-is-killing-peoples-minds-says-apple-boss-tim-cook>.
Rawls, John. 1971. A Theory of Justice. Cambridge, MA: Harvard University Press.
Rawls, John. 1993. Political Liberalism. New York: Columbia University Press.
Rawls, John. 1999. A Theory of Justice, rev. edn. Cambridge, MA: Harvard University Press.
Rawls, John. 2001. Justice as Fairness: A Restatement. Cambridge MA: Harvard University
Press.
Raz, Joseph. 1986. The Morality of Freedom. Oxford: Clarendon Press.
Raz, Joseph. 1990. ‘Facing Diversity: The Case of Epistemic Abstinence’. Philosophy &
Public Affairs 19: 3–46.
Rehfeld, Andrew. 2009. ‘Representation Rethought: On Trustees, Delegates and
Gyroscopes in the Study of Political Representation and Democracy’. American
Rehm, Diane; with Major Garrett, Abby Phillip, Olivier Knox. 2016a. ‘Friday News
Roundup—Domestic’. The Diane Rehm Show, 18 Nov.; available at <http://thedia-
nerehmshow.org/shows/2016-11-18/friday-news-roundup-domestic>.
Rehm, Diane; with James Fallows, Margaret Sullivan, Glenn Thrush, Mark Baldwin, and
Scottie Nell Hughes. 2016b. ‘How Journalists Are Rethinking Their Role Under a Trump
References 411
Presidency’. The Diane Rehm Show, 30 Nov.; available at <http://thedianerehmshow.

org/shows/2016-11-30/how-journalists-are-rethinking-their-role-under-
a-trump-presidency>.
Rehnquist, William. 1991. Opinion of the US Supreme Court. Payne v Tennessee,
501 U.S. 808.
Reuters. 2017. ‘Germany Investigating Unprecedented Spread of Fake News Online’.
Guardian, 9 Jan.; available at <https://www.theguardian.com/world/2017/jan/09/
germany-investigating-spread-fake-news-online-russia-election>.
Rhodes, R. A. W. and Paul ‘t Hart, eds. 2014. Oxford Handbook of Political Leadership.
Rich, Frank. 2016. ‘Ronald Reagan Was Once Donald Trump’. New York Magazine,
30 May; available at <http://nymag.com/daily/intelligencer/2016/05/ronald-reagan-
was-once-donald-trump.html>.
Riffkin, Rebecca. 2015. ‘Americans’ Trust in Media Remains at Historical Low’. Gallup,
15 Sept.; available at <http://www.gallup.com/poll/185927/americans-trust-media-
remains-historical-low.aspx>.
Riker, William H. 1982. Liberalism against Populism. San Francisco: W. Freeman.
Riker, William H. 1992. ‘The Justification of Bicameralism’. International Political
Robert, Henry M. 1876/1951. Robert’s Rules of Order. Rev. edn. Chicago: Scott, Foresman;
Rogers, James R. 2001. ‘An Informational Rationale for Congruent Bicameralism’.
Journal of Theoretical Politics, 13: 123–51.
Rokkan, Stein. 1961. ‘Mass Suffrage, Secret Voting and Political Participation’. Archives
Européenes di Sociologie, 2: 132–52.
Rosen, Jay; Erik Wemple; and James Downie. 2017. ‘Should the Media Boycott the
White House?’ Washington Post, 1 Mar.; available at <https://www.washingtonpost.
com/opinions/should-the-media-boycott-the-white-house/2017/03/01/57d3968e-
fed2-11e6-8ebe-6e0dbe4f2bca_story.html?hpid=hp_no-name_opinion-card-
f%3Ahomepage%2Fstory&utm_term=.2ff3d3e90b12>.
Rosenblum, Nancy L. 2008. On the Side of the Angels: An Appreciation of Parties and
Partisanship. Princeton, NJ: Princeton University Press.
Rosenblum, Nancy L. 2014. ‘Partisanship and Independence: The Particular Moralism
of American Politics’. Critical Review of International Social and Political Philosophy,
17: 267–88.
Rothschild, Emma. 2001. Economic Sentiments: Adam Smith, Condorcet and the
Enlightenment. Cambridge, MA: Harvard University Press.
Rothstein, Bo and Dietlind Stolle. 2008. ‘The State and Social Capital: An Institutional
Theory of Generalized Trust’. Comparative Politics, 40: 441–59.
Rousseau, Jean-Jacques. 1762/1997. The Social Contract. Pp. 39–152 in The Social Contract
and Other Later Political Writings, trans. and ed. Victor Gourevitch. Cambridge:
Cambridge University Press; originally published 1762.
Rubin, Jennifer. 2017. ‘Maybe Trump Isn’t “Lying” ’. Washington Post, 25 Jan.; available at
<https://www.washingtonpost.com/blogs/right-turn/wp/2017/01/25/maybe-trump-
isnt-lying/?hpid=hp_no-name_opinion-card-e%3Ahomepage%2Fstory&utm_
term=.b61a454e07d8>.
412 References
Rucker, Philip; John Wagner; and Greg Miller. 2017. ‘Trump, in CIA Visit, Attacks
Media for Coverage of His Inaugural Crowds’. Washington Post, 21 Jan.; available at
<https://www.washingtonpost.com/politics/trump-in-cia-visit-attacks-media-for-
coverage-of-his-inaugural-crowds/2017/01/21/f4574dca-e019-11e6-ad42-f3375f271c9c_
story.html?hpid=hp_hp-bignews3_pp-cia-1141am%3Ahomepage%2Fstory&utm_
term=.50f428a5db9d>.
Ryan, Frances. 2016. ‘Martin’s Already Lost Almost Everything—He Voted Leave to
Spread the Pain’. Guardian, 30 June; available at <https://www.theguardian.com/
commentisfree/2016/jun/30/martin-nothing-lose-vote-leave-unemployed-benefits-
sanctioned>.
Ryle, Gilbert. 1949. The Concept of Mind. London: Hutchinson.
Ryle, Gilbert. 1958. ‘On Forgetting the Difference between Right and Wrong’. Pp. 147–59
in Essays in Moral Philosophy, ed. A. I. Melden. Seattle, University of Washington Press.
Saad, Lydia. 2017. ‘Approval of Trump Transition Still Low as Inauguration Nears’.
Gallup, 12 Jan.; available at <www.gallup.com/poll/201833/approval-trump-transition-
low-inauguration-nears.aspx>.
Safire, William. 1988. ‘Watch What We Do’. New York Times, 14 Nov.; available at <http://
www.nytimes.com/1988/11/14/opinion/essay-watch-what-we-do.html>.
Safire, William. 2008. Safire’s Political Dictionary. Rev. edn. New York: Oxford University
Press.
Samuelson, William F. and Richard Zeckhauser. 1988. ‘Status Quo Bias in Decision
Making’. Journal of Risk and Uncertainty, 1: 7–59.
Sargent, Greg. 2017a. ‘GOP Voters Know Trump Is Telling Them the Truth, and the
Media Is Lying’. Washington Post, 22 Feb.; available at <https://www.washingtonpost.
com/blogs/plum-line/wp/2017/02/22/gop-voters-know-trump-is-telling-them-the-truth-
and-the-media-is-lying-to-them/?hpid=hp_no-name_opinion-card-f%3Ahomepage
%2Fstory&utm_term=.a7f2d3cb0411>.
Sargent, Greg. 2017b. ‘Trump’s Lies Are Working Brilliantly. This New Poll Proves It’.
Washington Post, 27 Ap.; available at <https://www.washingtonpost.com/blogs/plum-
line/wp/2017/04/27/trumps-lies-are-working-brilliantly-this-new-poll-proves-it/
?hpid=hp_no-name_opinion-card-f%3Ahomepage%2Fstory&utm_term=
.dd5bee4f7032>.
Satterthwaite, Mark A. 1975. ‘Strategy-proofness and Arrow’s Conditions’. Journal of
Economic Theory, 10: 187–217.
Saunders, Peter. 1995. ‘Privatization, Share Ownership and Voting’. British Journal of
Sawyer, Keith. 2007. Group Genius: The Creative Power of Collaboration. New York:
Basic Books.
Schaffner, Brian and Samantha Luks. 2017. ‘This Is What Trump Voters Said When
Asked to Compare His Inauguration Crowd with Obama’s’. Monkey Cage, 25 Jan.;
available at <https://www.washingtonpost.com/news/monkey-cage/wp/2017/01/25/
we-asked-people-which-inauguration-crowd-was-bigger-heres-what-they-
said/?hpid=hp_hp-top-table-main_trumpfacts805p%3Ahomepage%2Fstory&
utm_term=.ac44a79ef094>.
Scharpf, Fritz W. 1999. Governing in Europe: Effective and Democratic? Oxford: Oxford
University Press.
References 413
Schattschneider, E. E., chair. 1950. ‘Toward a More Responsible Two-party System’.

Report of the Committee on Political Parties of the American Political Science
Association. American Political Science Review, 44 (no. 3, pt 2, Supplement: Sept).
Schauer, Frederick. 1987. ‘Precedent’. Stanford Law Review, 39: 571–605.
Schlesinger, Arthur M., Jr. 1965. A Thousand Days: John F. Kennedy in the White House.
Boston, MA: Houghton Mifflin.
Schlesinger, Arthur M., Jr. 1973. The Imperial Presidency. Boston, MA: Houghton Mifflin.
Schmitt, Carl. 1922/2005. Political Theology, trans. by G. Schwab, Chicago: University of
Chicago Press; originally published 1922.
Schrag, Peter. 1969. ‘The Forgotten American’. Harper’s, Aug., pp. 27–34. Available at
<http://web.mit.edu/21h.102/www/Primary%20source%20collections/Vietnam/
Schrag,%20Forgotten_American.htm>.
Schumpeter, Joseph A. 1942. Capitalism, Socialism, and Democracy. New York: Harper.
Schwartzberg, Melissa. 2015. ‘Epistemic Democracy and Its Challenges’. Annual Review
of Political Science, 18: 187–203.
Schwartzberg, Melissa. 2016. ‘Aristotle and the Judgment of the Many: Equality, Not
Collective Quality’. Journal of Politics, 78: 733–45.
Schwartzman, Micah J. 2011. ‘The Sincerity of Public Reason’. Journal of Political
Sclanfani, Jennifer. 2017. Talking Donald Trump: A Sociolinguistic Study of Style,
Metadiscourse and Political Identity. London: Routledge, Chapman & Hall.
Searle, John R. 1964. ‘How to Derive “Ought” from “Is” ’. Philosophical Review,
73: 43–58.
Seeberg, Henrik Bech. 2017. ‘How Stable Is Political Parties’ Issue Ownership? A Cross-
time, Cross-national Analysis’. Political Studies, 65: 475–92.
Seeley, Thomas. 1996. The Wisdom of the Hive. Cambridge, MA: Harvard University
Press.
Segal, Jeffrey A. and Harold J. Spaeth. 1996. ‘The Influence of Stare Decisis on the Votes
of United States Supreme Court Justices’. American Journal of Political Science,
40: 971–1003.
Sen, Amartya. 2009. The Idea of Justice. Cambridge, MA: Harvard University Press.
Sennett, Richard and Jonathan Cobb. 1993. The Hidden Injuries of Class. New York:
Norton.
Servan-Schreiber, Emile J. 2012. ‘Prediction Markets: Trading Uncertainty for Collective
Wisdom’. Pp. 21–37 in Collective Wisdom: Principles and Mechanisms, ed. Hélène
Landemore and Jon Elster. Cambridge: Cambridge University Press.
Shane, Scott. 2017. ‘From Headline to Photograph, a Fake News Masterpiece’. New York
Times, 18 Jan.; available at <https://www.nytimes.com/2017/01/18/us/fake-news-
hillary-clinton-cameron-harris.html?hp&action=click&pgtype=Homepage&
clickSource=story-heading&module=b-lede-package-region&region=top-news&
WT.nav=top-news&_r=1>.
Shapin, Steven. 1994. A Social History of Truth: Civility and Science in Seventeenth-
century England. Chicago: University of Chicago Press.
Shapiro, Robert Y. and Yaeli Bloch-Elkon. 2008. ‘Do the Facts Speak for Themselves?
Partisan Disagreement as a Challenge to Democratic Competence’. Critical Review,
20: 115–39.
414 References
Shapley, Lloyd S. and Bernard Grofman. 1984. ‘Optimizing Group Judgmental Accuracy
in the Presence of Interdependencies’. Public Choice, 43: 329–43.
Shepsle, Kenneth A. 1978. The Giant Jigsaw Puzzle: Democratic Committee Assignments
in the Modern House. Chicago: University of Chicago Press.
Shepsle, Kenneth A. 1979. ‘Institutional Arrangements and Equilibrium in Multi
dimensional Voting Models’. American Journal of Political Science, 23: 27–59.
Shepsle, Kenneth A. and Barry R. Weingast. 1981. ‘Structure-induced Equilibrium and
Legislative Choice’. Public Choice, 37: 503–19.
Shepsle, Kenneth A. and Barry R. Weingast. 1987. ‘The Institutional Foundations of
Committee Power’. American Political Science Review, 81: 84–104.
Shils, Edward. 1981. Tradition. Chicago: University of Chicago Press.
Ship, Chris. 2016. ‘ITV EU Referendum Debate: Boris Johnson Focus of “Remain”
Attacks’. ITV News, 9 June; available at <http://www.itv.com/news/2016-06-09/boris-
focus-of-remain-attacks-in-fiery-eu-referendum-debate/>.
Shklar, Judith N. 1991. American Citizenship: The Quest for Inclusion. Cambridge, MA:
Harvard University Press.
Siddique, Haroon. 2017. ‘Teach Schoolchildren How to Spot Fake News, Says OECD’.
Guardian, 18 Mar.; available at <https://www.theguardian.com/media/2017/mar/18/
teach-schoolchildren-spot-fake-news-says-oecd>.
Sidgwick, Henry. 1907. The Methods of Ethics, 7th edn. London: Macmillan.
Silver, Nate. 2012. The Signal and the Noise. New York: Penguin.
Silver, Nate. 2016. ‘How Much Did Comey Hurt Clinton’s Chances?’ 538 Politics, 6 Nov.;
available at <https://fivethirtyeight.com/features/how-much-did-comey-hurt-clintons-
chances/>.
Silverman, Craig. 2016. ‘This Analysis Shows How Viral Fake Election News Stories
Outperformed Real News on Facebook’. BuzzFeed, 17 Nov.; available at <https://
www.buzzfeed.com/craigsilverman/viral-fake-election-news-outperformed-real-
news-on-facebook?utm_term=.idRXzGwDJz#.hkWvre5x6r>.
Simmel, Georg. 1950. ‘The Phenomenon of Outvoting’. Pp. 239–49 in The Sociology of
Georg Simmel, ed. Kurt Wolff. Glencoe, IL: Free Press.
Simon, Herbert A. 1982. Models of Bounded Rationality. 2 vols. Cambridge, MA: MIT
Press.
Simpson, A. W. B. 1973. ‘The Common Law and Legal Theory’. Pp. 77–99 in Oxford
Essays in Juisprudence, 2nd series, ed. A. W. B. Simpson. Oxford: Clarendon Press.
Sinderbrand, Rebecca. 2017. ‘How Kellyanne Conway Ushered in the Era of “Alternative
Facts” ’. Washington Post, 22 Jan.; available at <https://www.washingtonpost.com/
news/the-fix/wp/2017/01/22/how-kellyanne-conway-ushered-in-the-era-
of-alternative-facts/?hpid=hp_hp-top-table-main_fix-conwaytranscript-138pm%
3Ahomepage%2Fstory&utm_term=.6e9cfda619b6>.
Skocpol, Theda and Morris Fiorina, eds. 1999. Civic Engagement in American Democracy.
Washington, DC: Brookings Institution.
Skocpol, Theda; Marshall Ganz; and Ziad Munson. 2000. ‘A Nation of Organizers: The
Institutional Origins of Civic Voluntarism in the United States’. American Political
Skocpol, Theda and Alexander Hertel-Fernandez. 2016. ‘The Koch Network and
Republican Party Extremism’. Perspectives on Politics, 14: 681–99.
References 415
Smith, Daniel A. and Caroline Tolbert. 2004. Educated by Initiative: The Effects of Direct
Democracy on Citizens and Political Organizations in the American States. Ann
Arbor, MI: University of Michigan Press.
Smith, Graham. 2009. Democratic Innovations: Citizen Participation in Political Decision
Making. Cambridge: Cambridge University Press.
Smith, Lones and Peter Sorensen. 2000. ‘Pathological Outcomes of Observational
Learning’. Econometrica, 68: 371–98.
Smith, Michael. 1994. The Moral Problem. Oxford: Blackwell.
Sniderman, Paul M. 2000. ‘Taking Sides: A Fixed Choice Theory of Political Reasoning’.
Pp. 67–84 in Elements of Reason, ed. Arthur Lupia, Matthew D. McCubbins and
Samuel L. Popkin. New York: Cambridge University Press.
Sniderman, Paul M.; Richard A. Brody; and Philip E. Tetlock. 1991. Reasoning and
Choice. New York: Cambridge University Press.
Solon, Olivia. 2017. ‘Tim Berners-Lee Calls for Tighter Regulation of Online Political
Advertising’. Guardian, 12 Mar.; available at <https://www.theguardian.com/
technology/2017/mar/11/tim-berners-lee-online-political-advertising-regulation>.
Somin, Ilya. 2006. ‘Knowledge about Ignorance: New Directions in the Study of Political
Information’. Critical Review, 18: 255–78.
Somin, Ilya. 2013. Democracy and Political Ignorance: Why Smaller Government is
Smarter. Stanford, CA: Stanford University Press.
Sommerlad, Fiona and Iain McLean. 1989. The Political Theory of Condorcet. Working
paper 1/89. Oxford: Social Studies Faculty Centre.
Sorensen, Theodore C. 1965. Kennedy. London: Hodder and Stoughton.
Soroka, Stuart; Blake Andrew; Toril Aalberg; Shanto Iyengar; James Curran; Sharon
Coen; Kaori Hayashi; Paul Jones; Bianpetro Mazzoleni; June Woong Rhee; David
Row; and Rod Tiffen. 2013. ‘Auntie Knows Best? Public Broadcasters and Current
Affairs Knowledge’. British Journal of Political Science, 43: 719–39.
Sparrow, Andrew. 2016a. ‘David Davis Questioned by MPs over Brexit’. Guardian,
14 Sept.; available at <https://www.theguardian.com/politics/blog/live/2016/sep/13/
boundary-review-corbyn-says-inner-city-seat-shouldnt-be-enlarged-politics-
live>.
Sparrow, Andrew. 2016b. ‘Thirty Things You Didn’t Know about the EU Referendum’.
Guardian, 22 Dec.; available at <https://www.theguardian.com/politics/blog/2016/
dec/22/thirty-things-you-didnt-know-about-the-eu-referendum>.
Spenkuch, Jörg L. 2017. ‘Expressive vs. Pivotal Voters: An Empirical Assessment’.
Working paper, Kellogg School of Management, Northwestern University.
Available at <http://www.kellogg.northwestern.edu/faculty/spenkuch/research/
voting.pdf>.
Spenkuch, Jörg L. and David Toniatti. 2016. ‘Political Advertising and Election
Outcomes’. CESinf Working Paper Series No. 5780. Available at <https://papers.ssrn.
com/sol3/papers.cfm?abstract_id=2753512>.
Sperber, Dan; Fabrice Clément; Christophe Heintz; Oliver Mascaro; Hugo Mercier;
Gloria Origgi; and Deirdre Wilson. 2010. ‘Epistemic Vigilance’. Mind & Language,
25: 359–93.
Spiekermann, Kai. 2010. ‘Judgement Aggregation and Distributed Thinking’. AI and
Society, 25: 401–12.
416 References
Spiekermann, Kai and Robert E. Goodin. 2012. ‘Courts of Many Minds’. British Journal
of Political Science, 42: 555–72. Supplementary Material available at <https://static.
cambridge.org/resource/id/urn:cambridge.org:id:binary:20161020083143951-
0771:S000712341100041X:S000712341100041Xsup001.pdf>.
Spiekermann, Kai and Robert E. Goodin. 2014. ‘Constructing Epistemic Reference
Groups when Interests Diverge’. Unpublished ms., LSE/ANU.
Stasavage, David. 2007. ‘Polarization and Publicity: Rethinking the Benefits of Deliberative
Democracy’. Journal of Politics, 69: 59–72.
Stasser, Garold and William Titus. 1985. ‘Pooling of Unshared Information in Group
Decision Making: Biased Information Sampling During Discussion’. Journal of
Personality & Social Psychology, 48: 1467–78.
Steinberg, Tom. 2016. Facebook post, 24 June. Available at <https://www.facebook.
com/tom.steinberg.503/posts/10157028566365237>.
Stelter, Biran. 2016. ‘Following Up with Scottie Nell Hughes about Facts’. CNN
Money: Reliable Sources; available at <http://us11.campaign-archive1.com/?u=
47c9040f6ff957a59bd88396e&id=781d6ae935>.
Stich, Stephen G.W. 2014. ‘When Democracy Meets Pluralism: Landemore’s Epistemic
Argument for Democracy and the Problem of Value Diversity’. Critical Review, 26:
170–83.
Stone, Harlan. 1938. Dissenting Opinion. United States v Carolene Products. 304 U.S. 144.
Stone, Jon. 2016a. ‘Brexit Secretary David Davis Says EU Referendum Campaign Untruths
“Not Important” ’. The Independent, 5 Oct.; available at <http://www.independent.
co.uk/news/uk/politics/david-davis-brexit-secretary-eu-referendum-campaign-
lies-a7344611.html>.
Stone, Jon. 2016b. ‘Nigel Farage Backtracks on Leave Campaign’s “£350m for the
NHS” Pledge Hours After Result’. The Independent, 24 June; available at <http://
www.independent.co.uk/news/uk/politics/eu-referendum-result-nigel-farage-nhs-
pledge-disowns-350-million-pounds-a7099906.html>.
Strauss, David A. 1996. ‘Common Law Constitutional Interpretation’. University of
Chicago Law Review, 63: 877–935.
Stubager, Rune. 2003. ‘Preference-shaping: An Empirical Test’. Political Studies, 51: 241–61.
Sullivan, Margaret. 2016. ‘The Post-truth World of the Trump Administration is Scarier
than You Think’. Washington Post, 4 Dec.; available at <https://www.washingtonpost.
com/lifestyle/style/the-post-truth-world-of-the-trump-administration-is-scarier-
than-you-think/2016/12/02/ebda952a-b897-11e6-b994-f45a208f7a73_story.
html?utm_term=.fea321e6735e>.
Sullivan, Margaret. 2017a. ‘Don’t Get Fooled Again by Bogus Links, Bots and Pure
Bunk’. Washington Post, 31 Mar.; available at <https://www.washingtonpost.com/
lifestyle/style/dont-get-fooled-again-by-bogus-links-bots-and-pure-bunk-heres-
how/2017/03/31/b8758104-14bd-11e7-833c-503e1f6394c9_story.html?hpid=hp_hp-
more-top-stories_sullivan-1125am%3Ahomepage%2Fstor y&utm_term=
.98816e4ad387>.
Sullivan, Margaret. 2017b. ‘It’s Time to Retire the Tainted Term “Fake News” ’. Washington
Post, 8 Jan.; available at <https://www.washingtonpost.com/lifestyle/style/its-time-to-
retire-the-tainted-term-fake-news/2017/01/06/a5a7516c-d375-11e6-945a-76f69a399dd5_
story.html?hpid=hp_hp-top-table-main_sullivan-435pm%3Ahomepage%2Fstory
&utm_term=.e44981fb8ccc>.
References 417
Sunstein, Cass R. 1993. ‘On Analogical Reasoning’. Harvard Law Review, 106: 741–91.
Sunstein, Cass R. 2000. ‘Deliberative Trouble? Why Groups Go to Extremes’. Yale Law
Journal, 110: 71–119.
Sunstein, Cass R. 2001. Republic.com. Princeton, NJ: Princeton University Press.
Sunstein, Cass R. 2002. ‘The Law of Group Polarization’. Journal of Political Philosophy,
10: 175–95.
Sunstein, Cass R. 2003. Why Societies Need Dissent. Cambridge, MA: Harvard University
Press.
Sunstein, Cass R. 2006a. ‘Burkean Minimalism’. Michigan Law Review, 105: 353–408.
Sunstein, Cass R. 2006b. ‘Deliberating Groups versus Prediction Markets (or Hayek’s
Challenge to Habermas)’. Episteme, 3: 192–213.
Sunstein, Cass R. 2006c. Infotopia: How Many Minds Produce Knowledge. New York:
Sunstein, Cass R. 2009a. A Constitution of Many Minds. Princeton, NJ: Princeton
University Press.
Sunstein, Cass R. 2009b. Going to Extremes: How Like Minds Unite and Divide.
New York: Oxford University Press.
Sunstein, Cass R. 2017a. #Republic: Divided Democracy in the Age of Social Media.
Sunstein, Cass R. 2017b. ‘Deliberative Democracy in the Trenches’. Daedalus, 146 (3):
129–39.
Sunstein, Cass R. and Reid Hastie. 2014. Wiser: Getting Beyond Groupthink to Make
Groups Smarter. Cambridge, MA: Harvard Business Review Press.
Sunstein, Cass R. and Reid Hastie. 2015. ‘Garbage In, Garbage Out? Some Micro Sources
of Macro Errors’. Journal of Institutional Economics, 11: 561–83.
Sunstein, Cass R. and Edna Ullmann-Margalit. 2001. ‘Solidarity Goods’. Journal of
Political Philosophy, 9: 129–49.
Surowiecki, James. 2004. The Wisdom of Crowds. London: Abacus/Little, Brown.
Suskind, Ron. 2004. ‘Faith, Certainty and the Presidency of George W. Bush’. New York
Times Magazine, 17 Oct.; available at <http://www.nytimes.com/2004/10/17/magazine/
faith-certainty-and-the-presidency-of-george-w-bush.html?_r=0>.
Swaim, Barton. 2017. ‘At Least This Time We Don’t Have to Pretend the President’s
Good’. Washington Post, 19 Jan.; available at <https://www.washingtonpost.com/
blogs/post-partisan/wp/2017/01/19/at-least-this-time-we-dont-have-to-pretend-the-
presidents-good/?hpid=hp_no-name_opinion-card-e%3Ahomepage%2Fstory
&utm_term=.0e5d516ae904>.
Swan, Jonathan. 2016. ‘Trump: “Forget the Press, Read the Internet” ’. The Hill, 18 Oct.;
available at <http://thehill.com/blogs/blog-briefing-room/news/301634-trump-
forget-the-press-read-the-internet>.
Swift, Adam. 2014. Political Philosophy: A Beginner’s Guide for Students and Politicians.
3rd edn. Cambridge: Polity.
Swift, Jonathan. 1710. ‘The Art of Political Lying’. The Examiner, 9 Nov. Reprinted in Swift,
Works, ed. Thomas Roscoe (London: George Bell & Sons, 1880), vol. 2, pp. 402–5.
Talisse, Robert B. 2005. Democracy after Liberalism: Pragmatism and Deliberative
Politics. New York: Routledge.
Talisse, Robert B. 2009. Democracy and Moral Conflict. New York: Cambridge University
Press.
418 References
Talisse, Robert B. 2013. ‘Sustaining Democracy: Folk Epistemology and Social Conflict’.
Critical Review of International Social and Political Philosophy, 16: 500–19.
Talley, Eric. 1999. ‘Precedential Cascades: An Appraisal’. Southern California Law
Review, 73: 87–137.
Tangian, A. S. 2000. ‘Unlikelihood of Condorcet’s Paradox in a Large Society’. Social
Choice & Welfare, 17: 337–65.
Taylor, Michael and Michael Laver. 1973. ‘Government Coalitions in Western Europe’.
European Journal of Political Research, 1: 205–48.
Tetlock, Philip E. 2005. Expert Political Judgment: How Good Is It? How Can We Know?
Thompson, Abigail. 2014. ‘Does Diversity Trump Ability? An Example of the Misuse of
Mathematics in the Social Sciences’. Notices of the American Mathematical Society,
61: 1024–30.
Thompson, Abigail. 2015. ‘Response to Page and Megginson’. Notices of the American
Mathematical Society, 62: 9–10.
Thompson, Christopher. 2013. ‘A General Model of a Group Search Procedure, Applied
to Epistemic Democracy’. Synthese, 190: 1233–52.
Thompson, Dennis F. 1970. The Democratic Citizen. Cambridge: Cambridge University
Press.
Treynor, Jack. 1987. ‘Market Efficiency and the Bean Jar Experiment’. Financial Analysis
Journal, 43: 50–3.
Tribe, Laurence H. 1971. ‘Trial By Mathematics: Precision and Ritual in the Legal
Process’. Harvard Law Review, 84: 1329–93.
Trump, Donald J. 2016a. Full Text: Donald Trump 2016 RNC Draft Speech Transcript.
Available at: <http://www.politico.com/story/2016/07/full-transcript-donald-trump-
nomination-acceptance-speech-at-rnc-225974>.
Trump, Donald J. 2016b. Transcript of Speech at Roanoke, Virginia, Rally, Sept 24.
WHSV 3; available at <http://www.whsv.com/content/news/WATCH-Live-raw-
video-of-Trump-rally-in-Virginia-394694121.html>.
Trump, Donald J. 2017a. Donald Trump’s Full Inauguration Speech Transcript,
Annotated. Washington Post, 20 Jan.; available at <https://www.washingtonpost.
com/news/the-fix/wp/2017/01/20/donald-trumps-full-inauguration-speech-
transcript-annotated/?utm_term=.3ac8ad9f0b1d>.
Trump, Donald J. 2017b. Donald Trump’s News Conference: Full Transcript and Video.
New York Times, 11 Jan.; available at <https://www.nytimes.com/2017/01/11/us/politics/
trump-press-conference-transcript.html>.
Trump, Donald J. 2017c. Trump CIA Speech Transcript. CBS News, 21 Jan.; available at
<http://www.cbsnews.com/news/trump-cia-speech-transcript/>.
Tsebelis, George. 1995. ‘Decision Making in Political Systems: Veto Players in
Presidentialism, Parliamentarianism, Multicameralism and Multipartyism’. British
Journal of Political Science, 25: 289–325.
Tsebelis, George and Jeanette Money. 1995. ‘Bicameral Negotiations: The Navette System
in France’. British Journal of Political Science, 25: 101–29.
Tumulty, Karen and David Nakamura. 2017. ‘Trump’s Rallying Cry: Fear Itself ’.
Washington Post, 3 Feb.; available at <https://www.washingtonpost.com/politics/
trumps-rallying-cry-fear-itself/2017/02/03/7d2a0432-ea4a-11e6-bf6f-301b6b443624_
story.html?hpid=hp_hp-cards_hp-card-politics%3Ahomepage%2Fcard&utm_
term=.8177986689cd>.
References 419
Tumulty, Karen and Philip Rucker. 2016. ‘Shouting Match Erupts Between Clinton and
Trump Aides’. Washington Post, 1 Dec: available at <https://www.washingtonpost.com/
politics/shouting-match-erupts-between-clinton-and-trump-aides/2016/12/01/
7ac4398e-b7ea-11e6-b8df-600bd9d38a02_story.html?hpid=hp_rhp-top-table-
main_election750p%3Ahomepage%2Fstory&tid=a_inl&utm_term=.e65e055549>.
TwitterCounter. 2017. ‘Donald Trump Twitter Stats’. Available at <http://twittercounter.
com/realDonaldTrump?utm_expid=102679131-118.LzR3SMUURL24X3MIR-d_
FA.0&utm_referrer=http%3A%2F%2Fwww.google.com.au%2Furl%3Fsa%3Dt%
26rct%3Dj%26q%3D%26esrc%3Ds%26source%3Dweb%26cd%3D13%26ved%
3D0ahUKEwjg7ZXlmcrRAhULzLwKHUxVBdEQFghQMAw%26url%3Dhttp%
253A%252F%252Ftwittercounter.com%252FrealDonaldTrump%26usg%
3DAFQjCNENEdgsIForJmhhEnJZx961c4P9gw>.
Tynan, Dan. 2016. ‘How Facebook Powers Money Machines for Obscure Political “News”
Sites’. Guardian, 25 Aug.; available at <https://www.theguardian.com/technology/2016/
aug/24/facebook-clickbait-political-news-sites-us-election-trump>.
Uhr, John. 1998. Deliberative Democracy in Australia. Cambridge: Cambridge University
Press.
UK Electoral Commission. 2016a. ‘Electoral Commission Designates “Vote Leave Ltd”
and “The In Campaign Ltd” as Lead Campaigners at EU Referendum’. Available at
<http://www.electoralcommission.org.uk/i-am-a/journalist/electoral-commission-
media-centre/news-releases-referendums/electoral-commission-designates-vote-
leave-ltd-and-the-in-campaign-ltd-as-lead-campaigners-at-eu-referendum>.
UK Electoral Commission. 2016b. EU Referendum Results. Available at <http://www.
electoralcommission.org.uk/find-information-by-subject/elections-and-referendums/
past-elections-and-referendums/eu-referendum/electorate-and-count-information>.
UK Statistics Authority. 2016. The UK Contribution to the EU Budget. Available at
<http://visual.ons.gov.uk/uk-perspectives-2016-the-uk-contribution-to-the-eu-
budget/>.
Urbinati, Nadia. 2006. Representative Democracy: Principles & Genealogy. Chicago:
University of Chicago Press.
US Director of National Intelligence. 2017. Assessing Russian Activities and Intentions in
Recent US Elections. Intelligence Community Assessment ICA 2017-01D, 6 January 2017.
Washington, DC: Office of the Director of National Intelligence, National Intelligence
Council. Available at <https://www.dni.gov/files/documents/ICA_2017_01.pdf>.
US Federal Election Commission (US FEC). 2017. ‘Official 2016 Presidential
General Election Results’. Available at <https://transition.fec.gov/pubrec/fe2016/
2016presgeresults.pdf>.
Verba, Sidney and Norman H. Nie. 1972. Participation in America: Political Democracy
and Social Equality. New York: Harper & Row.
Vermeule, Adrian. 2005. ‘Submajority Rules: Forcing Accountability upon Majorities’.
Journal of Political Philosophy, 13: 74–98.
Vermeule, Adrian. 2007. Mechanisms of Democracy: Institutional Design Writ Small.
Vermeule, Adrian. 2009a. Law and the Limits of Reason. Oxford: Oxford University Press.
Vermeule, Adrian. 2009b. ‘Many-minds Arguments in Legal Theory’. Journal of Legal
Analysis, 1: 1–44.
Vermeule, Adrian. 2009c. ‘The Parliament of the Experts’. Duke Law Journal,
58: 2231–75.
420 References
Vermeule, Adrian. 2010. ‘The Limits of Epistemic Legalism: A Reply’. Jerusalem Review
of Legal Studies, 2: 47–53.
Vermeule, Adrian. 2011a. ‘Second Opinions and Institutional Design’. Virginia Law
Review, 97: 1435–74.
Vermeule, Adrian. 2011b. The System of the Constitution. Oxford: Oxford University Press.
Vermeule, Adrian. 2015. ‘Open-secret Voting’. Pp. 215–30 in Secrecy and Publicity in
Votes and Debates, ed. Jon Elster. Cambridge: Cambridge University Press.
Vicario, Michela Del; Alessandro Bessi; Fabiana Zollo; Fabio Petroni; Antonio Scala;
Guido Caldarelli; H. Eugene Stanley; and Walter Quattrociocchi. 2016. ‘The Spreading
of Misinformation Online’. Proceedings of the National Academy of Sciences, 113:
554–9.
Viehoff, Daniel. 2016. ‘Authority and Expertise’. Journal of Political Philosophy, 24: 406–26.
Vine, Sarah. 2016. ‘“Gosh, I Suppose I Better Get Up!” Sarah Vine (aka Mrs Gove)
Reveals what Her Husband Said when He Learned Leave Had Won the Referendum . . .
and How PM’s Resignation Was “Absolutely” Not Intended’. Daily Mail, 29 June;
available at <http://www.dailymail.co.uk/news/article-3665146/SARAH-VINE-
Victory-vitriol-craziest-days-life.html>.
Viner, Katharine. 2016. ‘How Technology Disrupted the Truth’. Guardian, 12 July;
available at <https://www.theguardian.com/media/2016/jul/12/how-technology-
disrupted-the-truth>.
Voss, Stephen. 2016. Hitchens Prize: ‘Washington Post Editor Marty Baron Has a
Message to Journalists in the Trump Era’. Washington Post, 30 Nov.; available at
<http://www.vanityfair.com/news/2016/11/washington-post-editor-marty-baron-
message-to-journalists>.
Vyse, Graham. 2016. ‘Newt Gingrich Is Already Deifying Donald Trump’. New Republic,
14 Dec.; available at <https://newrepublic.com/article/139321/newt-gingrich-already-
deifying-donald-trump>.
Wagner, John. 2017. ‘Trump Agrees with Bannon’s Assessment that the Media is
the “Opposition Party”’. Washington Post, 27 Jan.; available at <https://www.
washingtonpost.com/news/post-politics/wp/2017/01/27/trump-agrees-with-
bannons-assessment-that-the-media-is-the-opposition-party/?hpid=hp_hp-cards_
hp-card-politics%3Ahomepage%2Fcard&utm_term=.4f0881e0a70d>.
Waldron, Jeremy. 1989. ‘Democratic Theory and the Public Interest: Condorcet and
Rousseau Revisited’. American Political Science Review, 83: 1322–8.
Waldron, Jeremy. 1995. ‘The Wisdom of the Multitude: Some Reflections on Book 3,
Chapter 11 of Aristotle’s Politics’. Political Theory, 23: 563–84.
Waldron, Jeremy. 1999a. The Dignity of Legislation. Cambridge: Cambridge University
Press.
Waldron, Jeremy. 1999b. Law & Disagreement. Oxford: Oxford University Press.
Waldron, Jeremy. 2000. ‘Legislation by Assembly’. Loyola Law Review, 46: 507–34.
Waldron, Jeremy. 2016. Political Political Theory. Cambridge, MA: Harvard University
Press.
Walker, Jack L. 1969. ‘The Diffusion of Innovations among the American States’.
Wang, Amy B. 2016. ‘ “Post-truth” Named 2016 Word of the Year by Oxford Dictionaries’.
Washington Post, 16 Nov.; available at <https://www.washingtonpost.com/news/
References 421
the-fix/wp/2016/11/16/post-truth-named-2016-word-of-the-year-by-oxford-
dictionaries/?utm_term=.281432aa9404>.
Warren, Mark E. and John Gastil. 2015. ‘Can Deliberative Minipublics Address the
Cognitive Challenges of Democratic Citizenship?’ Journal of Politics, 77: 562–74.
Warren, Mark E. and Hilary Pearse, eds. 2008. Designing Deliberative Democracy: The
British Columbia Citizens’ Assembly. Cambridge: Cambridge University Press.
Washington, George. 1746. Rules of Civility and Decent Behaviour in Company and
Conversation. Bedford, MA: Applewood Books, 1988.
Washington, George. 1796/1966. Farewell Address. Available at: <http://www.ourdocu-
ments.gov/doc.php?doc=15&page=transcript>. Reprinted pp. 192–210 in An American
Primer, ed. Daniel J. Boorstein. Chicago: University of Chicago Press, 1966.
Washington Post Editorial Board. 2016. ‘D.C. Pizzeria Is Casualty of Trump’s Safe
Zone for Fact-free Discourse’. Washington Post, 6 Dec.; available at <https://www.
washingtonpost.com/opinions/dc-pizzeria-is-a-casualty-of-trumps-safe-zone-
for-fact-free-discourse/2016/12/06/b68708ba-bb1f-11e6-ac85-094a21c44abc_
story.html?utm_term=.1a53d69df803>.
Washington Post Editorial Board. 2017a. ‘The Absurd Conspiracy Theory around Seth
Rich’s Death Causes Real Harm’. Washington Post, 22 May; available at <https://www.
washingtonpost.com/opinions/the-absurd-conspiracy-theory-around-seth-richs-
death-causes-real-harm/2017/05/22/520ba572-3f21-11e7-9869-bac8b446820a_story.
html?hpid=hp_no-name_opinion-card-d%3Ahomepage%2Fstory&utm_term=
.b791a4a3c2f2>.
Washington Post Editorial Board. 2017b. ‘ “Some Have Classified It as Treason,” but
These Opposing-party Lawmakers are Collaborating’. Washington Post, 29 May;
available at <https://www.washingtonpost.com/opinions/some-have-classified-it-
as-treason-but-these-opposing-party-lawmakers-are-collaborating/2017/05/29/
21529412-422e-11e7-8c25-44d09ff5a4a8_story.html?hpid=hp_no-name_opinion-
card-f%3Ahomepage%2Fstory&utm_term=.7452c138048d>.
Weber, Max. 1947. The Theory of Social and Economic Organization, trans. A. M. Henderson
and Talcott Parsons. New York: Oxford University Press.
Weedon, Jen; William Nuland; and Alex Stamos. 2017. Information Operations and
Facebook. Mountain View, CA: Facebook. Available at: <https://fbnewsroomus.files.
wordpress.com/2017/04/facebook-and-information-operations-v1.pdf>.
Weingarten, Gene. 1996. ‘Read it and Veep’. Washington Post, 4 Feb.: F4.
Weingast, Barry R. and William Marshall. 1988. ‘The Industrial Organization of
Congress: or, Why Legislatures, Like Firms, Are Not Organized as Markets’. Journal
of Political Economy, 96: 765–800.
Weinstock, Daniel. 1999. ‘Democracy, Value and Truth: Saving Deliberation from
Justification’. Available at <http://mapageweb.umontreal.ca/lepagef/dept/cahiers/
Weinstock_Deliberation.pdf>.
Weinstock, Daniel. 2016. ‘How Democratic Is Civil Disobedience?’ Criminal Law and
Weisberg, Michael and Ryan Muldoon. 2009. ‘Epistemic Landscapes and the Division
of Cognitive Labor’. Philosophy of Science, 76: 225–52.
Wemple, Erik. 2017a. ‘Guess Who’s Now Accusing Others of Spreading “Fake News” ’.
Washington Post, 11 Jan.; available at <https://www.washingtonpost.com/blogs/
422 References
erik-wemple/wp/2017/01/11/guess-whos-now-accusing-others-of-spreading-fake-
news/?utm_term=.ada825fdcabd>.
Wemple, Erik 2017b. ‘The Life-sucking Tedium of Covering Official Lies’. Washington
Post, 9 Feb.; available at <https://www.washingtonpost.com/blogs/erik-wemple/wp/
2017/02/09/the-life-sucking-tedium-of-covering-official-lies/?hpid=hp_no-name_
opinion-card-e%3Ahomepage%2Fstory&utm_term=.4f6bf2b275bc>.
Weymark, John A. 2015. ‘Cognitive Diversity, Binary Decisions, and Epistemic Democracy’.
Episteme, 12: 497–511.
Whelan, Frederick G. 1983. ‘Democratic Theory and the Boundary Problem’. Pp. 13–47
in Nomos XXV: Liberal Democracy, ed. J. R. Pennock and J. W. Chapman. New York:
New York University Press.
White, Jonathan and Lea Ypi. 2011. ‘On Partisan Political Justification’. American
White, Jonathan and Lea Ypi. 2016. The Meaning of Partisanship. Oxford: Oxford
University Press.
Williams, Bernard. 2002. Truth and Truthfulness. Princeton, NJ: Princeton University
Press.
Williams, David. 2004. Condorcet and Modernity. New York: Cambridge University
Press.
Wilson, Deirdre and Dan Sperber. 2012. Meaning and Relevance. Cambridge: Cambridge
University Press.
Wingfield, Nick; Mike Isaac; and Katie Benner. 2016. ‘Google and Facebook Take Aim
at Fake News Sites’. New York Times, 14 Nov.; available at <https://www.nytimes.
com/2016/11/15/technology/google-will-ban-websites-that-host-fake-news-from-
using-its-ad-service.html>.
Wittman, Donald. 1995. The Myth of Democratic Failure. Chicago: University of Chicago
Press.
Wolfers, Justin and Eric Zitzewitz. 2004. ‘Prediction Markets’. Journal of Economic
Perspectives, 18 (2): 107–26.
Wolff, Jonathan. 1994. ‘Democratic Voting and the Mixed-Motivation problem’. Analysis,
54: 193–6.
Wollheim, Richard. 1962. ‘A Paradox in the Theory of Democracy’. Pp. 71–87 in
Philosophy, Politics & Society, 2nd series, ed. P. Laslett & W. G. Runciman. Oxford:
Blackwell.
Woodhouse, A.S.P. 1938. Puritanism and Liberty. London: Dent.
YouGov. 2016a. Times survey results, 20–22 June; available at <https://d25d2506sfb94s.
cloudfront.net/cumulus_uploads/document/atmwrgevvj/TimesResults_160622_
EVEOFPOLL.pdf>.
YouGov. 2016b. Today Programme survey results, 13–14 June; available at <https://
d25d2506sfb94s.cloudfront.net/cumulus_uploads/document/x4iynd1mn7/
TodayResults_160614_EUReferendum_W.pdf>.
Young, H. Peyton. 1988. ‘Condorcet’s Theory of Voting’. American Political Science
Review, 82 (no. 4: Dec): 1231–44.
Yuhas, Alan. 2016. ‘How Does Donald Trump Lie? A Fact Checker’s Final Guide’.
Guardian, 7 Nov.; available at https://www.theguardian.com/us-news/2016/nov/07/
how-does-donald-trump-lie-fact-checker.
References 423
Yuhas, Alan. 2017. ‘Smoke and Mirrors: How Trump Manipulates the Media and
Opponents’. Guardian, 18 Jan.; available at <https://www.theguardian.com/us-news/
2017/jan/18/donald-trump-media-manipulation-tactics>.
Zaigraev, Alexander, and Serguei Kaniovski. 2012. ‘Bounds on the Competence of a
Homogeneous Jury’. Theory and Decision, 72: 89–112.
Zaller, John. 1992. Nature and Origins of Mass Opinion. New York: Cambridge University
Press.
Zaller, John. 2012. ‘What Nature and Origins Leaves Out’. Critical Review, 24: 569–642.
Zito, Salena. 2016. ‘Taking Trump Seriously, Not Literally’. Atlantic, 23 Sept.; available at
<https://www.theatlantic.com/politics/archive/2016/09/trump-makes-his-case-in-
pittsburgh/501335/>.
Zollman, Kevin J.S. 2010. ‘The Epistemic Benefit of Transient Diversity’. Erkenntnis,
72: 17–35.
Index
access Aristotle 3n, 5, 143n, 350

to cues 188 Arneson, R. 227n, 313n
to evidence 71, 73, 75–6, 82, 160, 185–6, Asch, S. 108
187–8 Asymptotic Result 20, 24, 26, 29, 54, 61, 64,
to facts and information 3, 82, 99–100, 105, 73, 78n, 166–7, 209, 310n (see also
249, 261 Condorcet Jury Theorem)
to true state of the world 10, 71, 73, 79, Austen-Smith, D. 47n, 50n
81, 186 Austin, J. L. 305
to voters 341 Australian Senate 268n, 285
accountability 40, 248, 291n, 299, 300, 326 available evidence 8, 16, 71–3, 94, 106, 210n,
Achen, C. H. 179n, 185n 306, 313–16, 320, 347 (see also evidence)
Adams, John 266
advisory committees 260, 289, 290, 293 Bächtiger, A. 137n
agenda Bacon, Francis 354n
correct option not on agenda 42, 43–4, 63, Bagehot, Walter 226
112–14, 119, 143, 288–90 balance of power 267
expansion 13, 63, 107, 111–14, 119–20, 142, 317 ballot, pre-election 214 (see also block voting;
(see also brainstorming; deliberation; epistemic solidarity)
localized search; search; search parties) bandwagon effects 284, 286 (see also cascades)
incomplete 111–14, 203 (see also agenda, Banerjee, A. V. 154n
correct option not on agenda; agenda, Banks, Arron 326, 343
expansion) Banks, J. S. 47n
manipulation 45, 63, 65, 135n (see also Bannon, Stephen 359, 361, 362
agenda, setting) Bardella, Kurt 361n
random 45 Bartels, Larry 88–9, 90, 179n, 182, 185n
reduction 120, 121–8, 290–1, 317–18 Bayes’ theorem 4, 5n, 32, 154n, 233, 306
(see also subvotes) Beatty, J. 71n
setting 8, 64n, 301, 317, 320 (see also beauty contests 220
agenda, expansion; agenda, reduction) bees, behaviour of 111
well formulated 63–6, 143 Begin, Menachem 183n
aids see institutional aids Bendor, J. 200n
Ailes, Roger 323n Benkler, Y. 356n
Allcott, H. 360n Bentham, Jeremy 247n, 260n, 296, 299
‘alternative facts’ 323n, 325n, 342–3, 359 Berelson, B. R. 175
(see also deception; lies; ‘post-truth’; Berlusconi, Silvio 168–9, 297
sincerity; Sincerity Assumption; truth Berners-Lee, Tim 356
claims) best available evidence 10–11, 16, 209n, 306,
alternatives, adding see agenda, expansion 347 (see also evidence)
Althaus, Scott 89n, 90 best responder 8, 62, 72–82, 135, 137, 185, 195,
Ambrus, A. 139n 236–7, 246n, 288, 290, 293, 314n, 316n
American Founders 7, 139n, 144, 208, 226, deliberation to improve best responder
246n, 248–9, 252, 266–7, 277n, 278–9, performance 141–2
281, 282, 285n (see also Franklin, Best Responder Corollary 11, 16, 58n, 73,
Benjamin; Hamilton, A.; Jefferson, T.; 76–9, 81, 135n, 141–2, 168n, 186, 231n,
Madison, J.) 237–8, 246n, 253, 306, 314, 316
Ancell, A. 119n cue-taking and 186
Anderson, Elizabeth 73 direct vs representative democracy
Arendt, Hannah 105n, 306n, 323n 246, 253
aristocracy 226, 321n evidence-limited case 80, 293
426 Index
Best Responder Corollary (cont.) Bullock, J. 200n
guide to notations used 370 Burke, Edmund 149–50, 152
multiple common causes 80–1, 231n Burkean paradox 154
single common causes 79–80 Bush, George W. 327n
Betz, G. 144n
bias 11, 53–8, 62, 68, 83 89, 140, 182, 186–8, Caillaud, B. 273n
221, 237, 239, 247, 285, 296, 298, 305, Calhoun, J. C. 109
317–18, 319, 320, 324, 356, 364n Canadian Senate 285n
(see also heuristics; prejudice) Caplan, B. 86n
‘confirmation bias' 140n, 181n, 340n cascades 55, 151, 154–9, 162, 247, 286
(see also motivated reasoning) (see also decision-making, sequential)
cue-taking and 182, 184n, 186–7 guide to notation used 370
expert vs lay bias 82 Cassam, Quassim 347
news media 71, 94, 98–9 Chamley, C. P. 154n
‘overconfidence bias’ 111n China 183n, 294
psychological heuristics 184, 187–8 choice situation 8, 26, 31, 62–4, 170, 189, 196n
(see also heuristics) (see also agenda; binary choice;
sincerity 49n decision situation; options)
‘status quo bias’ 117, 277–83 options are ill formulated 63–6
bicameralism 261n, 264, 267–9, 270, 285, systematically misleading 37, 62
287n, 289 (see also legislatures; lower Christiano, T. 46–7n, 290n
house; upper house) Churchill, Winston 170n
‘big lie’ 323 (see also deception; lies; truth Churchland, P. 179n
claims) CIA 358, 361
Bikhchandani, S. 154n citizens’ initiatives 294
binary choice 8, 17n, 18, 23, 26–34, 64n, 65, civic education 8, 83, 85, 92, 94, 298, 320
98n, 122, 126–7, 170n, 189n, 196n Clark, Andy 179n
binomial distribution 20 class 10, 170, 184n, 186, 195n, 208, 209n, 226n,
biological systems 112, 179n 237, 239, 295, 338, 344 (see also elites;
Black, Duncan 38 factionalism; interests)
Blackstone, William 279n, 280n, 295n consciousness 226n
block voting 213–16 (see also factionalism; ‘hidden injuries of class’ 344
party line; party whips) climate change denial 304, 322n
Bohman, J. 105n Clinton, Hillary 329, 331, 332, 333, 336, 339,
Borda, Jean-Charles, chevalier de 7, 33, 36n 340, 343
Borda count 33n, 34, 35, 36, 331n clones 98, 169 (see also Independence
bottlenecks see epistemic bottlenecks Assumption)
Bradley, R. 233n, 234n coalitions 261, 262, 263, 272, 273–6
brainstorming 136n, 143n (see also agenda Cohen, Joshua 5, 68n
expansion) Cohn, Ray 323n
Breitbart 325n, 360, 362, 364 collective competence 13, 21–2, 30n, 74–5, 85,
Brennan, Jason 54n, 86n, 247n, 312n, 313n 87, 138, 143, 167, 186, 192–3, 230, 241–3,
Brexit referendum 302, 322, 324, 331 250, 251–3, 256–9, 260, 263–4, 269–71,
affective explanations 341–2, 343, 345 274, 298–9, 310 (see also Condorcet
differing priorities 335–6, 338 Jury Theorem; group competence)
epistemic agnosticism 366 estimated by Monte Carlo simulation
epistemic democracy under threat 367 371–2
epistemic insouciance 348, 349, 352, 354 guide to notation used 369
epistemic malevolence 358n common causes 16, 54, 57, 59–60, 61, 62, 68,
Facebook reactions 356 70–1, 74–82, 83, 88n, 98–100, 104, 108,
false claims by Vote Leave 325–7, 339 140, 141, 165, 189, 236, 237–9, 253, 257,
strategic signalling vs sincere voting 280, 295, 296, 314–15, 317, 321
332–3, 341 (see also bias; cascades; cues;
Brooks, Mel 347n conditionalizing; evidence; heuristics;
Bryce, J. 170–1 ideology; opinion leaders;
bubbles 354n, 355, 356, 357 psychological heuristics)
Index427
multiple independent common causes 74–6, knowledge vs competence 91
77n, 80–1, 99–100, 139, 177, 187n, 190, multi-alternative competence 27–31, 121–3
231, 237–8, 315, 319 (see also competence, varies with
single common cause 79–80, 98–9 number of options)
common knowledge 137, 154 random 9–10, 18, 28n, 44, 45, 48n, 50, 52,
common law 97n, 150, 156 53, 57, 58, 59, 65, 76, 88, 90, 129–31, 165,
Comparative Manifesto Project 206n 173, 182, 189, 191, 192, 204, 305, 315, 343
competence 9–10, 17–18, 23–6, 50–4, 85–95 topic-specific 25–6
(see also Competence Assumption; unequal factional competence 200–2,
collective competence; competence 211–16 (see also competence,
improvement; Deliberation Effect; asymmetries)
government competence; group varying with number of options 30–1,
competence; incompetence; 121–3, 126–7, 292
institutional aids; institutional worse than random 9–10, 37, 52–3, 57–9,
hindrances; random, better than) 190, 192–3, 242, 246, 314, 351
agenda relative 65 (see also failures of competence)
asymmetries in 25–6, 200 (see also Competence Assumption 1–2, 8, 9–10
competence, unequal factional (see also competence; Condorcet
competence) Jury Theorem)
better than random 9, 18, 23, 26, 28n, 29, classic CJT framework 17–18
30, 45, 50, 52, 59–60, 65–6, 69, 77, 85, uniform voter competence 209–11, 242
88, 92, 94, 103, 111, 121–2, 126, 129, weakening the Competence
134n, 138, 143, 169, 173, 180, 188, 191, Assumption 23
209, 213–14, 227, 230, 234, 242, 246–7, mean competence among heterogenous
257, 263, 297, 298, 305, 306, 310, 315, voters 23–5
316, 332, 336–7, 343 lower threshold in multi-alternative
in choosing representatives 246–51 choices 27–8
(see also Selection Effect) topic-specific competence 25–6
coping with incompetent voters 9–10, competence improvement 83, 85, 94–5
23–5, 86–92, 94–5, 135–8, 144–5, (see also civic education; deliberation;
178–82, 188–94, 229–35, 239–43, diversity; division of epistemic labour)
246–7, 298–9, 316 deliberation and 135–8
epistemic effects of expanding the institutional aids 298–9
electorate learning from experience 240–2
each individual’s competence is limiting the number of options 121–8
known 229–30 (see also agenda, reduction)
enfranchising batches of voters with political engagement 95
heterogeneous individual priorities in 92–4
competence 230 competent cues see cue-taking
enfranchising voters with heterogeneous concurrent majority 109
knowledge bases 231, 238–9 conditionalizing 23n, 26, 37, 48, 68–78, 80–1,
only average individual competence is 97n, 100, 101, 141, 152, 169, 172, 174n,
known 230 186n, 187n, 189n, 210, 269n, 274n, 294,
epistemic logic of enfranchising the less 297n, 311, 314
competent 231–3 on all common causes 74–6
failures of competence 50–4 (see also on the decision situation 77–8
competence, worse than random) on the evidence 72–3
guide to notations used 369 on opinion leader’s interpretation of
heterogeneous competence 15, 23–5, evidence 73–4
229–31, 233–4 Condorcet, Nicolas Caritat, Marquis de 6–7,
mean 23–5, 54n, 83n, 92, 94, 227n, 9, 32, 46n, 51n, 52, 53, 56n, 60, 63, 110,
231, 298 138, 143n, 208, 225, 226, 232n, 237, 246,
improving individual competence 254n, 266, 278n, 312
see competence improvement Condorcet Jury Theorem (CJT)
interaction with independence 55–60, 236 aggregation vs deliberation 132
isocompetence 20, 228 best- and worst-case scenarios 314–15
428 Index
Condorcet Jury Theorem (CJT) (cont.) mean competence among heterogeneous
classic framework 5–7, 12, 15, 17–21, 23–7, voters 23–5
29, 31, 32, 36, 37, 42, 43–7, 50–1, 56, topic-specific competence 25–6
57–60, 61, 62n, 67, 70, 73, 76–9, 83, weakening the Independence
105, 108n, 135, 140, 141–2, 143, 147, Assumption 60–2, 67–81
163n, 172, 187, 191–3, 195, 196–7, 229, Best Responder Corollary 76–81
236, 238, 246n, 253, 294–5, 297n, 306, conditionalizing 69–76
314, 319 Condorcet pairwise comparison 33n, 34, 35,
calculation 20–1 36, 123, 331n
Competence Assumption 17–18 conformity 108, 139–40, 159n
(see also Competence Assumption) consensus 5, 38, 294, 310
group competence 21–2 overlapping 203n
Independence Assumption 8, 10–11, 12, perceived 140
15–16 (see also Independence Consensus Conferences 3, 293–4
Assumption) conventionalism see moral conventionalism;
Sincerity Assumption 19 (see also social constructivism
Sincerity Assumption) constituencies 46, 255–6 (see also
theorem 19–20 representative democracy)
extensions 7–8, 15, 23 (see also weakening constructivism see social constructivism
the Competence Assumption; consumer sovereignty 164
weakening the Independence ‘contestatory deliberation’ 137n (see also
Assumption) deliberation; opposition parties)
more than two alternatives 26–31 context-dependent agenda limitations 112–14,
factional interpretation see factionalism 116–17, 119
false information and 330–1, 340–1 conversable groups 144, 236, 244, 269, 270,
guide to notations used 369–70 288–9, 320 (see also deliberation)
institutional design implications Converse, Philip 91, 181–2, 247n
see institutional aids; institutional Conway, Kellyanne 349n, 353, 358–9, 365n
hindrances Cook, Tim 364n
large vs small groups 316–18 cooling-off periods 283–7
objections to 5, 8–12 cooperative behaviour 49, 50n, 52, 299n
competence 8, 9–10, 15, 37, 50–4 correctness 18, 38n, 39, 52, 62, 88, 105,
(see also Competence Assumption) 144, 145n, 150, 163, 185, 188n, 196, 309,
facts vs values 11–12, 37, 38 (see also 313, 324
value judgements) as closeness to correct answer 18, 28n, 44,
failures of sincerity 37, 45 (see also 119n, 144
Sincerity Assumption) courts 3n, 55, 87, 111, 136n, 149–63, 170n, 189n,
independence 8, 10–11, 12, 15, 16, 240n, 260n, 275n, 279, 280–1, 309n,
37, 54 (see also Independence 325n, 339, 346, 362n (see also judicial
Assumption) behaviour; juries; precedents; stare
options are ill-formulated 63–6 decisis)
systematically misleading choice Cowen, T. 359n
situation 37, 62 (see also choice criminal trials 71, 280, 325n (see also judicial
situation; decision situation) behaviour; juries; magistrates)
truth claims 37, 42 (see also truth claims; Cronkite, Walter 354n
truth-tracking) crowdsourcing 3, 4, 143n
using different decision rules 31–2 Cuban Missile Crisis 143n
Bayesian parallel 32 cue-taking 9, 13, 55, 71, 74–7, 81–2, 91, 104,
Borda count 33n, 34, 35, 36 139, 147, 178–94, 231, 264n, 291n, 298,
Condorcet pairwise comparison 33n, 34, 301, 315, 316, 319, 320, 345
35, 36 bias 184n, 186–7
impartial culture model 33, 34n calculating epistemic effects 185–9
plurality rule 27, 29, 33, 34, 35–6, 63 baseline calculation 189–90
weakening the Competence incompetent voters 193–4
Assumption 23–6 relative insensitivity to the rate of
differing decision situations 81–3 cue use 191
Index429
sensitivity to the number and reliability delegate-style representation 13, 53n, 244,
of independent cues 190 248n, 254–6, 268, 270, 272n,
varying individual voter 295n, 301, 316 (see also representative
competence 191–3 democracy)
cueing incompetents 193–4 mixed assemblies with both delegates and
effectiveness trustees 257–9
evidence from large-scale surveys 181–3 deliberation 6, 8, 9, 12, 63, 83, 105n, 132–45,
experimental evidence 180–1 223, 236, 245, 249, 282, 289n, 290n,
unreliable cues 183–5 293–4, 299, 307, 317, 320, 321 (see also
intentional selection 184–5, 189 Deliberation Effect; deliberative
endorsements 179, 181, 298 (see also news democracy; deliberative institutions;
media; political parties) Deliberative Poll)
guide to notations 369 aggregation vs 132
heuristics 11, 55, 71, 76, 103, 104, 106, 139, benefits 134–5, 317
184, 187–8, 231, 301 (see also bias) changing the decision problem 142–4
low-information rationality 178–80 improving best responder
misleading cues 183–4 performance 141–2
party identification and label 91, 179, increasing individual competence 135–8
183–4, 298 increasing truth-conduciveness 135,
141–2
Dahl, Robert 39n, 321n inducing sincerity 50, 135, 140–1,
Dancey, L. 184n 299–300
Daugherty, A. F. 154n reducing positively-correlated votes 135,
Davis, David 327n, 366n 138–40
deception 323, 324n, 327n (see also changing the decision problem 142–3
‘alternative facts’; lies; ‘post-truth’; ‘contestatory deliberation’ 137n
Sincerity Assumption; strategic ideal and practice 133–4
signalling; strategic voting; truth jury 287n
claims) small conversable groups 144, 236, 244,
decision-making 2, 3n, 5, 12, 26n, 53, 63, 83, 269, 270, 288–9, 320
97, 110n, 123, 128, 144, 162, 163, 188, Deliberation Effect 144–5, 236 (see also
239, 254n, 259, 260, 264, 271, 276, 289, deliberation)
290, 301, 312n, 315, 317 direct democracy 249, 251–4
democratic 73, 227n, 301 representative democracy 254, 256–7, 258,
sequential 70, 151, 154–5, 286, 301 259, 275n
(see also cascades) small upper houses or committees 269–72
decision rules see vote aggregation rules deliberative democracy 132n, 134, 136n, 137n,
decision situation 8, 65, 77–82, 82, 137, 224, 251, 289–90, 300 (see also deliberation;
227n, 236–8, 288, 290, 293, 307, 314, democratic theory)
315, 316n, 318 (see also choice situation; deliberative institutions 293–4 (see also
institutional aids; institutional deliberation; institutional aids)
hindrances) Deliberative Poll 133n, 294 (see also
decision theory 4 deliberation; polls)
deference 131, 155n, 225, 270 (see also democratic theory 164, 223, 301, 316 (see also
experts; opinion leaders; decision-making; direct vs
precedents; pre-election ballot; representative democracy;
stare decisis) representative democracy)
consultation vs 178n deliberative democracy 134, 136n, 289–90
epistemic costs of 151–3, 319 epistemic democracy 7, 8, 12, 13, 15, 38,
partial deference 153–6 302, 321, 324–5, 367 (see also
to previous decisions see traditionalism Condorcet Jury Theorem)
representatives defer to electorate 248 epistocracy vs democracy see epistocracy
(see also trustee-style representation) matters that should not be put to a
to specialists 131, 265–6, 271, 319 vote 307–8
delays 117, 155n, 157–8, 267n, 283, 284, 289 participatory democracy 95, 239–40,
(see also cooling-off periods) 245, 254n
430 Index
democracy 26n, 32, 33–5, 40n, 42, 49n, 65, competent voters choosing even-more-
73, 86, 147, 164, 195–7, 199, 203–7, competent representatives: Selection
223, 225–43, 254, 260n, 290–1n, 295, Effect 248–51
296n, 297, 298, 301–2, 305, 307, 308, epistemic benefits of smaller groups:
309, 311, 312–14, 319, 321, 330 Deliberation Effect 251–4
(see also democratic theory; vote incompetent masses choosing competent
aggregation rules) representatives 246–8
associational 298 disagreement 38n, 40n, 41n, 112n, 137n, 179,
deliberative 132n, 134, 136n, 137n, 251, 188n, 205, 294, 304, 308, 324, 359n,
289–90, 300 (see also deliberation; 363n, 365n (see also dissent; diversity;
democratic theory) negatively correlated votes; opposition
direct 13, 95, 223, 240n, 244, 245, 247, 248, parties)
250, 253, 254, 255, 256, 259, 316, 317 discussion see deliberation
epistemic 7, 8, 12, 13, 15, 26n, 38, 41–2, 50, disjunction problem 30n, 52n, 64–5
52–3n, 54n, 94, 132n, 133n, 147, 165, dissent 107, 108, 140, 310–11 (see also
196–7, 199, 203–6, 225, 238, 244, 301–2, disagreement; free speech; negatively
305, 312–14, 316, 319, 321, 322, 324–5, correlated votes; opposition parties)
330, 367 (see also Condorcet Jury diversity 3, 12, 61, 62n, 83, 96–109, 115, 116,
Theorem) 139, 140, 195n (see also negatively
epistocracy vs 13, 111n, 223, 225–43, 244 correlated votes)
liberal 164, 208n, 211 clones vs 98
mass 9 (see also electorate; franchise) common causes vs 98–100
parliamentary 108, 262, 264 deliberation and 139–40
party 273n, 293 , 297 dissenting opinions 107, 108, 140, 310
participatory 95, 239–40, 245, 254n epistemic benefits 96–8, 104–7, 298
representative 13, 223, 244–60, 288–90, extending the franchise 231, 238–9
292, 316, 317, 345 (see also individuality 97
representative government) multiculturalism 96, 298, 299
democratically-epistemically correct promotion 107–9
outcome 41–2, 147, 196–7, 199, 203–6 trumps ability 97, 98n
(see also moral majoritarianism) division of epistemic labour 12, 106n, 110–31,
Denmark 293 236, 290n, 317
dependence 55, 60–2, 68, 70–1, 72, 82, 83, 98, devolving control over some
112–14, 116–19, 138–41, 239n, 287n, 295 dimensions 128–31
(see also common causes; evidence; localized search 111–12 (see also search)
independence; interdependence; diversified (or many random) search
Independence Assumption; opinion parties 114–16
leaders) incomplete agendas 112–14
devil’s advocate 107, 108 (see also opposition recognizing the best 119–20
parties) transition costs 116–19
devolving control over dimensions 128–31 narrowing the focus 120–1
Dietrich, Franz 67, 72n, 77n, 78n, 104n, 128n considering options a few at a
Dilnot, Sir Andrew 326 time 123–4
dimensions of a decision 128–31 experts vs whole groups 126–8
Dimock, M. A. 93n individual competence reduces with the
direct democracy 13, 95, 223, 240n, 244, 245, number of options 121–3
247, 248, 250, 253, 254, 255, 256, 259, subgroups vs whole group 124–6
316, 317 (see also direct vs Downs, Anthony 27, 86, 87, 178
representative democracy; Drew, Elizabeth 342n
democracy) Dryzek, J. S. 290n
direct vs representative democracy 13, 95, Dworkin, R. 39n
223, 240n, 244, 316, 317 (see also direct
democracy; representative echo chamber 94, 356n (see also bubbles)
democracy) education 39n, 82, 83, 97n, 157, 170, 226, 237,
smaller group of representatives vs larger 240, 316, 320, 364 (see also civic
group of voters 245 education)
Index431
egalitarian distribution of decision-making 321, 322, 324–5, 330, 367 (see also
power 26n Condorcet Jury Theorem)
election outcomes 88–90 (see also US election epistemic deference 155n, 319
outcomes) (see also experts; opinion leaders)
Electoral College 144, 145, 249, 250, 251, 252, epistemic enhancement 83, 85
253, 331, 332 (see also competence improvement;
electoral reform 110n, 238–9, 294, 296 institutional aids)
(see also electorate; franchise) epistemic injustice 137n
electorate 9, 21, 24, 25n, 31, 34n, 42, 49, 50–1, epistemic insouciance 347–54
56, 59–60, 62n, 65, 83, 85–6, 88, 92, 94, epistemic labour see division of epistemic
104, 107, 108, 124, 126–8, 130–1, 144–5, labour
147, 163, 168, 169, 171–2, 173, 176, 179n, epistemic malevolence 358–65
182, 190, 193–4, 198–9, 202–5, 209, epistemic modesty 306 (see also epistemic
210n, 212, 214, 223, 227–8, 233–4, 235n, abstinence; epistemic agnosticism)
237, 238–9, 245, 246n, 248–56, 259, epistemic solidarity 213–16
264, 291, 295, 307n, 310n, 316–18, differential abstention from 217
320–1, 324, 330, 332, 333n, 338, 340–1, epistemic vigilance 133n, 367 (see also
358n, 366 epistemic insouciance;
expanding 229–33 (see also electoral fact checkers)
reform) epistemology, social 133
elites 147, 179n, 208, 211–20, 223n, 227–9, epistocracy 13, 111n, 126n, 223, 225–7,
329, 335, 343 (see also class; 244, 309
epistocracy; experts; experts; beating the smartest clique of guys 228–9
factionalism; masses) beating the smartest single guy 227–8
epistemic 223n, 233–5, 227–9 (see also competence-weighted voting rules 233–5
epistocracy) differential benefits of learning from
endorsements 179, 181, 298 (see also experience 239
cue-taking; news media; political classical argument of participatory
parties) democrats 239–40
Enlightenment 7, 305–6 improving already competent
entertainment 345–7 voters 240–2
epistemic abstinence 304–5, 365–6n (see also political upshot 242–3
epistemic agnosticism; epistemic rendering initially incompetent voters
modesty) competent 242
epistemic agnosticism 365–7 (see also democracy vs 13, 111n, 223, 225–43, 244
epistemic abstinence; epistemic epistemic considerations beyond
modesty) competence
epistemic authority 41n, 308–11 larger groups might outperform smaller
epistemic bottlenecks 2n, 13, 223–4, 260, 319 ones 237–8
coalition government with small, pivotal political upshot 238–9
parties 274, 275–6 smaller groups might outperform larger
legislative committees 264, 265–6, 271 ones 235–7
party whips 275 epistemic effects of expanding the
presidential vetoes 276, 281–3 electorate
small upper houses 264–5, 266–9, 271–2 each individual’s competence is
strong leaders 261, 262n, 263–4 unknown 229–30
supermajority rules 276–81 enfranchising batches of voters with
epistemic competence see competence; heterogenous individual
Competence Assumption; competence competence 230
improvement; government enfranchising voters with heterogenous
competence; minimal competence knowledge bases 231, 238–9
requirement only average individual competence is
epistemic democracy 7, 8, 12, 13, 15, 26n, 38, known 230
41–2, 50, 52–3n, 54n, 94, 132n, 133n, epistemic logic of enfranchising the less
147, 165, 196–7, 199, 203–6, 225, 238, competent 231–3
244, 301–2, 305, 312–14, 321, 316, 319, equality see free and equal people
432 Index
errors 1n, 2, 4, 56n, 85, 89–90, 92, 94, differential group selection
100–4, 107, 125, 157, 284, 285n, 312, competence 217–20
315, 324n, 365n strategic leadership and
cognitive 46 (see also heuristics; coordination 220–1
psychological heuristics) unequal factional competence 211–12
‘fundamental attribution error’ 248 uniform voter competence 209–11
random 4, 6n, 135n, 196 fake news 339, 355n, 356, 357, 360, 361, 363,
systematic 45, 52, 54n, 56, 204, 205, 305 365 (see also ‘alternative facts’;
trial and 149 deception; lies; ‘post-truth’)
Eskridge, W. N. Jr 187–8 fallibility 46, 73, 97, 142, 282, 304, 306
Estlund, David 11n, 12, 26n, 38n, 52n, 53n, 56, Fallows, James 324n, 362
65, 213n, 225 false consciousness see interests, true
evidence 3, 5, 8, 10–11, 13, 16, 26, 28, 32, 55, 62, falsehoods see lies; truth claims
68, 71–6, 76–8, 80–3, 87, 94, 97n, 98, Farage, Nigel 325n, 327n
99–100, 104, 106, 107, 135, 136–7, 139, Fearon, J. 142–3n, 135n
140, 141–2, 152–3, 154, 157–9, 160, 166, Fedderson, T. 47n
169, 172, 185–8, 189n, 209n, 236, 237, federalism 115, 116, 118, 239
280, 288, 296, 304, 306, 313, 314–18, Federalist Papers 96, 98, 139n, 144, 208n, 226,
322, 325, 327, 337, 340, 347, 349, 357, 246n, 248, 249n, 250–3, 267n, 273n,
361, 362, 363, 365 (see also best 277n, 281, 282n, 285n, 291n (see also
responder; Best Responder Corollary; American Founders)
choice situation; decision situation; Feldman, S. 179n
deliberation) Ferejohn, J. A. 187–8
access to see access, to evidence filter bubbles 356 (see also bubbles)
attempts to discredit 322–3 (see also fitness landscapes 112–14
epistemic malevolence) Fishkin, J. S. 320n
improving the evidence base 293–4, 320 Flynn, Michael G. 339
opinion leader’s interpretation of 70, 73–4, ‘folk epistemology’ 133n
75, 76, 165–6, 170–1, 176 following leaders see opinion leaders
experts 3, 26, 81–2, 106n, 111, 131, 223, Ford, Gerald 180
237, 238, 239–43, 247n, 283, 316, forecasting 3–4
317, 319, 320, 321, 326, 358 Fox News 71, 98, 99, 100, 139, 176, 177,
(see also epistocracy) 323n, 364
bias 82, 237 franchise 10, 13, 95, 223, 229–33, 238–9, 240,
committees 136 299 (see also electorate)
epistemic deference 155n, 319 restricted 10, 230, 252n, 295–6
experts propose, whole groups Frankfurt, Harry 347, 348n
dispose 126–8 Franklin, Benjamin 7
panels 290n, 294 Frazer, M. 226n
think tanks 289 free and equal people 195, 223n, 234
expressive voting 87, 333, 335, 341–2 (see also egalitarian distribution
expressing emotions 342–3 of decision-making power)
expressing identity 343–5 free speech 108, 140, 297
having fun 345–7 free-riding 247
Fricker, M. 137
Facebook 341, 354–7, 360n, 363n, 365 fundamental attribution error 248
(see also social media)
fact checkers 328–9, 330, 349, 357, 358n, Galbraith, John Kenneth 164
363n, 364, 365 Galston, William 308
factionalism 13, 138, 147, 208–23 Galton, Francis 87, 139n
differing interests 208–9 game theory 47–50, 134n
epistemic solidarity and block ‘garbage in, garbage out’ 141, 260
voting 215–16 Garrett, Major 364n
guide to notation used 370 Gastil, J. 110n
uncertainties 216–17 Gaudet, H. 175
differential abstention from epistemic Gaus, G. 247n
solidarity 217 global politics 301
Index433
Gentzkow, M. 360n Hirshleifer, D. 154n
Gerson, Michael 338n Hochschild, J. L. 179n
Gingrich, Newt 338n, 342–3, 346, 349n Holmes, Oliver Wendell 149n
Goldman, A. I. 91n, 211n Hong, Lu 97–8, 104–7, 109, 115n
Gomis, Elvin 72n Horner, Paul 356n
good faith 19, 150, 322, 341 House of Commons see UK Parliament,
Google 355, 357 House of Commons
Goren, P. 179n House of Lords see UK Parliament, House
Gove, Michael 326, 327n, 332, 358n of Lords
‘government by discussion’ 137n (see also House of Representatives see US Congress,
deliberation) House of Representatives
government competence 2, 309, 312–13 Hughes, Scottie Nell 362
(see also competence; output Humberstone, I. L. 207n
legitimacy) Hume, David 355n
Grofman, Bernard 233n, 234, 235 Hume’s Law 38
Grofman–Dummkopf–Witkopf theorem 20, Hurley, N. L. 179n
228, 250n, 251n, 252n
Grotius, Hugo 304n ‘ideal speech situation’ 134
group competence 13, 19n, 20–2, 25–6, 36, 54, identity politics 149, 336, 343–5, 347
61, 75, 80–1, 83, 85, 87, 92–3, 102–3, (see also groups, identity; political
122–4, 135–6, 139, 141–2, 145, 163n, 177, parties, party identification)
190, 213n, 229, 232, 234–5, 242, 253n, ideology 55, 61, 71, 104, 181–2, 285
258, 297–9, 310 (see also Condorcet ignorance 2, 9, 53, 56n, 83, 86–91, 92, 151, 178,
Jury Theorem; collective competence) 192, 225–6, 321, 324n (see also
estimated by Monte Carlo competence; Competence Assumption;
simulation 371–2 incompetence; pluralistic ignorance)
formula 20–1 rational voter 86–7
guide to notation used 369 ‘impartial culture’ model 21n, 33, 34n
groups (see also classes; electorate; incompetence 1–2, 54, 57–60, 65, 88, 92,
factionalism; group competence; 102–3, 147, 166, 168, 193–4, 230, 238,
political parties) 242, 246, 299, 309, 314, 316 (see also
conformism 139, 159n (see also conformity) collective competence; competence;
formation 216–21, 226n Competence Assumption; government
identity 149, 179n, 185n, 217–20, 336, competence; random, worse than)
343–5, 347 (see also factionalism) incomplete agendas see agenda, incomplete
larger 237–8, 245–53, 316–17 (see also independence (see also Independence
Condorcet Jury Theorem) Assumption)
polarization 136, 139, 140n, 170–2 among votes 6, 28, 43, 50, 51, 68–9, 72, 143,
smaller 235–7, 245–53, 317–18 (see also 168–72, 192, 296 (see also common
Condorcet Jury Theorem) causes; dependence; Independence
subgroups 124–6 Assumption; interdependence among
‘Guardianship Argument’ 233 votes; opinion leaders)
conditional (see also conditionalizing)
Habermas, Jürgen 134, 141, 290, 304n on all common causes 74–6
Hale, Matthew 97n, 157n on the decision situation 77–8
Hamilton, Alexander 144n, 248–9, 252, 277n, on the evidence 72–3
281, 282 on opinion leader’s interpretation of
Hanson, R. 41n evidence 73–4
Harris, Cameron 360n on the state of the world 69–70
Hayden, Michael 363 failures of see bias; cascades; common
Hayek, Friedrich A. 149, 226n causes; cues; decision situation;
heterogenous competence see competence, evidence; ideology; dependence;
heterogeneous heuristics; opinion leaders;
heuristics 11, 55, 71, 76, 103, 104, 106, 139, 184, psychological heuristics
187–8, 231, 301 (see also bias; cue-taking) interaction with competence 55–60, 236
‘hidden profile’ paradigm 137 not unconditional 69
hindrances see institutional hindrances statistical 10, 18, 29, 37, 55, 68
434 Index
Independence Assumption 8, 10–11, 12, 15–16, weeding out bad, confusing
18–19, 26, 54, 67–82, 138–9, 140n, 151, alternatives 290–3
165, 187, 214n, 236–7, 295–6, 299, 314, institutional hindrances 260–87
340 (see also Best Responder coalition government with small, pivotal
Corollary; Condorcet Jury Theorem) parties 275–6
conditionalizing legislative committees 264, 265–6
independence is not unconditional 69 mitigating factors: Selection and
on all common causes 74–6, 77n Deliberation Effects 270–1
on available evidence 71–3 mitigation: political cooling-off periods
on opinion leader’s interpretation of the 283–7
evidence 73–4, 75, 76 party whips 272–5
on state of the world 69–70, 71, 72 presidential vetoes 277, 281–3
deliberation to reduce dependence 138–40 small upper houses 264–5, 266–9
experts vs lay persons 81–2 mitigating factors: selection and
failures of independence 314 (see also Deliberation Effects 269–72, 285
dependence; interdependence among strong leaders 261, 280
voters; common causes) mitigating factors: many independent
living with dependence 60–2 leaders 263–4
relation between competence and party leaders dictating party policy 262–3
independence 55–60, 236 supermajority rules 276–81
worries over independence 54–5 interdependence 8, 11, 71, 97, 295, 314
independence is not absence of (see also common causes;
interaction 68–9 Independence Assumption)
mechanisms to increase interests 9, 13, 42, 47, 56n, 63n, 82, 87, 89, 92,
independence 294–8 117, 147, 150n, 180n, 181n, 195–7, 200n,
interdependence among votes 8, 11, 71, 97, 98, 206n, 231, 233n, 239, 240, 247n, 248,
139, 153, 295, 314 (see also common 254, 277n, 280, 282, 291, 296n, 299,
causes; Independence Assumption; 307, 309, 312, 321, 332, 335, 336, 337–8,
opinion leaders) 345, 347, 355, 367
individuality 97 factional interests 208–21
inertia 112, 116–18 guide to notations 370
information, private 88, 91, 135n, 136–7, 154n, private interests 6n, 208, 299
162, 307 (see also private signal) public interest 208, 209, 282, 299
informational cascades 151, 154, 155, 156, sectoral interests 291, 299n
159n, 286 (see also cascades) true 200n, 209–20 (see also paternalism)
informational shortcuts 9, 178, 179, 183, 184n Internet 323, 339, 354–7, 358n, 361 (see also
(see also cue-taking) Facebook; Google; Twitter; social
‘information bubbles’ 354n, 355 media)
(see also bubbles) IQ 249, 321n
information pooling see pooling information isocompetence 20, 228
institutional aids 288–300
mechanisms to increase Jefferson, Thomas 7, 10, 150, 277n, 278–9, 295
competence 298–9 Jeffrey, A. 126n
mechanisms to increase Johnson, Boris 326, 327n, 339, 345
independence 294–5 judges 65n, 71, 136n, 149–63, 260n, 303
proliferation of independent opinion (see also courts; judicial behaviour;
leaders 297–8 precedents; stare decisis;
public funding of elections and public traditionalism)
broadcasting 296–7 judicial behaviour 71, 150–63, 188–9n, 260n,
restricted franchise 295 303n (see also precedent; stare decisis)
secret ballots 296 juries 38n, 97n, 133n, 136n, 139n, 142, 247n,
mechanisms to increase sincerity 278, 280
299–300 jury theorems (see also Best Responder
mechanisms to make decision situation Corollary; Condorcet Jury theorem)
more truth conducive 8, 81–2, 288 with heterogeneous voter competence
finding new, better alternatives 288–90 23–5
improving the evidence base 293–4, 320 with many alternatives 26–31
Index435
Kaniovski, S. 100n, 101n, 104n Lippert-Rasmussen, K. 229n
Kant, Immanuel 39, 320 List, Christian 67, 72n, 128n, 209–10,
Karotkin, D. 229n, 230n 259n, 278n
Keillor, Garrison 334n localized search 96–7, 98n, 110–16, 119–20,
Keller, Bill 358n 291, 317 (see also search; search parties)
Kennedy, J. F. 88n, 143n, 284, 323n diversified (or many random) search
Keynes, John Maynard 220 parties 114–16
know how 4 incomplete agendas 112–14
knowledge vs competence 91 recognizing the best 119–20
Kuklinski, J. H. 179n transition costs 116–19
Locke, John 277n
Ladha, K. K. 62n, 140n Lockhart, Joe 353n
Landemore, H. 12, 40n, 231n Lomasky, L. 247n
Lane, M. 41n lower house 265, 269, 285, 287n (see also
Laplace, Pierre Simon 32n bicameralism; legislatures; UK
Lau, R. R. 121 Parliament, House of Commons; US
law of large numbers 29, 50–1, 68, 198, 214 Congress, House of Representatives)
(see also Condorcet Jury Theorem) low-information rationality 178–80, 182
Lazarsfeld, Paul 175, 177 Lu, R. 333n
leaders see elites; opinion leaders Lukes, Steven 164
leadership 8, 104, 164–5, 171, 175–6, 206, Lupia, A. 87, 91n, 179n
220–1, 261, 262n, 264, 327n, 340–1, 361
(see also opinion leaders; strong Maass, A. 262n, 289n
leaders) Macdonald, Sir John A. 285n
learning from experience 95, 223, 239–2 Madison, James 7, 139n, 226, 246n, 267n,
legal theorists 97, 150, 156–7, 309 277n, 285n
legislative committees 7, 264–6, 289–90, 293, magistrates 188–9n (see also judicial
300, 317 behaviour)
epistemic bottlenecks 264, 265–6 majoritarianism, moral see moral
Selection and Deliberation Effects 270–1 majoritarianism
legislatures 3n, 101, 123, 163, 250, 254, 260, majority rule 17, 20, 23, 25n, 27, 31, 32, 34, 41,
261n, 262, 264, 293, 300, 345, 355, 364 47, 196, 209n, 213n, 214n, 226n, 227,
(see also UK Parliament; US 235n, 277n, 278n, 280, 284, 322
Congress) (see also vote aggregation rules)
bicameral 261n, 264, 266–70, 271–2, 285, plurality rule as extension with more than
287n, 289 (see also lower house; upper two options 17n, 27, 34
house) special majorities see supermajority rule
coalitions 261, 262, 263, 272, 273–6 majority tyranny 196, 284n
committees 265–6, 268–71, 290–1, 317 matters that should not be put to the
party whips 272–5 vote 307, 309
procedures 276–87 majority vote 5, 9, 10, 19–24, 26n, 27, 28–32,
legitimacy 13, 53n, 135n, 164, 238, 308–9, 34n, 41–4, 47, 48, 51, 52, 53n, 54, 57–61,
312n, 313, 323n 72–3, 76–9, 81, 87n, 89–90, 92, 99,
democratic 260n, 321 101–3, 107, 110n, 127, 135n, 143, 152,
minimal competence requirement 1, 154–61, 166–9, 171–7, 182, 186–7,
312–13, 321 (see also government 189–94, 196n, 197, 199, 201, 202, 209,
competence) 210–14, 218, 225, 226n, 227–9, 231–9,
‘output legitimacy’ 309, 312 241, 246–7, 250–1, 255–9, 261–4, 266,
Lewandowski, Corey 351 267n, 268–70, 273–6, 277, 280, 281,
liberal democracy 164, 208n, 211 (see also 283, 292–3, 305–7, 309, 310–11, 314–16,
democracy) 318–19, 330–2, 336, 338, 340, 343–5,
lies 49, 323n, 324, 325–31, 339–40, 347–51, 348 (see also majority rule; vote
359, 362–7 (see also ‘alternative facts’; aggregation rules)
deception; ‘post-truth’; Sincerity Marti, J. L. 135n
Assumption; strategic signalling; Marx, Karl 208, 209
strategic voting; truth claims) mass democracy 9 (see also democracy;
Lindsay, A. D. 137n electorate; franchise)
436 Index
masses 53, 86, 141, 147, 208–21, 246–8, 251 Negroponte, Nicholas 355
(see also class; electorate; elites) news media 71, 94, 98–9, 139, 168–9, 177, 179,
McGovern, George 1, 2 354–7, 358n, 360–1, 362, 363, 364
mean competence 23–5, 54n, 83n, 92, 94, (see also fake news)
227n, 231, 298 (see also competence; broadcast 296–7, 328, 354, 360n
Competence Assumption) public broadcasting 296–7
media see news media; social media television 177, 354n, 357n, 363, 364n
Mencken, H. L. 86 (see also Fox News)
Mendelberg, T. 186n cable news 60, 346
Mercier, H. 137n, 140n, 294n entertainment 345–6
Mill, John Stuart 9, 26n, 53n, 95, 97, 133n, newspapers 10, 179, 220n, 297, 345, 354,
137n, 141, 233, 240, 245n, 293, 296, 299 360n, 363n, 365
Miller, David 38 online 354–5, 356, 357, 358n, 360–1,
Miller, Nicholas 209 363n, 365 (see also Breitbart; Google;
minimal competence requirement social media)
for government 1, 312–13, 321 Nixon, Richard 1, 2, 183n, 351, 359n
(see also government competence) ‘no criticism rule’ 136n, 137 (see also
for voters 9, 10, 66, 85, 111, 209, 239, 321 brainstorming)
(see also random, better than; ‘noise’ 87–8, 89, 98n, 179n, 349n
competence; Competence (see also random)
Assumption) Non-asymptotic Result 19, 24, 25
‘minipublics’ 110n, 290, 294, 300, 320n (see also Condorcet Jury Theorem)
(see also deliberative democracy; normal approximation 20
deliberative institutions) Novitz, D. 350n
Misak, C. 133n
Montesquieu, Charles-Louis de Secondat, Obama, Barack 4n, 308, 342, 357, 366n
Baron de 225 Obamacare 334, 337, 352
Money, J. 261n Ober, J. 128n
Monte Carlo simulation 371–2 ‘open-secret voting’ 139n, 299n
Monti, Mario 321 open voting 296, 299
moral claims 11–12, 38–42, 53n, 195n, 303, opinion leaders 11, 13, 54–5, 57, 58–60, 60–1,
354, 365n (see also value judgements) 68, 70–1, 74–6, 77, 81, 104, 139, 147,
moral conventionalism 40, 42 164–77, 189, 206, 221, 247, 314–15, 317,
moral majoritarianism 41–2, 147, 196, 211 319, 339–41, 367 (see also
moral realism 38–9, 42 Independence Assumption;
moral separability 40–1, 305n voter independence)
motivated reasoning 41n, 53, 140, 305n, 320, deliberate falsehoods 339–41 (see also
324, 340n (see also confirmation bias) truth claims)
Moynihan, Daniel Patrick 308 encouraging independent leaders 297–8
Muirhead, R. 179n guide to notations 369
multi-alternative competence 27–31, 121–3 independent voters and 169–72
(see also competence; Competence interpretation of the evidence 73–4, 75, 76
Assumption) local 175–7
multiculturalism 96, 298, 299 many multiply mediated opinion
multiple truths 44–5, 47, 97, 105n, 220 leaders 175–7
(see also ‘alternative facts’; multiple opinion leaders 168–9, 189, 264,
precedent, multiple precedents; 297–8, 315, 341
truth claims) negatively correlated 170–2, 315, 341
myopia 112–14, 116, 117, 120 positively correlated 169–70
partial followership 172, 264, 315
Nash equilibrium 47–9 competent opinion leaders 174–5
negatively correlated votes 97, 100–8, 135, 140, opinion leaders of purely random
236 (see also diversity) competence 173
epistemic benefits 100–3 single opinion leader 61, 74–6, 104, 165–8,
plausibility 103–4 186n, 189, 285
negatively correlated opinion leaders 170–2 strategic leadership and coordination 220–1
Index437
strong leaders 261, 263, 280, 314 party whips 272–6 (see also block voting;
mitigating factors: many independent political parties)
leaders 263–4 paternalism 304, 309
party leaders dictating party Patterson, T. E. 366n
policy 262–3 Peirce, C. S. 141
opinion polls 3, 180, 227n, 339, 340 (see also Pelosi, Nancy 262n
Deliberative Poll; polls) perceived consensus 140 (see also consensus)
opposition parties 108, 137n, 170, 291, 361, 362 Pesendorfer, W. 47n
(see also dissent; disagreement; Peter, F. 26n
political parties) Pinter, Harold 303
options pivotal actors
binary choice 8, 17n, 18, 23, 26–34, 64n, 65, coalition parties 272, 274–6
98n, 122, 126–7, 170n, 189n, 196n voters 47, 49, 86, 333, 335
competence varies with number of Plato 9
options 30–1, 121–3, 126–7, 292 pluralism 13, 147, 195–7, 208, 293, 319–20
considering a few at a time 123–4 (see also priorities; value judgements)
(see also agenda, reduction) democratic competition over values and
correct option missing 42, 43–4, 63, 112–14, priorities 206–7
119–20, 143, 288–90 differing priorities 205–6, 335–8
expanding see agenda expansion differing values 40n, 197
ill-formulated 63–6 baseline scenario 197–8
more than two 26–7, 32, 34, 204 (see also competence asymmetries 200
binary choice; disjunction problem) democratic upshot 205
multiple equally good 44–5, 220 overlapping consensus 203n
reducing see agenda, reduction variations 198–205
Orwell, George 323n, 358 institutional aids 293
Osborn, A. F. 136n ‘reasonable pluralism’ 195
out-of-court settlements 156n plurality votes 5, 15, 28n, 29–31, 36, 42, 64, 98n,
‘output legitimacy’ 309, 312 121–2, 127, 130, 131, 135n, 196, 203, 204,
‘overconfidence bias’ 111n (see also bias) 206, 209, 220, 232n, 254n, 278n, 292
overlapping consensus 203n plurality rule 16, 17n, 27–8, 29, 32, 33, 34,
35–6, 41, 63–4, 98n, 196, 203, 209n,
Packard, Vance 164 211n, 331n (see also vote aggregation
Page, Scott E. 97–8, 104–7, 109, 115n, 119n rules)
Paine, Tom 7 extension of majority rule 17n, 27, 34
Palin, Sarah 88n Plurality Vote Reliability Result 29
Pariser, E. 356n Plurality Vote Asymptotic Result 29
parliamentary democracy 108, 262, 264 polarization 136, 139, 170–2, 315
(see also democracy) (see also factionalism)
Paroush, J. 229n, 230n policy landscape 112–14, 116, 117, 118, 119, 274
partial information 105–6, 179n (see also political capital 117–18
agenda, incomplete; evidence) political cooling-off periods 267n, 282–7
participatory democracy 95, 239–40, 245, political engagement 95, 137, 142, 354n
254n (see also democratic theory) political feasibility 118, 119
partisanship 179n, 187, 283, 291, 340n, political inertia 112, 116, 117, 118
351, 354n political influence 178 (see also opinion
party democracy 273n, 293 , 297 (see also leaders)
democracy; parliamentary political leadership 164 (see also opinion
democracy; political parties) leaders)
party identification see political parties political myopia 112–14, 116–17, 120
party label 91, 179, 183–4, 298 (see also political parties
cue-taking; political parties, party endorsement see endorsements
identification) opposition parties 108, 137n, 170, 291,
party line 132n, 183, 213n, 267n, 272–3 361, 362
(see also block voting; cue-taking; parliamentary democracy 108, 262, 264
party whips; political parties) (see also democracy)
438 Index
political parties (cont.) presidential vetoes 96, 277, 281–3
partisanship 179n, 187, 283, 291, 340n, (see also bottlenecks; veto)
351, 354n Principle of Insufficient Reason 30n
party democracy 273n, 293, 297 priorities 13, 42, 47, 147, 195–7, 199, 205–7,
(see also democracy) 208–9, 292, 332, 335–9, 345, 347, 349,
party identification 179 357, 367 (see also interests; political
expressing identity 343–5 parties, priorities and manifestos;
party affiliation 181, 183–4 value judgements)
reliability of party labels 183–4, 298 democratic competition over 206–7, 357
party whips 272–5 guide to notations 370
pivotal coalition parties 272, 274–6 in improving voter competence 92–4
priorities and manifestos 206–7, 292, 357 political party 292, 357
search function 291–3 (see also localized Prisoner’s Dilemma 50n
search; search; search parties) private signal 32, 64n, 153–5, 157–62, 165, 286
strong leaders 262–3 (see also information, private)
political practices 147, 163, 314, 319–21 (see also probability theory 5–6 (see also Condorcet
traditionalism; opinion leaders; Jury Theorem; law of large numbers)
cue-taking; pluralism; factionalism) ‘production blocking’ 136
polls 3, 180, 227n, 287n, 333, 334, 335, 336, protest votes 333–5, 341 (see also expressive
339–40, 343, 349, 360n, 363n, 367 voting)
(see also Deliberative Poll; opinion polls) Przeworski, A. 296n
poll tax 1 psychological heuristics 53, 184, 187–8
pooling (see also crowdsourcing; vote (see also biases; heuristics)
aggregation rules) public broadcasting 296–7, 298 (see also
information 3, 4–5, 9, 54, 136, 141–2, news media)
144, 147, 150, 159, 213, 215, 321, 322, 323 public debate 82, 132n, 136, 137, 141, 294,
proposals 136n 306, 357 (see also deliberation)
voter judgements 15, 46, 49, 142n, 144, public ignorance see ignorance
154–5, 218–20, 262n, 264, 322 public funding of elections 296–7
Popkin, S. L. 91n, 93n public reason theories 330
popular sovereignty 164
populism 165n random 104, 230, 313, 345
epistemic 358 agenda 45
positively correlated 231n (see also common better than 9, 18, 23, 26, 28n, 29, 30, 45, 50,
causes; evidence; opinion leaders) 52, 59–60, 65–6, 69, 77, 85, 88, 92, 94,
opinion leaders 169–70 103, 111, 121–2, 126, 129, 134n, 138, 143,
votes 68, 69, 97, 98–100, 104, 135 169, 173, 180, 188, 191, 209, 213–14, 227,
Posner, R. A. 321n 230, 234, 242, 246–7, 257, 263, 297,
‘post-truth’ 324, 342 298, 305, 306, 310, 315, 316, 332, 336–7,
precedent 55, 149–63 343 (see also competence; Competence
epistemic costs of complete Assumption)
deference 151–3 error 4, 6n, 54n, 88, 135n
hiding precedents 156–7 noise 88n, 89
informative precedents 159–61 opinion leaders 173–4
multiple precedents 161–2 searches 114–16
partial deference 153–6 tie-breaking 20n, 34–5, 127n, 189, 265n
resisting precedent 157–9 votes 9–10, 18, 28n, 44, 45, 48n, 50, 52, 53,
stare decisis 150–1, 152, 163 57, 58, 59, 65, 76, 88, 90, 129–31, 165,
prediction markets 3–4 173, 182, 189, 191, 192, 204, 305, 315, 343
preference 5, 135n, 139n, 164, 165n, 206n, 207, voting order 286n
221, 255 worse than 9–10, 37, 52–3, 57–8, 190, 192–3,
beliefs vs preferences 43, 46–7, 163 230, 242, 246, 314–15, 351 (see also
collective 90–1 competence; Competence
policy 121n, 183n, 345 Assumption; incompetence)
rankings 21n, 33, 34, 35, 46 rational voter ignorance see ignorance,
prejudices 11, 53–5, 56n, 62n, 88n, 225, 226, rational voter
237, 239 (see also bias) Rawls, John 68n, 138n, 195, 203n, 283–4, 304
Index439
Raz, Joseph 309n Shapley, L. S. 233n, 234, 235
Reagan, Ronald 181n, 323n, 342n, 350 Sheagley, G. 184n
realism, moral see moral realism shortcuts, informational see informational
‘reasonable pluralism’ 195 (see also pluralism; shortcuts
value judgements) Sidgwick, Henry 365–6n
Rehfeld, A. 254n Silverman, C. 354n
Reinganum, J. F. 154n sincerity 5, 9, 10, 15, 19, 32, 37, 45–50, 135, 153,
religious convictions 195n, 304–5 196, 224, 279, 322, 330, 332, 333n, 335,
(see also value judgements) 341, 342 (see also deception; lies;
representative democracy 13, 223, 244–60, Sincerity Assumption; strategic
288–90, 316, 317, 345 (see also signalling; strategic voting)
democracy; direct vs representative deliberation induces sincerity 50, 140–1
democracy) failures of sincerity 37, 45–9
delegate versus trustee mechanisms to increase sincerity 299–300
representatives 254–9 sincerity as a default 49
Deliberative Effect among trustee-style ‘sincerity bias’ 49n
representatives 256–7 strategic voting 19, 46, 47n, 48, 49, 50
epistemic cost of bunching voters into (see also strategic signalling; strategic
constituencies 255–6 voting)
mixed assemblies with both delegates and Sincerity Assumption 19, 29, 45, 125n, 153, 261
trustees 257–9 (see also Condorcet Jury Theorem; lies;
respecting tradition see traditionalism deception; strategic signalling;
restricted franchise 295 (see also electoral strategic voting)
reform; electorate; epistocracy; small government 279
franchise) small groups 93, 124–8, 136, 234, 249, 259, 317
right to vote see franchise committee vs legislature 265–6
Riker, W. H. 27n constituencies vs whole electorate 255–6
Robespierre, Maximilien 7 Electoral College vs electorate 144–5,
Roosevelt, Franklin 293 249–51
Rousseau, Jean-Jacques 6n, 138, 208, 211, epistemic elites 223n, 233–5, 227–9
213n, 240n, 266 (see also epistocracy)
Rove, Karl 362 epistemically less reliable 22, 227–9, 250
rugged landscape 112–13, 117–18, 317 (see also Condorcet Jury Theorem)
(see also fitness landscapes) epistemically more reliable 124–6,
rule of 97n, 149n, 150 235–9, 317
‘rush to judgment’ 107, 284–6 expert committee 126–8
Russia 360, 361, 362, 363 legislature vs electorate 245–54
Rust Belt 336, 337, 344, 352 more conversable 143–5, 236, 244, 269,
270, 288–9, 320
Scharpf, F. W. 309 parties vs whole legislature 272–5
Schmitt, Carl 318n party leader vs whole party 262–3
Schumpeter, Joseph A. 164, 247n pivotal coalition parties 275–6
search engines 355 (see also Google; search; subgroups propose 124–6
search parties) upper vs lower chamber of
search 13, 49, 96–7, 98n, 110–20, 291, 317, 339, legislature 266–9
355–6 (see also localized search; search Smith, L. 154n
parties) social constructivism 303–4
diversified 114–16 social engineering 108–9, 229
search parties 110, 112, 114–16, 119, 120, 291, 317 social media 339 (see also Facebook, Twitter)
(see also localized search; search) social status 226 (see also class)
Searle, John 40 ‘hidden injuries of class’ 344
secret 2, 156, 232, 299n solidarity see block voting; epistemic
ballot 296, 299 solidarity; party line
votes 139n, 287, 296n, 299n Somin, I. 86n, 88n
Selection Effect 248–51, 269–72 Sorensen, Ted 154n, 284n
Sen, Amartya 313n sovereignty 325n, 336 (see also consumer
Shapin, Steven 133n sovereignty; popular sovereignty)
440 Index
Spenkuch, J. L. 49n multiple precedents 161–2 (see also
Sperber, D. 133n, 137n, 294n, 350n multiple truths)
Spicer, Sean 359n resisting precedent 157–9
stare decisis 150–1, 152, 163 (see also precedent; discerning traditionalists 159–61
traditionalism) epistemic costs 151–3, 163
Stasser, G. 137n partial deference 153–6
‘state of the world’ 10, 17–18, 23n, 37, 52, 55, precedent 150–1
69–75, 77–81, 97n, 99–101, 169, 174n, transition costs 116–19
185–6, 269n, 294 Trump, Donald
status quo 113, 117, 268n, 270n, 277–83, 287, Art of the Deal 350
338 (see also traditionalism) election victory 44n, 302, 322, 324
bias 117, 277–83 affective explanations 341–7
despair with the status quo 338 differing priorities 335–8
Steinberg, Tom 356 epistemic agnosticism 366–7
strategic signalling 332–5, 341 (see also epistemic democracy under threat 367
expressive voting; sincerity; Sincerity epistemic insouciance 347–54
Assumption) epistemic malevolence 358–65
strategic voting 19, 46n, 47–50, 261n, 299, false claims 327–30, 339–41
333, 335 (see also deception; lies; strategic signalling vs sincere
sincerity; Sincerity Assumption) voting 333–5, 341
strong leaders 261–4, 280 (see also opinion influences 323n
leaders) trustee-style representation 13, 244, 248n,
many independent leaders 263–4 254–7, 268, 270
party leaders dictating party policy 262–3 mixed assemblies with both delegates and
subvotes 123–4, 125n, 126 (see also agenda, trustees 257–9
reduction) truth claims 2, 4, 6, 10, 18, 26, 28, 32, 38–44,
Sunstein, Cass 3n, 96, 136n, 139n, 150, 154n, 46n, 47–53, 55, 69, 73–5, 78n, 80, 99,
165n, 184n, 355 137n, 141, 185, 186n, 226, 246, 294,
supermajority rules 276–82 (see also vote 295n, 303–4, 307–8, 313, 323n, 331, 339,
aggregation rules) 355n, 357–60, 361n, 362–7 (see also
Surkov, Vladislav 362 ‘alternative facts’; deception; lies;
surprises 307, 360 multiple truths; ‘post-truth’; sincerity)
Swaim, B. 353n Brexit lies 325–7, 339
Sweden 249 dangers of truth-seeking in politics 305–7
Swift, Jonathan 132n, 324n, 327n, 348n deliberate falsehoods 323–5, 341
Switzerland 240n undermining the CJT 330–1, 340–1
epistemic agnosticism 365–7
taking cues see cue-taking epistemic insouciance 347–54
Talisse, R. B. 38n, 133n, 137n epistemic malevolence 358–65
Talley, E. 154n, 159n fact checkers 349, 357n, 363, 364, 365
Tanasoca, Ana 239n fake news 339, 356, 357, 360n, 361n, 363
technocrats 321 (see also experts) honest opinion 108, 139, 333, 347, 348n,
Thatcher, Margaret 1, 2 363–4 (see also conformity; sincerity;
think tanks 289 (see also experts; strategic signalling; strategic voting)
technocrats) interests 200n, 209–20
Thompson, C. 119n, 233n, 234n limitations of truth-seeking in
Tirole, J. 273n politics 303–5
Titus, W. 137n persons with epistemic authority 308–11
tobacco industry 322–3 in politics 11–12, 13, 15, 24n, 39–43, 46n,
tolerance 293, 298, 299 208–9, 303–11, 312, 313–14, 321,
topic-specific competence 25–6 323–4, 326, 330, 332, 340, 341–2, 347,
(see also competence; Competence 348–51, 354, 356, 358, 359n, 361, 363–7
Assumption) (see also ‘alternative facts’; lies;
traditionalism 13, 147, 149–63 ‘post-truth’)
averting epistemic damage dangers of pursuing 301–2, 304, 305–11
discerning traditionalists 159–61 ‘post truth’ 324
hiding precedents 156–7 ‘true fictions’ 347n, 350–1
Index441
truth aptness 11, 43, 307 (see also differing priorities 335–8
preference; value judgements) epistemic agnosticism 366–7
truth value epistemic democracy under threat 367
of ‘clickbait’ 356 epistemic insouciance 347–54
of expressive voting 342 epistemic malevolence 358–65
of preferences 43, 355 false claims 327–30, 339–41
of values 11–12, 53n (see also moral popular vote vs Electoral College 331–2
conventionalism; moral realism) strategic signalling vs sincere voting
Trump lies 327–30, 339–41 333–5, 341
‘tyrannies of “truth” ’ 306
who should decide 308–11 value judgements 2, 11–12, 13, 15, 37, 38–42,
truth-conducive decision situations 8, 81–2, 43, 47, 53, 140, 147, 149, 195–207, 209,
314, 315, 316, 318, 320, 324, 292, 305, 307–8, 332, 335–8, 345, 347,
(see also Best Responder Corollary; 367 (see also truth claims)
institutional aids) matters that should not be put to a
guide to notations 370 vote 307–8
truthfulness see sincerity moral conventionalism 40
‘truthful hyperbole’ 350, 351 moral majoritarianism 41–2
truth-tracking 13, 15, 27, 34, 35, 39, 42, 43, 89, moral realism 38–9
184, 301–2, 314, 332 (see also best moral separability 40–1
responder; competence; random, truth of 11–12, 53n (see also moral
better than; Condorcet Jury Theorem; conventionalism; moral realism)
truth claims) value pluralism 40n, 195–205
matters that should not be put to a guide to notations 370
vote 307–8 Vermeule, Adrian 139n, 154, 155n, 159n, 223n,
multiple truths 44–5, 220 224n, 260, 262n, 264n, 268n, 271n,
questions that are not truth apt 43 287n, 288n, 299n, 309n
right answer is not on the agenda 43–4 veto 106n, 109, 260n, 264n (see also
truth value 11–12, 43, 53n, 342, 355, 356 bottlenecks)
(see also preferences; truth claims; bicameral 267–71, 284
value judgements) expert 111, 131, 289
Tsebelis, G. 261n presidential 96, 144, 277, 281–3
Turgot, Anne Robert Jacques 266 Vicario, M. Del 356n
Twitter 333n, 339, 341, 344, 350n, 355, 359, Voltaire, François Marie Arouet 7
361n, 363, 365 (see also social media) vote aggregation rules
Borda count 33n, 34, 35, 36, 331n
UK Parliament competence-weighted voting rule 233–5
House of Commons 252, 266n, 274, 284 Condorcet pairwise comparison 33n, 34,
House of Lords 266n, 268n, 284–5 35, 36, 123, 331n
unicameralism 266, 267n (see also majority rule 17, 20, 23, 25n, 27, 31, 32,
bicameralism) 34, 41, 47, 196, 209n, 213n, 214n,
upper house 268n, 289 (see also Australian 226n, 227, 235n, 277n, 278n, 280,
Senate; bicameralism; Canadian 284, 322
Senate; legislatures; UK Parliament, plurality rule 16, 17n, 27–8, 29, 32, 33, 34,
House of Lords; US Congress, Senate) 35–6, 41, 63–4, 98n, 196, 203, 209n,
epistemic bottlenecks 264–5, 266–9 211n, 331n
Selection and Deliberation Effects supermajority rule 276–82
269–72, 285 votes correlated see negatively correlated
US Congress 179, 213n, 262n, 265–70, 281–3, votes; positively correlated, votes
285n, 289n voter ignorance see ignorance
House of Representatives 226, 262n, 265, voter incompetence see competence;
266–7, 269, 270 Competence Assumption;
Senate 265–71, 281n, 285, 352 incompetence
US election outcomes 88–90, 180, 182, 252 voter independence 8, 10–11, 12, 15–16, 79
based on falsehoods 323n (see also common causes; independent
Trump victory 44n, 302, 322, 324 votes; Independence Assumption;
affective explanations 341–7 opinion leaders)
442 Index
voter interdependence 8, 11, 71, 97, 98, 139, Weinstock, D. 227n
153, 295, 314 (see also common causes; Welch, Edgar Maddison 339
independent votes; Independence Welch, I. 154n
Assumption; opinion leaders) whips 272–6
voting cycles 33–5, 265, 268 Williams, Bernard 305, 313n
voting, expressive see expressive voting Wilson, D. 350n
voting mechanisms 5, 7 (see also vote ‘wisdom of the ages’ 149, 150n, 279
aggregation rules) ‘wisdom of crowds’ 77, 127–8, 138, 261, 322
approval voting 65n ‘wisdom of the multitude’ 3, 5, 86, 127, 149,
block voting 213–16 262n, 305, 322
‘open secret voting’ 139n, 287n wishful thinking 46n, 53 (see also
open voting 296, 299 confirmation bias; motivated
secret vote 139n, 287n, 299n reasoning)
voting, strategic see strategic signalling; Wittman, D. 111n, 188n
strategic voting worse than random 9–10, 37, 52–3, 57–9,
190, 192–3, 242, 246, 314, 351
Waldron, Jeremy 132n (see also competence; Competence
Warren, M. E. 110n Assumption)
Watergate 1
Weber, Max 147 Zimmerman, Neetzan 356n

Robert E Goodin - Kai Spiekermann - An Epistemic Theory of Democracy-Oxford University Press, USA (2018)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Robert E Goodin - Kai Spiekermann - An Epistemic Theory of Democracy-Oxford University Press, USA (2018)

Uploaded by

Copyright:

Available Formats

OUP CORRECTED PROOF – FINAL, 13/03/18, SPi

This is a book in democratic theory, not applied mathematics. Much valuable

Analytic Table of Contents xi

Analytic Table of Contents

List of Figures xix

xii Analytic Table of Contents

4.1.3 Moral Separability 40

Analytic Table of Contents xiii

xiv Analytic Table of Contents

11. Following Leaders 164

Analytic Table of Contents xv

xvi Analytic Table of Contents

17.3.2 Coalition Government with Small, Pivotal Parties 275

Analytic Table of Contents xvii

21. Epilogue: What about Trump and Brexit? 322

2.1 Group competence Pn for different levels of pc. 22

8.1 Rugged policy landscape. 113

List of Figures xxi

4.1 Interaction of the Competence and Independence Assumptions. 56

xxiv List of Tables

17.1 The collective competence of assemblies divided into similarly

1.1 EPISTEMIC COMPETENCE: A MINIMAL

2 An Epistemic Theory of Democracy

1.2 ASK AROUND

9 Tetlock 2005. 10 Lane 1999. 11 NIH 2013. 12 Hayek 1945.

4 An Epistemic Theory of Democracy

Crowdsourcing is an increasingly popular tool of government and public

1.3 PO OLING INFORMATION AND

Not all knowledge is propositional knowledge. ‘Know how’ is as important as

1.4 THE COND ORCET JURY THEOREM, IN BRIEF

The history of philosophy contains many allusions to the ‘wisdom of the

6 An Epistemic Theory of Democracy

The process of aggregating votes is purely mechanical, to be sure. It elides

1.5 EXTENDING THE COND ORCET JURY THEOREM

8 An Epistemic Theory of Democracy

We will hardly confine ourselves to the analysis of voting mechanisms alone,

1.6 FORESTALLING FAMILIAR OBJECTIONS

35 Condorcet 1785/1976, pp. 49–50.

10 An Epistemic Theory of Democracy

1.6.3 The Truth Value of Values

Perhaps the most fundamental objection to the application of the Condorcet

12 An Epistemic Theory of Democracy

1.7 THE STRUCTURE OF THE B O OK

This book is divided into four nearly equal parts.

43 Estlund and Landemore 2018.

16 An Epistemic Theory of Democracy

may be less than virtually certain to be correct (because Independence is

The Classic Framework

In this chapter we introduce, briefly, the classic version of the Condorcet

2.1 THE SETUP

2.2 THE ASSUMPTIONS

The CJT rests on three assumptions as follow.

2.2.1 The Competence Assumption

18 An Epistemic Theory of Democracy

remains ‘undecided’). Informally, a voter’s competence is the probability that

2.2.2 The Independence Assumption

The Independence Assumption says:

The Classic Framework 19

2.2.3 The Sincerity Assumption

The Sincerity Assumption says:

2.3 THE THEOREM

20 An Epistemic Theory of Democracy

The second result that can be proven is this:

Analytic Table of Contents xi

Analytic Table of Contents xiii

Analytic Table of Contents xv

Analytic Table of Contents xvii

List of Figures xxi