Professional Documents
Culture Documents
By
Daniel J. Solove
&
Hideyuki Matsumi
ABSTRACT
This Essay critiques a set of arguments often made to justify the use of AI
and algorithmic decision-making technologies. These arguments all share
a common premise – that human decision-making is so deeply flawed that
augmenting it or replacing it with machines will be an improvement.
In this Essay, we argue that these arguments fail to account for the full
complexity of human and machine decision-making when it comes to
deciding about humans. Making decisions about humans involves special
emotional and moral considerations that algorithms are not yet prepared
to make – and might never be able to make.
It is wrong to view machines as deciding like humans do, but better because
they are supposedly cleansed of bias. Machines decide fundamentally
differently, and bias often persists. These differences are especially
pronounced when decisions have a moral or value judgment or involve
human lives and behavior. Some of the human dimensions to decision-
making that cause great problems also have great virtues. Additionally,
algorithms often rely too much on quantifiable data to the exclusion of
qualitative data. Whereas certain matters might be readily reducible to
quantifiable data, such as the weather, human lives are far more complex.
Having humans oversee machines is not a cure; humans often perform
badly when reviewing algorithmic output.
Introduction .................................................................................................. 4
I. Arguments About AI and Humans ............................................................ 6
A. The Awful Human Argument................................................................ 6
B. The Better Together Argument ............................................................. 6
C. The Humans in the Loop Argument ...................................................... 7
D. The Brighter Future Response ............................................................... 7
II. The Problems of Humans + Machines ..................................................... 8
A. The Limits of Quantification ................................................................. 8
1. Quantification Is Transformative ....................................................... 9
2. Emotion, Morality, and Value Judgments........................................ 10
3. The Challenge of Blending Machine and Human Decision-making 11
4. Mesmerized by Math and Machines ................................................. 12
B. Accurate – But on What Basis? ............................................................ 14
1. Is Accuracy the Right Goal? .............................................................. 14
2. The Conflict Between Accurate and Unbiased ................................. 15
III. Rethinking the Relationship Between Algorithms and Humans .......... 17
A. The Importance of Structure ............................................................... 17
B. The Virtues of Good Decisions ............................................................. 18
1Eugene L. and Barbara A. Bernard Professor of Intellectual Property and Technology Law,
George Washington University Law School. Thanks to JP Schnabel for excellent research
assistance.
2 Hideyuki (“Yuki”) Matsumi. PhD candidate/researcher at the Research Group on Law
Science, Technology and Society (LSTS) as well as at the Health and Ageing Law Lab (HALL)
of the Vrije Universiteit Brussel (VUB). Member of the New York Bar.
INTRODUCTION
Increasingly, sophisticated algorithms are being used in decisions about
people’s lives. Algorithms are deployed to help determine whether to hire
people or to lend them money—and if so, what interest rate to charge.
Algorithms are used in the criminal justice system for pretrial detention and
sentencing decisions. Algorithms are implemented in education and other
domains to make interventions and key determinations affecting people’s
opportunities and futures.3 These algorithms, often referred to as “artificial
intelligence” (AI) in popular discourse, are raising many concerns.
3See Hideyuki Matsumi & Daniel J. Solove, The Prediction Society: Algorithms and the
Problems of Forecasting the Future, https://ssrn.com/abstract=4453869.
In this Essay, we critique these arguments when they are applied to decisions
made about humans.4 We argue that these arguments fail to account for the
full complexity of human and machine decision. Although these arguments
have some dimension of truth, they are missing some key considerations that
make the reality far more complicated.
It is wrong to view machines as deciding like humans do, but better because
they are supposedly cleansed of bias. Machines decide fundamentally
differently, and bias often persists. These differences are especially
pronounced when decisions have a moral component or involve human lives
and behavior. Making decisions about humans involves special emotional
and moral considerations that algorithms are not yet prepared to make – and
might never be able to make.
5 Cary Coglianese & Lavi M. Ben Dor, AI in Adjudication and Administration, 86 Brook. L.
Rev. 791 (2021).
6 Id. at 1177-78
7 Cass R. Sunstein, Governing by Algorithm? No Noise and (Potentially) Less Bias, 71 Duke
L.J. 1175, 1177 (2022).
8 Wisconsin v. Loomis, 881 N.W.2d 749 (Wis. 2016).
9 Id. at 764.
Other courts have made similar arguments – that algorithmic tools are just
helping humans by giving them more information. For example, in
Malenchik v. State, the Indiana Supreme Court held that sentencing
assessment tools “enable a sentencing judge to more effectively evaluate and
weigh several express statutory sentencing considerations.”11
Many laws are premised on this argument, such as the European Union’s
General Data Protection Regulation (GDPR), which provides that “[t]he data
subject shall have the right not to be subject to a decision based solely on
automated processing, including profiling, which produces legal effects
concerning him or her or similarly significantly affects him or her.”12 When
a person challenges a decision based solely on automated processing, the
data controller must implement protections for the person’s “rights and
freedoms and legitimate interests,” and provide for “the right to obtain
human intervention on the part of the controller, to express his or her point
of view and to contest the decision.”13
Both the Better Together Argument and the Humans in the Loop Argument
posit that the combination of humans and algorithms will address the
deficiencies of each. These arguments are essentially the same, but from
opposite sides. The Better Together Argument posits that algorithms can be
brought in to help human decision-makers. The Humans in the Loop
Argument postulates that humans can be brought in to address any problems
with algorithms.
10 Id.
11 Malenchik v. State, 928 N.E.2d 564, 574 (Ind. 2010).
12 GDPR art. 22.
13 GDPR art. 22.
14 ORLY LOBEL, THE EQUALITY MACHINE: HARNESSING DIGITAL TECHNOLOGY FOR A BRIGHTER,
MORE INCLUSIVE FUTURE (2022).
15 Cass R. Sunstein, Governing by Algorithm? No Noise and (Potentially) Less Bias, 71 Duke
L.J. 1175, 1177 (2022).
16 Sam Corbett-Davies, Sharad Goel and Sandra González-Bailón, Even Imperfect Algorithms
Can Improve the Criminal Justice System, N.Y. Times, Dec. 20, 2017.
17 Carolin Kemper, Kafkaesque AI? Legal Decision-Making in the Era of Machine Learning,
24 Intell. Prop. & Tech. L.J. 251, 285 (2020) (“Algorithms can be designed to alleviate
biases.”); Mirko Bagaric, Dan Hunter, and Nigel Stobbs, Erasing the Bias Against Using
Artificial Intelligence to Predict Future Criminality: Algorithms Are Color Blind and Never
Tire, 88 U. Cin. L. Rev. 1037, 1040 (2019) (positing that it is “achievable” to ensure that
algorithms “do not discriminate directly or indirectly”).
1. Quantification Is Transformative
Algorithms focus on data that can readily be quantified, and this type of data
skews the output. Problems emerge when too much quantitative data is
relied upon to the exclusion of qualitative data. Not everything is readily
quantifiable.
In Seeing Like a State, James Scott argues that making things “legible”
(more capable of being understood and analyzed) involves simplification – a
“narrowing of vision” that “brings into sharp focus certain limited aspects of
an otherwise far more complex and unwieldy reality.” The simplification can
make things “more susceptible to careful measurement and calculation.”18
When modern institutions seek to “see” things by gathering quantifiable
data, they are seeing through a particular lens, and much is excluded from
the picture. Scott warns that “thin, formulaic simplifications” can lead to
tragic failures.19
18 JAMES C. SCOTT, SEEING LIKE A STATE: HOW CERTAIN SCHEMES TO IMPROVE THE HUMAN
CONDITION HAVE FAILED 11 (1998).
19 Id. at 309.
20 Quetelet quoted in CHRIS WIGGINS AND MATTHEW L. JONES, HOW DATA HAPPENED: A HISTORY
FROM THE AGE OF REASON TO THE AGE OF ALGORITHMS 26 (2023).
21 Laurence H. Tribe, Trial by Mathematics: Precision and Ritual in the Legal Process, 84
Harv. L. Rev. 1329, 1393 (1971).
22 CHRIS WIGGINS AND MATTHEW L. JONES, HOW DATA HAPPENED: A HISTORY FROM THE AGE OF
REASON TO THE AGE OF ALGORITHMS 57 (2023).
For example, ranking systems and scoring systems are designed to make
things more readily comparable. But they are reductive. Writing about the
IQ test, Professor Neil Postman argues that it “assume[s] that intelligence is
not only a thing, but a single thing, located in the brain, and accessible to the
assignment of a number.”23 Some have posited that we have multiple types
of intelligence, not just a singular type.24
Although emotion can certainly make decisions worse, it can also make them
better. The amount of desirable involvement of non-rational elements in
decision-making is a challenging question, one that is contextual and lacks a
10
definitive answer. But we posit that non-rational elements can play a positive
role when decisions are made about humans.
Humans are also capable of making decisions about moral and normative
issues in ways that machines cannot. Making normative choices can involve
tradeoffs and compromises between different values, which can be
incommensurate and conflicting.
It is easy to point to grades and scores as more objective, even though these
more quantifiable things can also be influenced by privilege and bias. In the
end, college admissions decisions, as well as many other types of decisions,
involve value judgments and tradeoffs that machines are currently unable to
make.
11
Policymakers and laws thus lack an effective plan for how humans and
machines will work together beyond vague hopes that somehow they’ll just
get along.
The hope that humans and machines can decide better together is not just
vague and unsubstantiated; in fact, strong evidence shows that there are
significant problems with combining humans and machines in making
decisions. Humans can perform poorly when using algorithmic output
because of certain biases and flaws in human decision-making. Far from
serving to augment or correct human decision-making, algorithms can
exacerbate existing weaknesses in human thinking, making the decisions
worse rather than better.
Many commentators have pointed out that people are prone to “automation
bias” – to trust algorithms without sufficient skepticism.31 As one court
declared in 1968: “Mathematics, a veritable sorcerer in our computerized
society, while assisting the trier of fact in the search for truth, must not cast
29 Rebecca Crootof, Margot E. Kaminski & W. Nicholson Price II, Humans in the Loop, 76
Vand. L. Rev. 429, 468 (2023).
30 Ben Green, The Flaws of Policies Requiring Human Oversight of Government Algorithms,
45 Computer Law & Security Rev. 1, 12 (2022).
31 Dan L. Burk, Algorithmic Legal Metrics, 96 Notre Dame L. Rev. 1147, 1161 (2020); Katrina
Geddes, The Death of the Legal Subject, 25 Vand. J. Ent. & Tech. L. 1, 36 (2023) ( “automation
bias” is “the tendency of human decision makers to accept, rather than challenge, quantitative
assessments”); Margot E. Kaminski & Jennifer R. Urban, The Right to Contest AI, 121 Colum.
L. Rev. 1957, 1961 (2021) (“Humans may exhibit an “automation bias” that creates
overconfidence in machine decisions, and an ensuing bias against challenges to those
decisions.”).
12
a spell over him.”32 Laurence Tribe contends that “the very mystery that
surrounds mathematical arguments - the relative obscurity that makes them
at once impenetrable by the layman and impressive to him - creates a
continuing risk that he will give such arguments a credence they may not
deserve and a weight they cannot logically claim.”33
The unfortunate result, then, of having humans use algorithmic output for
decisions is that humans might fail to appreciate the human influence within
that output. Of course, humans can be educated and warned about these
problems, but it is not clear that this can save the day. Humans can become
“complacent” and “overreliant” when working with an autonomous system—
a tendency that “seems to afflict experts as much as novices, and is largely
resistant to training.”39
Decision‑Making and the Control Problem, 29 Minds and Machines 555, 556 (2019).
13
Ultimately, much more study must be devoted to figuring out the proper
ways for humans and machines to interact. The mere combination of
humans and machines is far from a solution – it is a problem itself yet to be
solved.
40 Mirko Bagaric, Dan Hunter, and Nigel Stobbs, Erasing the Bias Against Using Artificial
Intelligence to Predict Future Criminality: Algorithms Are Color Blind and Never Tire, 88
U. Cin. L. Rev. 1037, 1038 (2019).
41 Matsumi & Solove, Prediction Society, supra note X, at X.
42 Matsumi & Solove, Prediction Society, supra note X, at X.
14
Quantifiable data does not consist of neutral facts; such data is created and
curated by humans, which introduces bias into the algorithm.
Hypothetically, AI optimists propose, the data could be scrubbed of all bias.
But even if the most obvious biases are removed from the data, the data still
comes from a society of humans where bias infects nearly everything.
We might value bias-free decisions rather than ones that accurately follow
from the training data. After all, one purpose of using algorithms is to
improve upon human decision-making. But this demonstrates that we might
not really want an “accurate” decision; we want an ideal decision.
Consider law school rankings. A magazine, US News and World Report, has
ranked law schools using a formula that has been vehemently criticized by
law schools as being reductive, inaccurate, and unfair. Schools have
attempted to boycott the rankings by withholding data, but the magazine still
continues to rank schools by using an even more simplistic algorithm.
15
partly because they simplify complex things and appear to make choices
easier by ignoring incommensurate dimensions. But this simplicity is just a
mirage; rankings distort in rather crude ways. One can reduce a
Shakespearean play to a simple plot summary, but this removes the most
important part – the artistry and beauty of the language. History can be
reduced to dates and events, but that does not make it useful or insightful.
Of course, people want the easy button, but life is not easy. The easy button
is a delusion. Unfortunately, many algorithmic tools are being used in this
way. They function to reduce the richness of life to certain quantifiable
elements.
Many decisions about humans are fraught with complex moral and policy
issues that elude an easy answer. For example, in sentencing, is recidivism
an appropriate factor to focus on? If so, how much weight should it be given?
Algorithms can try to predict recidivism. But they can’t answer the normative
question of how much recidivism should factor into the sentencing
calculation.
Far too often, though, the use of algorithms results in a lack of questioning
about these normative issues. The focus is solely on whether the algorithm is
accurately predicting recidivism rather than on how much recidivism should
be weighed in a sentencing decision. And the studies show that humans
might rely too much on the factors the algorithm computes because
algorithmic output is overly trusted.
Algorithms can make decisions appear more scientific, more “accurate,” and
more reliable and trustworthy. Even if the algorithms be used to make
decisions better (assuming this can be determined), they might be an overall
negative because they are dressing up a fraught activity in scientific-looking
vestments. During the heyday of phrenology – the false idea that physical
characteristics were linked to behavior – there was an attempt to dress it up
with science. This thin veneer of science was based on faulty studies.
Proponents raced forward with hope and enthusiasm; any contrary evidence
was ignored in their zeal to succeed.43
Today, algorithms are widely being used to make many decisions about
people, and it is far too soon to be doing so on such a largescale basis. This
is, in essence, a grand experiment involving millions of people—probably
hundreds of millions and perhaps even more—without the controls and
43 Pierre Schlag, Law and Phrenology, 110 Harv. L. Rev. 877 (1997).
16
When the law seeks to add a human in the loop, what does this mean?
Humans are already in the loop. Behind each machine, there are humans
who decide what data to use to train the model, what kind of output is
expected from the model, and how the algorithmic model will be used in the
decision-making process. Thus, what the law appears to mean by adding a
human is to have a human review algorithmic output at the end stage. But
this point might not be the ideal point for human involvement. Humans lack
the ability to process large datasets or understand specifically everything
algorithms are doing.46 Humans might thus not have the expertise or
cognitive ability to thoughtfully oversee algorithms. Perhaps the best place
for human oversight is in the creation and testing of algorithmic models.
Perhaps the humans involved in oversight should not be the humans at the
endpoint but somewhere at the beginning or middle. These humans are
already there, but the law often doesn’t hold them accountable.
If additional humans are to be added into the loop, the law must address
when humans should be added and what role these humans should have.
Regarding the role of additional humans in the loop, the law should identify
the goals that adding the humans will achieve and how the humans will
achieve these goals. In other words, there must be a meaningful gameplan
for how and why humans are to be added.
44Crootof, Kaminksi, & Price, Humans in the Loop, supra note X, at 437.
45Id.
46 John Zerilli, Alistair Knott, James Maclaurin, Colin Gavaghan, Algorithmic
Decision‑Making and the Control Problem, 29 Minds and Machines 555, 560 (2019)
(“Humans are often at a severe epistemic disadvantage vis-à-vis the systems they are tasked
with supervising.”).
17
Some decisions might call for more accuracy, but others less so. For landing
a plane, we want high accuracy, but for decisions about school admissions,
credit scoring, or criminal sentencing, other values are also quite important.
There is no one-size-fits-all approach to regulating AI, as the decisions it will
be employed to help make are quite different and demand different
considerations.
47 KAREN LEVY, DATA DRIVEN: TRUCKERS, TECHNOLOGY, AND THE NEW WORKPLACE SURVEILLANCE
148 (2023).
48 As Mireille Hildebrandt aptly notes, “profilers are not very interested in causes or reasons,
their interest lies in a reliable prediction, to allow adequate decision making.” Mireille
Hildebrandt, Defining Profiling: A New Type of Knowledge?, in PROFILING THE EUROPEAN
CITIZEN: CROSS-DISCIPLINARY PERSPECTIVES 17, 18 (Mirielle Hildebrandt & Seth Gutwirth eds.
2008).
18
eventually pass. It can also overlook students who work in public interest
jobs. It can dismiss students who could succeed at the bar if given the right
training or opportunities. It might fail to consider students who use the
degree to go into a non-legal career (politics, writing, journalism) or who go
into a low-paying public interest position.
Ultimately, the goal should be good decisions, and such decisions are quite
varied and contextual. As a general matter, the law does a very poor job in
guiding good decision-making, whether it be by human or machine. In the
end, there is no substitute for good judgment. Awful humans produce awful
results, whether alone or combined with a machine.
When we focus on how awful human decision-making can be, it’s tempting
to rush to embrace the machines. But we might be shortchanging the good
human elements in decisions. AI decides differently than humans. Machines
often standardize decisions, which brings about good as well as bad
consequences. Algorithms might bring greater uniformity but might also
lead to rigidity when flexible decisions might be preferable. There are
certainly times when we want a standardized decision but there are other
times when we want a Solomonic one.
In the end, good decisions depend upon good humans. We need much more
thought and study about when and how machine decision-making can
replace, help, or be mixed with human decision-making. Ultimately, the
success of the integration of AI into important decisions about people will
depend quite heavily on humans and how they think about algorithmic
decisions and their own decisions.
19