DATE DOWNLOADED: Mon Mar 18 15:26:26 2024

SOURCE: Content Downloaded from HeinOnline

Citations:
Please note: citations are provided as a general guideline. Users should consult their preferred
citation format's style manual for proper citation formatting.

Bluebook 21st ed.

Hubert L. Dreyfus, Artificial Intelligence, 412 Annals Am. Acad. Pol. & Soc. Sci. 21
(1974).

ALWD 7th ed.

Hubert L. Dreyfus, Artificial Intelligence, 412 Annals Am. Acad. Pol. & Soc. Sci. 21
(1974).

APA 7th ed.

Dreyfus, H. L. (1974). Artificial Intelligence. Annals of the American Academy of
Political and Social Science, 412, 21-33.

Chicago 17th ed.

Hubert L. Dreyfus, "Artificial Intelligence," Annals of the American Academy of
Political and Social Science 412 (1974): 21-33

McGill Guide 9th ed.

Hubert L. Dreyfus, "Artificial Intelligence" [1974] 412 Annals Am Acad Pol & Soc Sci
21.

AGLC 4th ed.

Hubert L. Dreyfus, 'Artificial Intelligence' [1974] 412 Annals of the American
Academy of Political and Social Science 21

MLA 9th ed.

Dreyfus, Hubert L. "Artificial Intelligence." Annals of the American Academy of
Political and Social Science, 412, 1974, pp. 21-33. HeinOnline.

OSCOLA 4th ed.

Hubert L. Dreyfus, 'Artificial Intelligence' (1974) 412 Annals Am Acad Pol & Soc Sci
21 Please note: citations are provided as a general guideline. Users
should consult their preferred citation format's style manual for proper citation
formatting.

-- Your use of this HeinOnline PDF indicates your acceptance of HeinOnline's Terms and
Conditions of the license agreement available at
https://heinonline.org/HOL/License
-- The search text of this PDF is generated from uncorrected OCR text.
-- To obtain permission to use this article beyond the scope of your license, please use:
Copyright Information
 Artificial Intelligence

By HUBERT L. DREYFUS

ABSTRACT: The belief in the possibility of artificial

intelligence (Al), given present computers, is the belief that
all that is essential to human intelligence can be formalized.
AI has not fulfilled early expectations in pattern recognition
and problem solving. These tasks cannot be formalized.
They necessarily involve a nonformal form of information
processing which is possible only for embodied beings
-where being embodied does not merely mean being able to
move and to operate manipulators. The human world, with its
recognizable objects, is organized by human beings using
their embodied capacities to satisfy their embodied needs.
There is no reason to suppose that a world organized in terms
of the body should be accessible by other means.

HubertL. Dreyfus, a graduateof HarvardUniversity, is Professorof Philosophyat

the University of California,Berkeley and has also taught at Harvard,Brandeis and
the Massachusetts Institute of Technology. He pursued the question of whether or
not digital computers could be programmed to behave intelligently, first as a con-
sultant at RAND and then, with NationalScience Foundationsupport,as a Research
Associate in Computer Sciences at the Harvard Computation Laboratory. He is
author of What Computers Can't Do: A Critique of Artificial Reason (1972) and of
numerous papers on artificialintelligence.

This article largely follows "Why Computers Must Have Bodies in Order To Be Intelligent"
(Review of Metaphysics, 40, no. 1 [September 1967]) and "Pseudo-Strides towards Artificial
Intelligence" (Theoria to Theory 2 [January 1968]). Some examples have been incorporated
from "Phenomenology and Artificial Intelligence" (in Phenomenology in America, ed. James
M. Edie [Chicago: Quadrangle Books, 1967]). Copyright* of all these articles is with Hubert L.
Dreyfus. A more detailed analysis of the problems discussed in this article, as well as ofthose in
the fields of game playing and language translation, can be found in the author's paperAlchemy
andArtificial Intelligence (RAND paper, p. 3244) and his book, What Computers Can't Do: A
Critique of Artificial Reason (New York: Harper and Row, 1972).
21
22 THE ANNALS OF THE AMERICAN ACADEMY

I TIS fitting to begin with a state-

ment made in 1957 by H. A.
programme at present, however, re-
mains a class C amateur.
Simon, one of the originators of the Similarly unfulfilled predictions
field of artificial intelligence: have been made in the areas of
pattern recognition and problem
It is not my aim to surprise or shock
you. . . . But the simplest way I can solving. However, philosophers
summarize is to say that there are now in have interests other than being the
the world machines that think, that learn conscience of a technical field which
and that create. Moreover, their ability to has been lax in critically evaluating
do these things is going to increase its failures. What should interest us
rapidly until-in a visible future-the is the philosophical significance of
range of problems they can handle will these unexpected difficulties: what
be co-extensive with the range to which underlying philosophical assump-
the human mind has been applied.1 tions lead workers in artificial intel-
The speaker predicts that within ten ligence (AI) to interpret their appar-
years a digital computer will be the ent failures as only temporary set-
world's chess champion and that backs and their modest success as
within ten years a digital computer justifying unbounded optimism?
will discover and prove an important Can these assumptions be justified?
mathematical theorem. If not, the stagnation of work in AI
We do not have time to go into the would cease to be surprising and,
deliberate confusions surrounding moreover, would give us new
the supposed proof of an important reasons to question the validity of
theorem. Suffice it to say that to date the assumptions on which such work
no important, or even original, is based.
theorem has been proved. The All Al work is done on digital
chess-playing story is also disap- computers because they are the only
pointing and typical: continued fail- all-purpose information processing
ure has been followed by optimistic devices which we know how to
predictions. 2 The best computer design, or even to conceive, at
present. All information with which
these computers operate must be
1. H. A. Simon and Allen Newell, "Heuris- represented in terms of binary
tic Problem Solving: The Next Advance in
Operation Research," Operations Research 6
digits-that is, in terms of a series of
(January-February 1958), p. 7-8. yes's and no's, of switches being
2. Allen Newell, J. C. Shaw and H. A. open or closed. The machine must
Simon, "Chess-Playing Programs and the operate on finite strings of these
Problem of Complexity," in Computers and determinate elements as a series of
Thought, ed. Edward A. Feigenbaum and
Julian Feldman (New York: McGraw-Hill, objects related to each other only by
1963); Allen Newell, J. C. Shaw and H. A. rules. Thus, psychologically, the
Simon, The Processes of Creative Thinking, computer is a model of the mind as
(RAND Corporation Paper, 1958), p. 1320. conceived of by associationists-for
Norbert Wiener, "The Brain and the Ma-
the elements-and intellectualists
chine," in Dimensions of Mind, ed. Sidney
Hook (New York: Crowell-Collier, 1960); -for the rules. Both associationists
Michael Scriven, "The Compleat Robot: A and intellectualists share the tradi-
Prolegomena to Androidology," in Dimen- tional conception of thinking as data
sions of Mind; H. A. Simon and Peter A. processing-a third-person process
Simon, "Trial and Error Search in Solving
Difficult Problems: Evidence from the Game in which the involvement of the
of Chess," Behavioral Science 7 (October processor plays no essential part.
1962). Moreover, since all information fed
 ARTIFICIAL INTELLIGENCE 23

into such machines must be in terms irony. Computer technology has

of bits, the belief that such machines been most successful in simulating
can be made to behave intelligently the so-called higher rational
presupposes that all relevant infor- functions-those which were once
mation about the world must be ex- supposed to be uniquely human.
pressible in an isolable, determinate Computers can deal brilliantly with
way. ideal languages and abstract logical
Thus, given digital computers, relations; for example, Wang's pro-
workers in Al are necessarily com- gramme has proved two hundred
mitted to two basic assumptions: (1) theorems from Principia Mathe-
an epistemological assumption that matica in less than three minutes.
all intelligent behavior can be It turns out that it is the sort of
simulated by a device whose only intelligence which we share with
mode of information processing is animals, such as pattern recognition,
that of a detached, disembodied, that has resisted machine simula-
objective observer; (2) the ontologi- tion.
cal assumption, related to logical Simon, who was only slightly
atomism, that everything essential to daunted by the failures of the first
intelligent behavior can in princi- decade of AL, still felt that "machines
ple be understood in terms of a will be capable, within twenty years,
determinate set of independent of doing any work that a man can
elements. do, " although he admits: "Automa-
In brief, the belief in the pos- tion of a flexible central nervous sys-
sibility of Al, given present com- tem will be feasible long before
puters, is the belief that all that is automation of a comparatively flex-
essential to human intelligence can ible sensory, manipulative, or loco-
be formalized. This formalist aim has motive system."4 However, what if
dominated philosophy since Plato, the work of the central nervous sys-
who set the goal by limiting the real tem, or what if the higher, determi-
to the intelligible and the intelligi- nate, logical and detached forms of
ble to that which could be made fully intelligence are necessarily derived
explicit so as to be grasped by any from, and guided by, global and
rational being. Leibniz pushed this involved lower forms? Then Simon's
position one step further by conceiv- optimism, as well as the two assump-
ing of a universal logical language tions underlying AI and traditional
capable of expressing everything in philosophy, would be unjustified. It
explicit terms which would permit is this existentialist thesis which I
thinking to achieve its goal of shall attempt to explain and defend.
becoming pure manipulation of this I shall consider two areas in which
formalism. Digital computers and work in AI has not fulfilled early
information theory have given us the expectations: pattern recognition
hardware and the conceptual tools to and problem solving. In each, I will
implement Leibniz's vision. We are try to account for the failure by
now witnessing the last act wherein arguing that the task in question
this conception of man as essentially cannot be formalized and by isolat-
rational-and rationality as essen- ing the nonformal form of infor-
tially calculation-will either tri-
umph or else reveal its inherent 3. Herbert Simon, The Shape of Automa-
tion for Men and Management (New York:
inadequacies. Harper and Row, 1965), p. 96.
It has already produced a certain 4. Ibid., p. 40.
24 THE ANNALS OF THE AMERICAN ACADEMY

mation processing necessarily in- ceed in some other way; indeed,

volved. Finally, I will try to show phenomenologists such as Aron Gur-
that the nonformalizable form of witsch, as well as gestalt psychol-
information processing in question ogists, have pointed out that our
is possible only for embodied recognition of ordinary spatial or
beings-where being embodied temporal objects does not seem to
does not merely mean being able to operate by checking off a list of
move and to operate manipulators. isolable, neutral, specific traits. For
example, in recognizing a melody
PATTERN RECOGNITION the notes get the values they have by
being recognized as part of the
Work in pattern recognition is melody rather than by the melody's
characteristic of work in all areas of being built up out of independently
Al. Some striking successes have recognized notes. Likewise, in the
been achieved; yet, they are based preception of objects there are no
on techniques which for practical neutral traits. The same hazy layer
reasons do not seem to be generaliz- which I would see as dust if I thought
able, and the important problems for I was confronting a wax apple might
pattern recognition, such as the rec- appear as moisture if I thought I
ognition of everyday objects or was seeing a fresh apple. The
speech, have so far proved intract- significance of the details and, in-
able. deed, their very look is determined
There are pattern recognition by my perception of the whole.
programmes now in operation which The recognition of spoken lan-
can recognize letters and numbers guage offers the most striking dem-
printed in various type fonts, and onstration of this global character
programmes which can be taught of our experience. From time to time
to recognize the handwriting of brash predictions have been made
specific persons. These all operate about mechanical secretaries into
by searching for certain topological which-or at whom-one could
features of the characters to be speak and whose programmes would
recognized and checking these fea- analyze the sounds into words and
tures against preset or learned type out the results. In fact, no one
definitions of each letter in terms of knows how to begin to make such a
these traits. The trick is to find rele- versatile device. Current work has
vant features-that is, those which shown that the same physical con-
remain generally invariant through- stellation of sound waves is heard as
out variations of size and orientation quite different phonemes depend-
and other distortions. This approach ing on the expected meaning. As
has been surprisingly successful Anthony Oettinger, of the Harvard
where recognition depends on a Computation Laboratory, has put it
small number of specific traits. in a paper published by Bell
However, the number of traits which Laboratories 5 :
can be looked up in a reasonable The essentially discrete and invariant
amount of time is limited, and nature of the phoneme, so evident to the
present programmes have already linguist concerned with the phonemic
reached this technological limit.
The restricted applicability of 5. Anthony Oettinger, Human Communi-
such programmes suggests that cation: A Unified View (New York:
human pattern recognition may pro- McGraw-Hill, 1973).
 ARTIFICIAL INTELLIGENCE 25

analysis . . . has turned out to be most spit it out. Or, if the right noema is
unexpectedly elusive in the absence of a found fast enough, one may recover
human agent. in time to recognize-that is, to
This leads Oettinger to the conclu- organize-the milk for what it is. Its
sion: other characteristics-whether it is
fresh or sour, buttermilk or skimmed
Perhaps . . . in perception as well milk-will then fall into place.
as in conscious scholarly analysis, the One might well wonder how it is
phoneme comes after the fact, namely,
possible to avoid looking for some
s . . it is constructed, if at all, as a con-
sequence of perception not as a step neutral features to begin this process
in the process of perception itself. of recognition. In fact, such a
description may seem so paradoxical
This would mean that the total as to make us try to explain the
meaning of a sentence or a melody phenomenon away. However, we
or a perceptual object determines must bear in mind that each meaning
the value to be assigned to the is given in a context which is already
elements. organized and on the basis of which
Oettinger goes on reluctantly to we have certain expectations. It is
suggest these conclusions: also important that we sometimes do
It may well be that an understanding of give the wrong meaning. In these
the meaning of a sentence is a precondi- cases the data coming in makes no
tion for . . . the analysis of the sen- sense at all, and we have to try a new
tence into phonemic components. The total hypothesis.
possibility is a frightening one to It is hard to imagine how a
face. . . . Yet the school boy asked to computer, which must operate on
parse a sentence proceeds neither like a completely determinate data accord-
machine nor like a generative grammar, ing to strictly defined rules, could be
at least there is no evidence that he does. programmed to use an underdeter-
On the contrary, the scant evidence there
is, suggests that he works backwards,
mined expectation of the whole in
going from meaning to structure. order to determine the elements of
that whole. Workers in AI might
The phenomenologist Edmund answer: even though people do
Husserl argued that, in recognizing use some sort of holistic approach
an object, we give a global mean- based on context which no one now
ing-a noema-to an otherwise in- knows how to program, there is no
determinant but determinable sen- reason, in principle, why some
suous matter. We then proceed to alternative approach could not be
make this open global meaning more discovered which would do the same
determinate by exploring what job. One could, for example, deal
Husserl called its inner horizon. more efficiently with a large number
This process can best be noticed of specific traits, or one could
when it is breaking down. If one develop a sort of anticipation which,
reaches for a glass of water and gets on the basis of certain traits in the
milk by mistake, on taking a sip the context, would assign an object to a
first reaction is total disorientation. class defined in terms of a large
One does not taste water, but one number of traits which would then
does not taste milk, either. One has serve as hypotheses.
a mouthful of what Husserl would This answer, however, ignores a
call pure sensuous matter-hyletic unique feature of human pattern
data-and, naturally, one wants to recognition: our ability to recognize
26 THE ANNALS OF THE AMERICAN ACADEMY
family resemblances where, as
Wittgenstein points out, two in-
dividuals recognized as belong-
ing to the same family need have
no exactly similar traits in common.
We can nonetheless recognize such
similarities by picking out a typical
case and introducing intermediate
cases. This use of paradigms and
context, rather than class definitions,
allows our recognition of patterns to
be open-textured in a way which is
impossible for recognition based on
a specific list of traits. Oettinger is
justified in concluding his paper
on a pessimistic note: "If indeed
we have an ability to use a global
context without recourse to for-
malization . . . then our optimistic
discrete enumerative approach is
doomed. . . ."
How, then, do human beings
operate with wholes, the elements
of which cannot exhaustively be
specified? Husserl has no answer
beyond the assertion that we do: that
transcendental consciousness has
the wunderbar capacity for giving
meanings and, thus, making possible
the perception, recognition and
exploration of enduring objects.
There is no way to criticize this view
except to say that it is frustrating;
it states a problem without propos-
ing any solution. For further help
we must turn to the existential
phenomenologists and, in particular,
to Merleau-Ponty who postulates
that it is the body which confers the
meanings discovered by Husserl.
Being prepared to feel silk, for
example, is to move or be prepared to
move our hand in a certain way and
to have certain expectations. As in
the case of the milk, if we have the
wrong expectations we experience
only confused sensations. p. 82.
7. Maurice Merleau-Ponty, Phenomenol-
It is easiest to become aware of ogy of Perception (London: Routledge and
the role of the body in taste and Kegan Paul, 1962), p. 142.
touch, but seeing, too, is a skill
which has to be learned. Focus-
ing, getting the right perspective
and picking out certain details,
all involve coordinated actions and
anticipations. As Piaget remarks,
"Perceptual constancy seems to be
the product of genuine actions,
which consist of actual or potential
movements of the glance or of the
organs concerned. . . ."6
Moreover, as Merleau-Ponty has
pointed out, the body is able to
respond as a whole to its environ-
ment. When the percipient acquires
a skill, he:
. . . does not weld together individual
movements and individual stimuli but
acquires the power to respond with a
certain type of solution to situations of a
certain general form. The situations may
differ widely from place to place, and the
response movements may be entrusted
sometimes to one operative organ,
sometimes to another, both situations
and responses in the various cases
having in common not so much a par-
tial identity of elements as shared
7
significance.
Thus, an anticipation of an object
does not arouse a single response or
specific set of responses but a
flexible skill which can be brought to
bear in an indefinite number of
ways. I can feel silk with either
hand or even with my feet. As
already noted, these anticipations
need not be completely specific,
but can become more specific in
the course of examining the object.
Thus, we give a global meaning to
our perceptual experience by bring-
ing to it a set of interdependent and
6. Compare, J. Piaget, Psychology of
Intelligence (New York: Humanities, 1966),
 ARTIFICIAL INTELLIGENCE 27

underdetermined skills which ex- recognition which would seem to be

perience gradually fills in and makes impossible for a system with a finite
more determinate. set of states.
A human perceiver, as does a
machine, needs feedback to find out PROBLEM SOLVING
if he has successfully recognized an Again, in problem solving there
object; however, there is an impor- was early success. In 1957 Newell,
tant difference in the feedback Shaw and Simon's logic theorist,
involved. A machine can, at best, using heuristically guided trial-
make a specific set of hypotheses and
and-error, proved thirty-eight out of
then find out if they have been
fifty-two theorems from Principia
confirmed or refuted by the data. The
Mathematica. Two years later the
body allows a much more flexible pretentiously named general prob-
criterion of what fulfills its expecta- lem solver (GPS), using more
tions. It need not check for specific sophisticated means-end analysis,
characteristics or a specific range solved the "cannibal and mission-
of characteristics, but simply for ary" problem and other problems of
whether, on the basis of its expecta- similar complexity.
tions, it is coping with the object. In 1961, after comparing a
Furthermore, coping need not be machine trace with a protocol that
defined by any specific set of traits,
matched the machine output to some
but rather by an ongoing mastery
extent, Newell and Simon jubilantly
which Merleau-Ponty calls max-
announced:
imum prise. Thus, whereas present
programmes call for a machine to Subsequent work has tended to confirm
recognize an object in order to [our] initial hunch, and to demonstrate
manipulate it, a human being can that heuristics, or rules of thumb, form
manipulate an object in order to the integral core of human problem-
solving processes. As we begin to
recognize it. understand the nature of the heuristics
To conclude: pattern recognition that people use in thinking, the mystery
is relatively easy for digital comput- begins to disolve from such (heretofore)
ers if there are a few specific traits vaguely understood processes as "intui-
which define the pattern, but com- tion" and "judgment."8
plex pattern recognition has proved
intractable using these methods. Yet, as we have seen in the case of
Transcendental phenomenologists pattern recognition, difficulties have
have pointed out that human beings an annoying way of reasserting
recognize complex patterns by pro- themselves. This time, the mystery
jecting a somewhat indeterminate of judgment reappears in terms of
whole which is progressively filled the organizational aspects of the
in by anticipated experiences. Exis- problem-solving programmes. In
tential phenomenologists have re- Some Problems of Basic Organiza-
lated this ability to our active, tion in Problem-Solving Pro-
organically interconnected body, set grammes Newell discusses some of
to respond to its environment in the problems which arise in organiz-
terms of a continual sense of its own ing the chess programme, the logic
functioning. This embodied sort of 8. H. A. Simon, Modelling Human Metal
information processing makes pos- Processes (RAND Corporation Paper, P-2221,
sible the open texture of pattern 20 February 1961), p. 12.
 28 THE ANNALS OF THE AMERICAN ACADEMY

theorist and especially the GPS we shall call "essential" and "inessen-
with a candour rare in the field and tial" operators, respectively. Essential
admits: "Most of them are unsolved operators are those which, when applied
to some extent, either completely, or to an expression, make "large" changes
because the solutions that have been in its appearance-change "P v P" to
adopted are still unsatisfactory in "P", for example. Inessential operators
are those which make "small" changes
one way or another." 9 No further
-e.g., change "P v Q" to "Q v P."
progress in solving these basic As we have said, the distinction is purely
problems has been reported. pragmatic. Of the twelve operators in
What is lacking is a way of organiz- this calculus, we have classified eight as
ing the problem so that one can see essential and four as inessential. 0
which operations are significant and
which, trivial. Wertheimer, in his No comment is necessary. We
classic work Productive Thinking, need only note that the classification
points out that the associationist of operators into essential and
account of problem solving excludes inessential-the function Werth-
the- most important aspect of eimer calls finding the deeper struc-
problem-solving behavior: a grasp ture, or insight-is not part of the
of the essential structure of the programme. It is introduced by the
problem, which he calls insight. In programmers before the so-called
this operation one breaks away from planning programme begins.
the surface structure and sees the No one has even tried to suggest
basic problem-what Wertheimer how a machine could perform this
calls the deeper structure-which structuring operation or how it could
enables one to recognize the steps be learned; in fact, it is itself one of
necessary for a solution. the conditions for learning from past
This gestaltist conception may experience. The ability to distin-
seem antithetical to the operational guish the essential from the inessen-
concepts demanded in artificial in- tial seems to be a uniquely human
telligence, but in fact this restructur- form of information processing, one
ing is surreptitiously presupposed not amenable to the mechanical
by the work of Newell, Shaw and search techniques which may oper-
Simon. In The Processes of Creative ate once this distinction has been
Thinking they introduce "the heuris- made. It is precisely this function of
tics of planning" to account for intelligence which resists further
characteristics of the subject's pro- progress in the problem-solving
tocol which are lacking in a simple field.
means-end analysis. The difficulty becomes even more
acute if one wishes to deal with
We have devised a programme . . . to
describe the way some of our subjects everyday problems rather than for-
handle logic problems, and perhaps the mal ones. With formal problems,
easiest way to show what is involved in planning is a matter of practical
planning is to describe the programme. necessity; in the case of ill-defined
On a purely pragmatic basis, the twelve problems, it is necessary in princi-
operators that are admitted in this system ple. Since there is no limit to the
of logic can be put in two classes, which amount of data which may be
9. Allen Newell, Some Problems of Basic 10. Allen Newell, J. C. Shaw and H. A.
Organization in Problem Solving Programs Simon, The Processes of Creative Thinking
(RAND Corporation Paper, RM 3283-PR, (RAND Corporation Paper, P-1320, 16 Sep-
December 1962), p. 4. tember 1958), pp. 43-44.
 ARTIFICIAL INTELLIGENCE
relevant for solving an ill-defined
problem, one cannot even in princi-
ple try all the permutations of
possibly relevant data in seeking a
solution. In such cases one must not
only determine which operations
bring about essential transforma-
tions, but which facts from the total
context are relevant.
. Even in a nonformal game, such
as playing the horses-which is
still much more systematic than
the everyday ill-structured prob-
lems which Simon once predicted
machines would be able to handle
-an unlimited set of conditions
becomes relevant. In placing a bet
we can usually restrict ourselves to
facts about the horse's age, jockey
and past performance. Perhaps, if
restricted to these facts, the machine
could do fairly well, possibly better
than an average handicapper; how-
ever, there are always other factors,
such as whether the horse is allergic
to goldenrod or whether the jockey
has just had a fight with the owner,
which may be decisive in some
cases. If the machine were to
examine explicitly each possibly
relevant factor as a determinate bit of
information in order to determine
whether to consider or ignore it, it
could never complete the calcula-
tions necessary to predict the out-
come of a single race. On the other
hand, if the machine systematically
excluded possibly relevant factors in
order to complete its calculations,
then it would sometimes be incapa-
ble of performing as well as an
intelligent human being.
The difficulties of simple means-
end analysis suggest that, in order for
the machine to structure its own data
in terms of significance and rele-
vance, it is not sufficient for it to have
an objective goal and to measure its
progress towards this preset end. We
have seen that pragmatic considera-
29
tions were necessary to structure the
logic programme before it could
begin and that the end must not only
serve as a test of progress, but must
modify our evaluation of the steps
which lead to it." Interest in the goal
is present at each moment and
structures the whole of experience
so that each detail is seen as relevant
or irrelevant to that end.
Moreover, an important feature of
pragmatic problem solving is com-
pletely neglected by workers in Al:
in creative problem solving we do
not know what our goal is until we
have achieved it. We do not have
a list of determinate objective
specifications which the solution
must fulfill. To understand these
features of problem solving we
require a concrete phenomenologi-
cal analysis of needs. In his thesis
Samuel Todes has provided just such
an analysis.' 2 According to Todes,
our bodily needs give us our sense of
the task at hand, in terms of which
our experience is structured as sig-
nificant or insignificant. These
needs, moreover, have a very special
structure. When we experience a
need we do not at first know what it
is we want. We must search to dis-
cover what allays our restlessness or
discomfort. This is not found by
comparing various objects and ac-
tivities with some objective, deter-
minate criterion, but through what
Todes calls our sense of gratification.
This gratification is experienced as
the discovery of that which we have
11. Samuel's checker programme does this,
but only by examining all permutations of
evaluations of moves and checking these
evaluations against the outcome of play in
each case. This is a successful ad hoc strategy,
but it is only possible where there is a small
and clearly defined set of evaluations and of
alternatives to be evaluated.
12. Samuel Todes, "The Human Body as
the Material Subject of the World" (Ph.D.
diss., Harvard University, 1963).
30 THE ANNALS OF THE AMERICAN ACADEMY

needed all along, but it is a seems to be based on bodily needs,

retroactive understanding and cov- the question of whether artificial
ers up the fact that we were unable to intelligence is possible boils down
make our need determinate without to the question of whether there can
first receiving the gratification. The be an artificial embodied agent. I
original fulfillment of any need is, will take up this question in the
therefore, what Todes calls a crea- remainder of this paper. Remember,
tive discovery. Only such an analysis however, that the question is
of human needs can both account for philosophically interesting only if
our ability to order our experience in we restrict ourselves to asking if one
terms of relevance and significance can make such a robot by using a
and at the same time allow determi- digital computer and mechanical
nation of the goal ofcreative problem hardware. I assume that, in princi-
solving to remain part of the ple, one could construct an artificial
problem-solving task. embodied agent if one used compo-
To summarize: the work in nents sufficiently similar to those
problem-solving programmes has which make up a human being.
shown that we must structure our A project to build such a digitally
problems into essential-necessary, controlled robot was undertaken at
indispensable, most needed-oper- the Massachusetts Institute of Tech-
ations and accidental operations nology; it is philosophically interest-
and select the relevant data. Only ing to consider its progress and its
an analysis of the gestalting process underlying assumptions. The project
by which human interests structure director, Marvin Minsky, modestly
human experience can account for tried to make only a mechanical
these abilities. The crucial role of shoulder, arm and hand, coordin-
interests in determining relevance ated with a TV eye; his none-too-
and significance has been neglected modest ambition was to make it
by Al workers who have unknow- use tools to construct things. The
ingly smuggled in their own in- first simple task was to programme
terests to do the job. the robot arm to pick up blocks.
Interests and goals cannot be This was indeed accomplished and
simulated on a digital machine represents the typical early success
whose only mode of existence is a one has learned to expect in the field.
series of determinate states and The problem which remained was,
which has, at best, specific targets as usual, that of generalizing the
rather than needs. Without this present, successful techniques. To
human form of information proces- bring a simple arm over to pick up a
sing, no digital computer can cope block requires locating the block in
with the indefinite number of pos- objective space, locating the arm in
sibly relevant facts in the every- the same space and then bringing
day world nor solve ill-defined the two together. This is already
problems. quite a feat. It took the machine
minutes just to pick up a block. A
more flexible arm endowed with
MECHANICAL BODIES
more degrees of freedom will in-
Since it turns out that pattern volve calculations requiring even
recognition is a bodily skill and since longer computations. If one adds
accident-essence discrimination to this the fact that in the case
 ARTIFICIAL INTELLIGENCE
of any skill which takes place in real
time, such as playing ping pong,
these calculations must be com-
pleted before the ball arrives, the
outlook is not very promising.
In the light of these difficulties,
what encourages researchers to de-
vote their research facilities to such
a project? They simply have the
conviction that since we are, as
Minsky puts it, meat machines and
are able to play ping pong, there
is no reason-in principle or in
practice-that a metal machine
should not do so, also. Before
jumping to such a conclusion, par-
ticularly when time and money are at
stake, the robot makers ought first to
examine their underlying assump-
tion that no essential difference
exists between meat machines and
metal machines, between being
embodied and controlling movable
manipulators. How do human beings
play ping pong or, to make the matter
simpler, how do human beings use
tools?
Heidegger, Merleau-Ponty and
Michael Polanyi have devoted a
great deal of thought to this question.
They discuss the important way in
which our experience of a tool we are
using differs from our experience of
an object. A blind man who runs his
hand along a stick he uses to grope
his way will be aware of its objective
characteristics. When he is using it,
however, he is not aware of its
objective traits nor of the pressure in
the palm of his hand. Rather, the
stick has become, like his body, a
transparent access to objects.
Polanyi writes:
While we rely on a tool or a probe,
these are not handled as external
objects . . . they remain on our
side . . . forming part of ourselves, the
operating persons. We pour ourselves
3I
out into them and assimilate them as
parts of our existence. We accept them
existentially by dwelling in them.13
In this way we are able to bring the
probe into contact with an object in
physical space without needing to be
aware of the physical location of the
probe. Merleau-Ponty notes that:
The whole operation takes place in the
domain of the phenomenal; it does not
run through the objective world, and
only the spectator, who lends his
objective representation to the living
body of the active subject, can believe
that . . . the hand moves in objective
space. 14
As Merleau-Ponty admits, this abil-
ity seems magical from the point of
view of science and, rather than have
no explanation of what people are
able to do, the scientist quite
justifiably embraces the assumptic
that people are unconsciously run-
ning with incredible speed through
the enormous calculation which
would be involved in programming a
computer to perform a similar task.
However implausible, this view
gains persuasiveness from the ab-
sence of an alternative account.
To make embodiment an accepta-
ble alternative we will have to
explain how one could perform
physical tasks without in any way
appealing to the principles of
physics or geometry. Consider the
act of randomly waving my hand in
the air. I am not trying to place my
objective hand at an objective point
in space. To perform this waving I
need not take into account the
geometry, since I am not attempting
any specific achievement. Now sup-
pose that, in this random thrashing
13. Michael Polanyi, Personal Knowledge
(London: Routledge and Kegan Paul, 1958),
p. 59.
14. Merleau-Ponty, Perception, p. 106.
32 THE ANNALS OF THE AMERICAN ACADEMY

about, I happen to touch something For the AI researcher it seems that

and that this satisfies and thereby intelligent behavior can be pro-
makes partially determinate a need, duced only by running through the
a need to cope with things. I can then calculations necessary to describe
repeat whatever I did-this time in the objective performance. How-
order to touch something-without ever, as we have seen, being
appealing to the laws necessary to embodied creates a second possibil-
describe it as a physical motion. I ity: an active, involved agent can
now have a way of bringing two build up skills and assimilate in-
objects together in objective space struments as extensions of his body.
without appealing to any principle Thus, an embodied agent can dwell
except: do that again. This is the way in the world in such a way as to avoid
skills are built up. The important the infinite task of trying to formalize
thing about skills is that, although everything.
science requires that the skilled
performance be described according CONCLUSION
to rules, these rules need in no way
be involved in producing the per- The force of my argument, in so far
formance. as it is an impossibility argument,
Human beings are further capable depends on the open texture of
pattern recognition, the infinity of
of remembering, refining and reor-
facts which may be relevant in
ganizing these somewhat indeter-
problem solving and the correlative
minate motor schemata. Piaget has
amassed an enormous amount of flexibility of bodily skills. If experi-
ence really has this open character,
evidence tracing this development
and has come to a gestaltist con- then any specific, human, intelligent
clusion: performance could indeed be simu-
lated on a computer after the fact, but
The specific nature of operations . . fully intelligent behavior would be
.

depends . . . on the fact that they impossible, in principle, for a digital

never exist in a discontinuous state. machine. This does not mean that
. . . A single operation could not be some limited sort of artificial intel-
an operation because the peculiarity of
ligence is impossible or even im-
operations is that
they form systems.
Here we may well protest vigorously practical. It remains an open ques-
against logical atomism . . . a grievous tion to which extent human perform-
hindrance to the psychology ofthought.1 s ance can be simulated after the fact
by finding rules to describe that
15. Piaget, Intelligence, p. 35. These motor performance and then programming
schemata must have their muscular and neural them.
basis, but there is no reason to suppose that I have, however, shown several
these physical correlates are reducible to a
series of determinate states. Both the global serious difficulties in current AL
and undetermined character of the motor work which suggest that, whatever
schemata argue against this possibility. D. M. rule-like way of processing informa-
MacKay, the one theoretician of Al who tion may yet be found, present
claims that a robot's model of the external
world must be stored as motor schemata,
recognizes this point. He warns: "We on the M. MacKay, "A Mind's Eye View of the
circuit side had better be very cautious before Brain," in Progress in Brain Research:
we insist that the kind of information Cybernetics of the Nervous System 17
processing that a brain does can be replicated (Amsterdam, 1965), p. 16. So far, the only
in a realizable circuit. Some kind of 'wet' entity which can meet these specifications is
engineering may turn out to be inevitable"; D. the human body.
 ARTIFICIAL INTELLIGENCE 33

techniques are inadequate. To simu- niques will ever be powerful enough

to cope with the amount of data
late the gestalt character of pattern
recognition and problem solving on involved.' 6 In fact, it would be more
a digital machine would require, at reasonable to suppose they will
the very least, the storage and easynever be. The human world with its
accessing of vast amounts of data. recognizable objects is organized by
Only in this way could the computer human beings using their embodied
begin to simulate the use of past capacities to satisfy their embodied
experience and present context to needs. There is no reason to suppose
structure present experience-a that a world organized in terms of the
process which seems to be necessary body should be accessible by other
for any complex pattern recognition means.
or problem solving.
No such data-processing tech- 16. There is no reason to take any comfort
niques exist at present and, once the from the fact that human beings perform these
traditional philosophical assump- remarkable tasks, since the body is certainly
not a digital computer, and the latest work in
tions underlying work in Al have neurophysiology has produced convincing
been called into question, there is no evidence that the digital computer is not an
reason to suppose that such tech- adequate model of the brain.