The Soul Gained and Lost:
Artificial Intelligence as a Philosophical Project

Philip E. Agre
Department of Information Studies
University of California, Los Angeles
Los Angeles, California 90095-1520
USA

pagre@ucla.edu
http://polaris.gseis.ucla.edu/pagre/

This paper appeared in a slightly different form in a special issue of the Stanford Humanities Review, entitled "Constructions of the Mind: Artificial Intelligence and the Humanities", edited by Guven Guzeldere and Stefano Franchi. The official citation is Stanford Humanities Review 4(2), 1995, pages 1-19.

9400 words.

 

1 Introduction

When I was a graduate student in artificial intelligence, the humanities were not held in high regard. They were vague and woolly, they employed impenetrable jargons, and they engaged in "meta-level bickering that never decides anything". Although my teachers and fellow students were almost unanimous in their contempt for the social sciences, several of them (not all, but many) were moved to apoplexy by philosophy. Periodically they would convene impromptu Two-Minute Hate sessions to compare notes on the arrogance and futility of philosophy and its claims on the territory of AI research. "They've had two thousand years and look what they've accomplished. Now it's our turn." "Anything that you can't explain in five minutes probably isn't worth knowing." They distinguished between "just talking" and "doing", where "doing" meant proving mathematical theorems and writing computer programs. A new graduate student in our laboratory, hearing of my interest in philosophy, once sat me down and asked in all seriousness, "Is it true that you don't actually do anything, that you just say how things are?" It was not, in fact, true, but I felt with great force the threat of ostracism implicit in the notion that I was "not doing any real work".

These anecdotes may provide some sense of the obstacles facing any attempt at collaboration between AI and the humanities. In particular, they illustrate certain aspects of AI's conception of itself as a discipline. According to this self-conception, AI is a self-contained technical field. In particular, it is a practical field; to do AI is to prove theorems, write software, and build hardware whose purpose is to "solve" previously defined technical "problems". The whole test of these activities lies in "what works". The criterion of "what works" is straightforward, clear, and objective in the manner of engineering design; arguments and criticisms from outside the field can make no claim at all against it. The substance of the field consists in the "state of the art" and its history is a history of computer programs. The technical methods underlying these programs might have originated in other fields, but the real work consisted in formalizing, elaborating, implementing, and testing those ideas. Fields which do not engage in these painstaking activities, it is said, are sterile debating societies which do not possess the intellectual tools -- most particularly mathematics -- to do more than gesture in the general direction of an idea, as opposed to really working it out.

I will be thought to exaggerate. Technicians will protest their respect for great literature and the attitudes I have reported will be put down to a minority of fundamentalists. Yet the historical record makes plain that interactions between AI and the humanities have been profoundly shaped by the disciplinary barriers that such attitudes both reflect and reproduce. Serious research in history and literature, for example, has had almost no influence on AI. This is not wholly due to ignorance on the part of technical people, many of whom have had genuine liberal educations. Rather AI, as a technical field, is constituted in such a way that its practitioners honestly cannot imagine what influence those fields could have.

Philosophy has had a little more influence. Research on AI's constitutive questions in the philosophy of mind is widely read and discussed among AI research people, and is sometimes included in the curriculum, but these discussions are rarely considered part of the work of AI, judging for example by journal citations, and any influence they might have had on the day-to-day work of AI has been subtle at best. Contemporary ideas from the philosophies of language and logic have been used as raw material for AI model-making, though, and philosophers and technical people have collaborated to some degree in specialized research on logic.

Perhaps the principal humanistic influence on AI has derived from a small number of philosophical critics of the field, most particularly Hubert Dreyfus (1972). For Dreyfus, the project of writing "intelligent" computer programs ran afoul of the critique of rules in Wittgenstein (1968 [1953]) and Heidegger's (1961 [1927]) analysis of the present-at-hand way of relating to beings (in this case, symbolic rules). The use of a rule in any practical activity, Dreyfus argues, requires a prior participation in the culturally specific form of life within which such activities take place. The attempt to fill in the missing "background knowledge" through additional rules would suffer the same problem and thus introduce a fatal regress [1]. Although most senior AI researchers of my acquaintance stoutly deny having been affected by Dreyfus' arguments, a reasonable amount of research has addressed the recurring difficulties with AI research that Dreyfus predicted. One of these is the "brittleness" of symbolic, rule-based AI systems that derives from their tendency to fail catastrophically in situations that depart even slightly from the whole background of operating assumptions that went into the system's design. For the most part, the response of AI researchers to these difficulties, and to Dreyfus' analyses generally, is to interpret them as additions to AI's agenda that require no fundamental rethinking of its premises [2].

Within the field itself, critical reflection is largely a prerogative of the field's most senior members, and even these papers are published separately from the narrowly technical reports, either in non-archival publications like the AI Magazine or in special issues of archival publications devoted to the founders' historical reflections. In 1990 I received a referee's report on a AAAI (American Association for Artificial Intelligence) conference paper that read in part,

In general, avoid writing these "grand old man" style papers until you've built a number of specific systems & have become a grand old man.
The boundaries of "real AI research", in short, have been policed with great determination.

Yet this is now changing. In part the current changes reflect sociological shifts in the field, in particular its decentralization away from a few heavily funded laboratories and the resulting, albeit modest, trend toward interdisciplinary pluralism. But the change in atmosphere has also been influenced by genuine dissatisfactions with the field's original technical ideas. AI's practices of formalizing and "working out" an idea constitute a powerful method of inquiry, but precisely for this reason they are also a powerful way to force an idea's internal tensions to the surface through prolonged technical frustrations: excessive complexity, intractable inefficiency, difficulties in "scaling up" to realistic problems, and so forth. These patterns of frustration have helped clear the ground for a new conception of technical work, one that recognizes the numerous, deep continuities between AI and the humanities. Although these continuities reach into the full range of humanistic inquiry, I will restrict myself here to the following five assertions about AI and its relationship to philosophy:

  1. AI ideas have their genealogical roots in philosophical ideas.
  2. AI research programs attempt to work out and develop the philosophical systems they inherit.
  3. AI research regularly encounters difficulties and impasses that derive from internal tensions in the underlying philosophical systems.
  4. These difficulties and impasses should be embraced as particularly informative clues about the nature and consequences of the philosophical tensions that generate them.
  5. Analysis of these clues must proceed outside the bounds of strictly technical research, but they can result in both new technical agendas and in revised understandings of technical research itself.
In short, AI is philosophy underneath. These propositions are not entirely original, of course, and some version of them underlies Dreyfus' early critique of the field. My own purpose here is to illustrate how they might be fashioned into a positive method of inquiry that maintains a dialogue between the philosophical and technical dimensions of AI research. To this end, I will present a brief case study of one idea's historical travels from philosophy through neurophysiology and into AI, up to 1972. Although much of this particular story has been told many times, some significant conclusions from it appear to have escaped analysis. It is an inherently difficult story to tell, since it requires a level of technical detail that may intimidate the uninitiated without nearly satisfying the demands of initiates. It is a story worth telling, though, and I will try to maintain a firm sense of the overall point throughout. I will conclude by briefly discussing recent developments that have been motivated in part by critical reevaluations of this tradition, and by sketching the shape of the new, more self-critical AI that is emerging in the wake of this experience.

2 Rene Descartes: Criteria of intelligence

In a famous passage in his Discourse on Method, Descartes summarizes a portion of his suppressed treatise on The World as follows:

... the body is regarded as a machine which, having been made by the hands of God, is incomparably better arranged, and possesses in itself movements which are much more admirable, than any of those which can be invented by man. ... if there had been such machines, possessing the organs and outward form of a monkey or some other animal without reason, we should not have had any means of ascertaining that they were not of the same nature as those animals. On the other hand, if there were machines which bore a resemblance to our body and imitated our actions as far as it was morally possible to do so, we should always have two very certain tests by which to recognize that, for all that, they were not real men. The first is, that they could never use speech or other signs as we do when placing our thoughts on record for the benefit of others. For we can easily understand a machine's being constituted so that it can utter words, and even emit some responses to action on it of a corporeal kind, which brings about a change in its organs; for instance, if it is touched in a particular part it may ask what we wish to say to it; if in another part it may exclaim that it is being hurt, and so on. But it never happens that it arranges its speech in various ways, in order to reply appropriately to everything that may be said in its presence, as even the lowest type of man can do. And the second difference is, that although machines can perform certain things as well as or perhaps better than any of us can do, they infallibly fail short in others, by the which means we may discover that they did not act from knowledge, but only from the disposition of their organs. For while reason is a universal instrument which can serve for all contingencies, these organs have need of some special adaptation for every particular action. From this it follows that it is morally impossible that there should be sufficient diversity in any machine to allow it to act in all the events of life in the same way as our reason causes us to act (page 116).
It is worth quoting Descartes' words at such length because they contained the seeds of a great deal of subsequent intellectual history. Distinctions between the body and the soul were, of course, of great antiquity, as was the idea that people could be distinguished from animals by their reasoned use of language. Descartes, though, extended these ideas with an extremely detailed physiology. His clearly drawn dualism held that automata, animals, and the human body could all be explained by the same mechanistic laws of physics, and he set about partitioning functions between body and mind [3] [4]. In establishing this partition, one of the tests was the conventional distinction between animal capabilities, which reside in the body, and specifically human capabilities, which required the exercise of the soul's faculties of reason and will. Thus, for example, automata or animals might utter isolated words or phrases in response to specific stimuli, but lacking the faculty of reason they could not combine these discrete units of language in an unbounded variety of situationally appropriate patterns. The soul itself has ideas, but it has no physical extent or structure. Thus, as Descartes explains in The Passions of the Soul (Articles 42 and 43), memory is a function of the brain; when the soul wishes to remember something, it causes animal spirits to propagate to the spot in the brain where the memory is stored, whereupon the original image is presented once again to the soul in the same manner as a visual perception.

The attraction of Descartes' proposals lay not in their particulars, many of which were dubious even to his contemporaries. Rather, Descartes provided a model for a kind of theory-making that contrasted with late scholastic philosophy in every way: it was specific and detailed, it was grounded in empirical physiology, and it was written in plain language.

3 Karl Lashley: Language as a model for action

The American cognitivists of the 1950's often modeled themselves after Descartes, and they intended their research to have much of the same appeal. Despite the intervening three centuries, the lines of descent are indeed clear. This is evident in the case of Chomsky, for example, who argued in explicitly Cartesian terms for a clear distinction between the physiology of speech, including the biological basis of linguistic competence, and the capacity for actually choosing what to say. While not a dualist, Chomsky nonetheless epitomized his conception of human nature in terms of "free creation within a system of rule" (1971: 50). Miller (1956) used Descartes' Rules for the Direction of the Mind to motivate his search for ways that people might more efficiently use their limited memories.

The first and most influential revival of research into mental mechanisms, Karl Lashley's 1951 paper "The problem of serial order in behavior", did not acknowledge any sources beyond the linguistics, psychology, and neurophysiology of the 1940's. Nonetheless, the underlying continuities are important for the computational ideas that followed. Despite his own complex relationship to behaviorism, Lashley's paper argued clearly that behaviorist psychology could not adequately explain the complexity of human behavior. Lashley focused on a particular category of behavior, namely speech. He pointed out that linguists could demonstrate patterns to the grammar and morphology of human languages that are hard to account for using the theory of "associative chains", whereby each action's effects in the world give rise to stimuli that then trigger the next action in turn. The formal structures exhibited by human language, then, were sufficient reason to restore some notion of mental processing to psychology.

Moreover, Lashley suggested that all action be understood on the model of language. He regarded both speech and physical movement as having a "syntax", and he sought the physiological basis of both the syntax of movement and the choice of specific movements from among the syntactically possible combinations. This suggestion was enormously consequential for the subsequent development of cognitivist psychology, and particularly for AI. Lashley summarized the idea in this way:

It is possible to designate, that is, to point to specific examples of, the phenomena of the syntax of movement that require explanation, although those phenomena cannot be clearly defined. A real definition would be a long step toward solution of the problem. There are at least three sets of events to be accounted for. First, the activation of the expressive elements (the individual words or adaptive acts) which do not contain the temporal relations. Second, the determining tendency, the set, or idea. This masquerades under many names in contemporary psychology, but is, in every case, an inference from the restriction of behavior within definite limits. Third, the syntax of the act, which can be described as an habitual order or mode of relating the expressive elements; a generalized pattern or schema of integration which may be imposed upon a wide range and a wide variety of specific acts. This is the essential problem of serial order; the existence of generalized schemata of action which determine the sequence of specific acts, acts which in themselves or in their associations seem to have no temporal valence (Lashley 1951: 122).
Two things are new to cognitive theorizing here, grammar as a principle of mental structure and the generalization of grammatical form to all action. But a great deal in Lashley's account is continuous with that of Descartes. To start with, it is an attempt at an architecture of cognition. Indeed, it is considerably less detailed than Descartes' architecture, although Descartes provided no account of the mechanics of speech. Both Lashley and Descartes assign the ability to speak individual words -- or in Lashley's case, to make individual discrete physical movements -- to individual bits of machinery, without being very specific about what these bits of machinery are like. And they both view the human capacity for putting these elements together as the signature of the mind. To be sure, Lashley's argument rests on the formal complexity of speech whereas Descartes points at the appropriateness of each utterance to the specific situation. In each case, though, what counts is the capacity of the mind to order the elements of language in an unbounded variety of ways.

The continuities go deeper. Lashley, as a neurophysiologist, shows no signs of believing in an ontological dualism such as Descartes'. Yet the conceptual relations among the various components of his theory are analogous to those of Descartes. In each case, the brain subserves a repertoire of bodily capacities, and on every occasion the mind orders these in accord with its choices, which themselves are not explained. For Descartes the mind's choices simply cannot be explained in causal terms, though its operations can be described in the normative terms of reason, as for example in his Rules for the Direction of the Mind. Lashley does not express any overt skepticism about his "determining tendency", but neither does he have anything very definite to say about it; the concept stays nebulous throughout. This is not simply an incompleteness of Lashley's paper but is inherent in its design: the purpose of the determining tendency is not to have structure in itself but to impose structure upon moment-to-moment activities from the repertoire of action schemata made available to it by the brain.

In retrospect, then, Lashley's paper makes clear the shape of the challenge that the cognitivists had set themselves. They wished to rout their sterile behaviorist foes in the same way that Descartes had routed the schoolmen, by providing a scientific account of cognitive processes. The problem, of course, is that Descartes was not a thoroughgoing mechanist. So long as the cognitivists retained the relational system of ideas that they had inherited from Descartes, and from the much larger tradition of which Descartes is a part, each of their models would include a component corresponding to the soul. No matter how it might be squeezed or divided or ignored, there would always remain one black box that seemed fully as intelligent as the person as a whole, capable of making intelligent choices from a given range of options on a regular basis. As the field of AI developed, this recalcitrant box acquired several names. Dennett (1978: 80-81), for example, spoke of the need for "discharging the homunculus", something he imagined to be possible by dividing the intelligent homunculus into successively less intelligent pieces, homunculi within homunculi like the layers of an onion, until one reached a homunculus sufficiently dumb to be implemented in a bit of computer code. AI researchers' jargon spoke of subproblems as being "AI-complete" (an analogy: so-called NP-complete computational problems are thought to be unsolvable except through an enumeration of possible solutions -- an efficient algorithm for any one such problem would yield efficient algorithms for all of them). And several exceedingly skilled programmers devised computer systems that were capable of reasoning about their own operation -- including reasoning about their reasoning about their own operation, and so on ad infinitum (e.g., Smith 1985). In each case, the strategy was reducing the soul's infinite choices to finite mechanical means.

But beyond sketching the shape of a future problem, Lashley also sketched the principal strategy of a whole generation for solving it. The operation of the determining tendency might be a mystery, but the general form of its accomplishment was not. While the linguistic metaphor for action envisions an infinite variety of possible actions, it also imposes a great deal of structure on them. In mathematical terms the possible actions form a "space". The generative principle of this space lies in the "schemata of action", which are modeled on grammatical rules. A simple schema for English sentences might be

Sentence -> NounPhrase IntransitiveVerb .
That is, roughly speaking "one way to make a sentence is to utter a noun phrase followed by an intransitive verb". Other rules might spell out these various "categories" further; for example,
NounPhrase -> Article Noun
Article -> a
Article -> the
Noun -> cat
Noun -> dog
IntransitiveVerb -> slept
IntransitiveVerb -> died
These mean, roughly, "one way to make a noun phrase is to utter an article followed by a noun", "some possible articles are "a" and "the" ", "some possible nouns are "cat" and "dog" ", and "some possible intransitive verbs are "slept" and "died" ". And there might be other ways to make sentences; for example,
Sentence -> NounPhrase TransitiveVerb NounPhrase
TransitiveVerb -> saw
TransitiveVerb -> ate
This particular set of grammatical rules generates a finite space of English sentences; for example,
the cat saw a dog
a dog ate a dog
The process of "deriving" a sentence with these rules is simple and orderly. One begins with the "category" Sentence, and then at each step one makes two choices: (1) which category to "expand", and (2) which rule to apply in doing so, until no categories are left. For example, one might proceed as follows:
1. Sentence
2. NounPhrase TransitiveVerb NounPhrase
3. NounPhrase TransitiveVerb Article Noun
4. NounPhrase saw Article Noun
5. Article Noun saw Article Noun
6. the Noun saw Article Noun
7. the Noun saw Article dog
8. the cat saw Article dog
9. the cat saw a dog
The space of possible sentences, then, resembles a branching road with a definite set of choices at each point. The process of choosing a sentence is reduced to a series of much smaller choices among a small array of alternatives. The virtue of this reduction becomes clearer once the grammar generates an infinite array of sentences, as becomes the case when the following grammatical rules are added to the ones above:
Sentence -> NounPhrase CognitiveVerb that Sentence
CognitiveVerb -> thought
CognitiveVerb -> forgot
It now becomes possible to generate sentences such as
the cat thought that the dog forgot that a cat slept
Chomsky (1965: 8) in particular made a great deal of this point; following Humboldt, he spoke of language as making "infinite use of finite means". And although he believes that the mind ultimately has a biological (and thus mechanical) explanation (1979: 66, 97), he has focused his research on the level of grammatical competence rather than trying to uncover this explanation himself.

4 Allen Newell and Herbert Simon: The mechanization of the soul

Instead, the first steps in mechanizing this idea of a generative space were due to Newell and Simon (1963). Whereas Chomsky was concerned simply with the precise extent of the generative space of English grammar, Newell and Simon's computer program had to make actual choices within a generative space. And whereas Lashley posited the existence of a "determining tendency" whose genealogical origins lay in a non-mechanical soul, Newell and Simon had to provide some mechanical specification of it. Here the generative structure of the space was crucial. Newell and Simon did not employ linguistic vocabulary. Nonetheless, just as grammatical rules and derivations provide a simple, clear means of generating any grammatical sentence, the application of "operators" provided Newell and Simon with a simple, easily mechanized means of generating any possible sequence of basic actions. Choosing which sequence of actions to adopt was a matter of "search". The mechanism that conducted the search did not have to make correct choices all the time; it simply had to make good enough choices eventually as it explored the space of possibilities.

Newell and Simon placed enormous significance on this idea (see, for example, Newell's comments in Agre (1993: 418)), and justifiably so. While maintaining the system of conceptual relations already found in Descartes, Lashley, and Chomsky, their program nonetheless embodied a serious proposal for the mechanization of the soul (cf Gallistel 1980: 6-7). Their strategy was ingenious: rather than endow the soul with an internal architecture -- something incomprehensible within the system of ideas they inherited -- they effectively proposed interpreting the soul as an epiphenomenon. Ironically, given Descartes' polemics against scholastic philosophy, the idea is approximately Aristotelian: the soul as the form of the person, not a discrete component. More specifically, rather than being identified with any particular device, the soul was contained by the generative structure of the search space and manifested through the operation of search mechanisms. These search mechanisms were "heuristic" in the sense that no single choice was ever guaranteed to be correct, yet the overall effect of sustained searching was the eventual discovery of a correct outcome. Despite the simplicity and limitations of their early programs, Newell and Simon were willing to refer to these programs' behavior as "intelligent" because they met this criterion. And they regarded their proposal as promising because so many human activities could readily be cast as search problems.

Up to this point, the story of the mechanization of the soul is a conventional chapter in the history of ideas: to tell this story, we trace the unfolding of an intellectual project within an invariant framework of continuities or analogies among idea-systems. With Newell and Simon's program, though, the story clearly changes its character. But how exactly? So far as the disciplinary culture of AI is concerned, the formalization and implementation of an idea bring a wholly new day -- a discontinuity between the prehistory of (mere) questions and ideas and the history, properly speaking, of problems and techniques. Once this proper history has begun, technical people can put their proposals to the test of implementation: either it works or it does not work.

Yet despite this conception, and indeed partly because of it, the development of technical methods can be seen to continue a along trajectory largely determined by the defining projects and internal tensions of the ancestral systems of ideas. In particular, these projects and tensions continue to manifest themselves in the goals and tribulations of AI's technical work. In the case of Newell and Simon's proposal, the central goals and tribulations clustered around the "problem" of search control -- that is, making heuristic search choices well enough -- not perfectly, just well enough -- to allow the search process to "terminate" with an acceptable answer within an acceptable amount of time. An enormous AI subliterature addresses this problem in a wide variety of ways. Within this literature, searches are said to "explode" because of the vastness of search spaces. It should be emphasized that mediocre search control ideas do not kill a mechanism; they only slow it down. Yet this research has long faced a troubling aporia: the more complicated the world is, the more choices become possible at each point in the search, and the more ingenuity is required to keep the search process under control. The metaphors speak of a struggle of containment between explosion and control. Such a struggle, indeed, seems inherent in any theory for which action is said to result from formal reason conducted by a finite being (Cherniak 1986).

Newell and Simon's achievement thus proved tenuous. So long as AI's self-conception as a self-sufficient technical discipline has remained intact, however, these difficulties are readily parsed as technical problems seeking technical solutions. An endless variety of solutions to the search control problem has indeed been proposed, and each of them more or less "works" within the bounds of one or another set of "assumptions" about the world of practical activity.

5 Richard Fikes and Nils Nilsson: Mechanizing embodied action

To watch the dynamics of this process unfold, it will help to consider one final chapter: the STRIPS program (Fikes and Nilsson 1971). The purpose of STRIPS is to automatically derive "plans" for a robot to follow in transporting objects around in a maze of rooms. The program constructs these plans through a search process modeled on those of Newell and Simon [5]. The search space consists of partially specified plans, with each "operator" adding another step to the plan. Returning to the linguistic metaphor, the authors of STRIPS understand the robot's action within a grammar of possible plans. They refer to the units of action that Lashley called "expressive elements" as "primitive actions", and the "syntax of the act" strings these actions into sequences that can be "executed" in the same manner as a computer program. In Descartes' terms, the soul's faculty of reason specifies an appropriate sequence of bodily actions, each of which may well be complicated, its faculty of will decides to undertake them, and the body then physically performs them.

To those who have had experience getting complex symbolic programs to work, the STRIPS papers make intense reading. Because the authors were drawing together so many software techniques for the first time, the technically empathetic reader gets a vivid sense of struggle -- the unfolding logic of what the authors unexpectedly felt compelled to do, given what seemed to be required to get the program to work. A detailed consideration of the issues would take us much too far afield, but the bottom line is easy enough to explain. As might be expected, this bottom line concerns the technical practicalities of search control. A great deal is at stake: if the search can be controlled without making absurdly unrealistic assumptions about the robot's world, then the program can truly be labeled "intelligent" in some non-trivial sense.

Consider, though, what this search entails. The STRIPS program is searching for a correct plan -- that is, a plan which, if executed in the world as it currently stands, would achieve a given goal. This condition -- achieving a given goal -- is not simply a property of the plan; it is a property of the robot's interactions with its world. In order to determine whether a given plan is correct, then, the program must effectively conduct a simulation of the likely outcome of each action. For example, if a candidate plan contains the primitive action "step forward", it matters whether the robot is facing a wall, a door, a pile of rubbish, or an open stretch of floor. If "step forward" is the first step in the plan then the robot can predict its outcome simply by activating its video camera and looking ahead of itself. But if "step forward" is the seventh step in the plan, subsequent to several other movements, then complex reasoning will be required to determine its likely outcome.

This is a severely challenging problem, and Fikes and Nilsson approached it, reasonably enough given the state of computer technology in 1971, through brute force: they encoded the robot's world in the form of a set of formulae in the predicate calculus, and they incorporated into STRIPS a general-purpose program for proving predicate-calculus theorems by means of a search through the space of possible formal proofs. This approach "works" in the same sense that any search method works: if the search ever terminates then the answer is correct, but how long this takes depends heavily on the perspicacity of the program's search control policies. And adequately perspicuous search control policies are notoriously elusive. As programmers like Fikes and Nilsson quickly learned, the trick is to design the world, and the robot's representations of the world, in such a way that long, involved chains of reasoning are not required to predict the outcomes of actions.

Yet predicting the outcomes of actions was, as programmers say, only the "inner loop" of the plan-construction process. Recall that the overall process of choosing possible actions is also structured as a search problem; extending Lashley's linguistic metaphor, it is as if the grammaticality of a spoken sentence depended on the listener's reaction to each successive word. Moreover, the space of possible plans is enormous: at any given time, the robot can take any of about a dozen primitive actions, depending on its immediate circumstances, and even a simple plan will have several steps. Once again, search control policies are crucial. At each point in the search process, the program must make two relatively constrained choices among a manageable list of options: it must choose a partially specified plan to further refine, and it must choose a means of further refining it -- roughly speaking, it must add another primitive action to the plan.

As with any search, making these choices correctly every time would require "intelligence" that no mechanism could probably possess. The point, instead, is to make the choices correctly often enough for the search to settle on a correct answer in a reasonable amount of time. This, once again, is the appeal of heuristic search: intelligent action emerges from a mass of readily mechanizable decisions. In other words, the problem for Fikes and Nilsson was that they had to write bits of code whose outcomes approximated two hopelessly uncomputable notions: "partially specified plan most likely to lead to a correct plan" and "best primitive action to add to this subplan". Their solution to these problems was unsurprising in technical retrospect, and the details do not matter here. Briefly, they chose whatever partially specified plan seemed to have gotten the furthest toward the goal with the smallest number of primitive actions, and they chose a new primitive action that allowed the theorem-proving program to make further progress toward proving that the goal had been achieved. Both of these criteria are virtually guaranteed to lead the plan-construction process down blind alleys (such as telling the robot to head for the door before getting the key). The important point is that these blind alleys did not hurt the robot; they only kept the robot waiting longer to be given a plan to execute.

How big a step was the STRIPS program toward mechanized intelligence? Reasonable people could disagree. It is certainly an impressive thing to watch such a program in operation -- provided you have long enough to wait. But the question of search control was daunting. To the AI research people of that era, search control in STRIPS-like plan-construction programs was a "problem" to be addressed through a wide variety of technical means. Yet this approach accepts as given the underlying structure of the situation: a steep trade-off between the complexity of the world and the practicality of the search control problem. If the robot can perform many possible actions, or if the results of these actions depend in complex ways on the circumstances, then the search space grows rapidly -- in mathematical terms, exponentially -- in size. And if it is impossible to predict the outcomes of actions -- say because the robot is not the only source of change in the world -- then the search space will have to include all of the possible outcomes as well. In a prescient aside in the sequel to the original STRIPS paper, Fikes, Hart, and Nilsson (1972) pointed this out:

One of the novel elements introduced into artificial intelligence research by work on robots is the study of execution strategies and how they interact with planning activities. Since robot plans must ultimately be executed in the real world by a mechanical device, as opposed to being carried out in a mathematical space or by a simulator, consideration must be given by the executor to the possibility that operations in the plan may not accomplish what they were intended to, that data obtained from sensory devices may be inaccurate, and that mechanical tolerances may introduce errors as the plan is executed.

Many of these problems of plan execution would disappear if our system generated a whole new plan after each execution step. Obviously, such a strategy would be too costly, so we instead seek a plan execution scheme with the following properties:

(1) When new information obtained during plan execution implies that some remaining portion of the plan need not be executed, the executor should recognize such information and omit the unneeded plan steps.

(2) When execution of some portion of the plan fails to achieve the intended results, the executor should recognize the failure and either direct reexecution of some portion of the plan or, as a default, call for a replanning activity (1972: 268).

Thus, although they recognized the tension that was inherent in the system of concepts they had inherited, the technical imagination of that time provided Fikes, Hart, and Nilsson with no other way of structuring the basic question of intelligent action. It was fifteen years before the inherent dilemma of plan-construction was given definite mathematical form, first by Chapman (1987) and then more compactly by McAllester and Rosenblitt (1991). This kind of research does not decisively discredit the conceptual framework of planning-as-search; rather it clarifies the precise nature of the trade-offs generated by that framework. And indeed, productive research continues to this day into the formal structure of plan-construction search problems.

6 Beyond the Cartesian soul

The previous sections offer a critical reconstruction of a single strand of intellectual history, a single intellectual proposition worked out in increasingly greater technical detail so that its internal tensions become manifest. To diagnose the resulting impasse and move beyond it, it will be necessary to transcend AI's conception of itself as a technical, formalizing discipline, and instead to reconsider the larger intellectual path of which AI research has been a part. No matter how esoteric AI literature has become, and no matter how thoroughly the intellectual origins of AI's technical methods have been forgotten, the technical work of AI has nonetheless been engaged in an effort to domesticate the Cartesian soul into a technical order in which it does not belong. The problem is not that the individual operations of Cartesian reason cannot be mechanized (they can be) but that the role assigned to the soul in the larger architecture of cognition is untenable. This incompatibility has shown itself up in a pervasive and ever more clear pattern of technical frustrations. The difficulty can be shoved into one area or another through programmers' choices about architectures and representation schemes, but it cannot be made to go away.

This impasse, though, is not a failure. To the contrary, tracing the precise shape of the impasse allows us to delineate with particular confidence the internal tensions in the relational system of ideas around the Cartesian soul. According to this hypothesis, the fundamental embarrassment of Descartes' theory does not lie in the untenability of ontological dualism. Rather, it lies in the soul's causal distance from the world of practical action. As this world grows more complex (or, more precisely, as one's representational schemes reflect this world's complexity more fully), and as one becomes more fully immersed in that world, the soul's job becomes astronomically difficult. Yet Descartes performed his analysis of the soul in sedentary conditions: introspecting, visualizing, and isolating particular episodes of perception. When he did discuss complex activities, he focused not on the practicalities of their organization but on the struggles they engendered between the body and the soul (see, for example, Passions of the Soul, Article 47).

In order to impose intelligent order on its body's actions, the Cartesian soul faces a stern task. For example, to visualize a future course of events, the soul must stimulate the brain to assemble the necessary elements of memory. The reasoning which guides this visualization process must be based in turn upon certain knowledge of the world, obtained through the senses -- enough information to visualize fully the outcomes of the individual's planned sequence of actions. Our judgement that such a scheme places an excessive burden on the soul -- or, as technical people would say, makes the soul into a "bottleneck" -- is not a logical refutation; it is only an engineer's embodied judgement of the implausibility of a design. But within the logic of Descartes' project that is a lot.

The underlying difficulty takes perhaps its clearest form in Lashley. At the beginning of his lecture, he opposes his own view to the behaviorist and reflexological tale of stimuli and responses as follows:

My principal thesis today will be that the input is never into a quiescent or static system, but always into a system which is already actively excited and organized. In the intact organism, behavior is the result of interaction of this background of excitation with input from any designated stimulus. Only when we can state the general characteristics of this background of excitation, can we understand the effects of a given input (page 112).
In contradistinction to a scheme that focuses upon the effects of an isolated stimulus, Lashley proposes giving due weight both to a stimulus and to the ongoing flux of brain activity in which the stimulus intervenes. People, in other words, are always thinking as well as interacting with the world. Having said this, though, he immediately gives priority to the internal "background" of neural activity, and his paper never returns to any consideration of external stimuli and their effects. As with his silence about the nature of the determining tendency, this is not a simple omission but is intrinsic to his relational system of concepts. His analysis of action on the model of speech portrays speakers as laying out a complex series of sounds through internal processing and then producing them in a serially ordered sequence, without in any way interacting with the outside world [6]. As we have already seen in the case of STRIPS, this obscurity about the relationship between "planning" (of action sequences) and "interaction" (with the world while those actions are going on) structured cognitive theorizing about action, and AI research in particular, for many years afterward [7].

It is precisely this pattern of difficulty that has impelled an emerging interdisciplinary movement of computational modelers to seek a conception of intelligent behavior whose focal point is the fullness of embodied activity, not the reticence of thought. An organizing theme of this movement is the principled characterization of interactions between agents and their environments, and the use of such characterizations to guide design and explanation. When the "agents" in question are robots, this theme opens out onto a systems view of robotic activity within the larger dynamics of the robot's world. When the "agents" are animals, it opens out onto biology, and specifically onto a conception of ethology in which creatures and their behavior appear thoroughly adapted to the dynamics of a larger ecosystem. When the "agents" in question are people, it opens out onto philosophical and anthropological conceptions of human beings as profoundly embedded in their social environments. In lieu of detailed references to these directions of research, allow me to me direct the reader to an issue of Artificial Intelligence on Computational Theories of Interaction and Agency that will appear in 1995.

7 AI and the humanities

I have argued that AI can become sterile unless it maintains a sense of its place in the history of ideas -- and in particular unless it maintains a respect for the power of inherited systems of ideas to shape our thinking and our research in the present day. At the same time, AI also provides a powerful means of forcing into the open the internal structure of a system of ideas and the internal tensions inherent in the project of getting those ideas to "work". Thus, AI properly understood ought to be able to participate in a constructive symbiosis with humanistic analyses of ideas.

Putting this mode of cooperative work into practice will not be easy. The obstacles are many and varied, but I believe that the most fundamental ones pertain to the use in AI of mathematics and mathematical formalization. This is not the place for a general treatment of these topics, but it is possible at least to outline some of the issues. The most obvious issue, perhaps, is the symbolic meaning attached to mathematics in the discursive construction of technical disciplines. Technical people frequently speak of mathematics as "clean" and "precise", as opposed to the "messy" and "vague" nature of the social world and humanistic disciplines. These metaphors clearly provide rich points of entry for critical research, but the important point here is that their practical uses go beyond the simple construction of hierarchies among disciplines. Most particularly, the notion of mathematics as the telos of reason structures AI researchers' awareness in profound ways.

To see this, let us briefly consider the role that mathematics plays in AI research. The business of AI is to build computer programs whose operation can be narrated in language that is normally used in describing human activities (Collins 1990). Since the function of computers is specified in terms of discrete mathematics, the daily work of AI includes the complex and subtle discursive practice of talking about human activities in ways that assimilate them to mathematical structures [8]. In the case of the computational models of action described above, this assimilation is achieved by means of a linguistic metaphor for action. This metaphor is not specific to AI; in fact it structures a great deal of the practice of applied computing (Agre 1994) [9]. And this fact in turn points to an inherent source of intellectual conservatism in AI: the field is not restricted a priori to speaking of human beings in particular terms, but it is restricted to speaking in terms that someone knows how to assimilate to mathematical structures that can be programmed on computers. In this way, the existing intellectual infrastructure of computing -- its stock of discursive forms and technical methods -- drags like an anchor behind any project that would reinvent AI using language drawn from alternative conceptions of human beings and their lives.

This observation goes far toward explaining the strange appearance that AI presents to fields such as literature and anthropology that routinely employ much more sophisticated and critically reflective conceptions of human life. The first priority for AI research is to get something working on a computer, and the field does not reward gnawing doubts about whether the conceptions of human life being formalized along the way are sufficiently subtle, accurate, or socially responsible -- thus the emphasis, mentioned at the outset, on "doing" as opposed to "just talking". Critical methods from the humanities are likely to appear pointless, inasmuch as they do not immediately deliver formalizations or otherwise explain what programs one might write. AI people see formalization as a trajectory with an endpoint, in which the vagueness and ambiguity of ordinary language are repaired through mathematical definition, and they are not greatly concerned with the semantic violence that might be done to that language in the process of formal definition. A word like "action" might present real challenges to a philosophical project that aims to respect ordinary usage (e.g., White 1968), but the assimilation of action to formal language theory reduces the word to a much simpler form: a repertoire of possible "actions" assembled from a discrete, finite vocabulary of "expressive elements" or "primitives". Having thus taken its place in the technical vocabulary of AI, the word's original semantic ramifications are lost as potential resources for AI work. The ideology surrounding formalization accords no intrinsic value to these left-over materials. As a result, formalization becomes a highly organized form of social forgetting -- and not only of the semantics of words but of their historicity as well. This is why the historical provenance and intellectual development of AI's underlying ideas claim so little interest among the field's practitioners.

What would a reformed AI look like? It would certainly not reject or replace mathematics. Rather, it would draw upon critical research to cultivate a reflexive awareness of how mathematical formalization is used as part of the engineer's embodied work of building things and seeing what they do. In particular, it would cultivate an awareness of the cycle of formalization, technical working-out, the emergence of technical impasses, the critical work of diagnosing the impasse as reflecting either a superficial or a profound difficulty with the underlying conception of action, and the initiation of new and more informed rounds of formal modeling. The privilege in this cycle does not lie with the formalization process, nor does it lie with the critical diagnosis of technical impasses. Rather, it lies with the cycle itself, in the researcher's "reflective conversation with the materials" of technical and critical work (Schon 1983).

Humanistic critical practice can take up numerous relationships, cooperative or not, to this cycle of research. My own analysis in this paper has employed a relatively old-fashioned set of humanistic methods from the history of ideas, tracing the continuity of certain themes across a series of authors and their intellectual projects. Since formalization is a fundamentally metaphorical process, discursively interrelating one set of things with another, mathematical set, it can be particularly fruitful to trace the historical travels of a given metaphor among various institutional sites in society, technical and otherwise (Martin 1987, McReynolds 1980, Mirowski 1989). The purpose in doing so is not simply to debunk any claims that technical institutions might make to an ahistorical authority, but to prevent the passage to formalism from forgetting the underlying commitments that a given way of speaking about human activities draws from its broader cultural embedding [10]. This contextual awareness will be crucial when the technical research reaches an impasse and needs to be diagnosed as a manifestation of internal tensions of the underlying system of ideas. Any given set of ideas will be more easily given up when they are seen as simply one path among many others not taken. Indeed this awareness of context will be crucial for recognizing that an impasse may have occurred in the first place. Viewed in this way, technical impasses are a form of social remembering, moments when a particular discursive form deconstructs itself and makes its internal tensions intelligible to anyone who is critically equipped to hear them.

The cycle of reaching and interpreting technical impasses, moving back and forth between technical design and critical inquiry, can be practiced on a variety of scales, depending upon the acuity of one's critical methods. The example I traced in the body of this paper was extremely coarse: whole decades of research could be seen in hindsight to have been working through a single, clear-cut intellectual problem. The difficulty was not that AI practitioners were insulated from the philosophical critiques of Cartesian reason that might have provided a diagnosis of their difficulties and defined the contours of alternative territories of research. To the contrary, Hubert Dreyfus was articulating some of these critiques all along. The real difficulty was that the critical apparatus of the field did not provide its practitioners with a living, day-to-day appreciation for the contingent nature of their formalisms. Although they viewed formalization as conferring upon language a cleanliness and precision that it did not otherwise possess, the effect was precisely the reverse. Lacking a conscious awareness of the immense historicity of their language, they could not understand it as it called out to them the very things they had discovered. A reformed technical practice would employ the tools of critical inquiry to engage in a richer and more animated conversation with the world.

* Footnotes

[1] For a detailed analysis of this argument see Preston (1993).

[2] Dreyfus, in joint work with Stuart Dreyfus, has been cautiously supportive of one alternative AI research program, the "connectionist" attempt to build simulations of neural circuitry without necessarily formulating "knowledge" in terms of symbolic "rules" (Dreyfus and Dreyfus 1988). But as the Dreyfuses points out, this research program still faces a long, difficult learning curve and will not be discussed here.

[3] The terms "mechanism" and "mechanistic" require further analysis than space permits here. Suffice it to say that a mechanism is a physical object whose workings are wholly explicable in causal terms. To speak of something as a mechanism, furthermore, is to insert it into a rhetoric of engineering design, whether divine or human, and whether on the model of the clockmaker or the computer programmer. For the modern mathematical intepretations of the term, which are obviously relevant to the foundations of computing if not immediately to the genealogy being traced here, see Webb (1980).

[4] Note that the intellectual culture of Descartes' day did not distinguish between "mind" and "soul", and the two terms continue to be used interchangeably in Catholic philosophy to this day; see for example Holscher (1986). Even in the present day, these terms are usually not so much opposed as simply employed in different discourses with overlapping genealogies.

[5] Chapman (1987) presents a genealogy of the AI "planning" systems in this lineage.

[6] Actual human speakers frequently do interact with their addressees and others during the real-time production of their utterances (Goodwin 1981), but this fact is rarely taken into account in cognitive theories of grammar and speech.

[7] It is particularly clear in the opening chapter of Miller, Galanter, and Pribram's influential book Plans and the Structure of Behavior (1960).

[8] This is obviously an attribute that AI shares with a wide variety of other fields, for example mathematical economics, and much of the analysis here applies to these other fields as well. It should be noted that AI people themselves place great emphasis on a distinction between "neat" forms of AI, which openly avow their commitment to mathematical formalization and employ large amounts of mathematical notation in their papers, and "scruffy" forms, which do not (Forsythe 1993). My argument, though, applies equally to both forms of AI research. Regardless of whether its author was consciously thinking in terms of mathematics, a computer program is a notation whose operational semantics can be specified in mathematical terms. While the formalizations in "neat" research are frequently more consistent, systematic, and explicit than those of "scruffy" research, the design of any computer program necessarily entails a significant level of formalization.

[9] Different linguistic metaphors for human action are obviously possible, if perhaps equally problematic; see for example Ricoeur (1971).

[10] For an impressive cultural analysis of the origins of AI, see Edwards (1996).

* Acknowledgements

This paper has been improved by comments from Harry Collins, Guven Guzeldere, Scott Mainwaring, Beth Preston, and Joszef Toth.

* References

Philip E. Agre, Interview with Allen Newell, Artificial Intelligence 59(1-2), 1993, pages 415-449.

Philip E. Agre, Surveillance and capture: Two models of privacy, The Information Society 10(2), 1994, pages 101-127.

David Chapman, Planning for conjunctive goals, Artificial Intelligence, 32(3), 1987, pages 333-377.

Christopher Cherniak, Minimal Rationality, Cambridge: MIT Press, 1986.

Noam Chomsky, Aspects of the Theory of Syntax, Cambridge: MIT Press, 1965.

Noam Chomsky, Problems of Knowledge and Freedom: The Russell Lectures, New York: Pantheon, 1971.

Noam Chomsky, Language and Responsibility, translated from the French by John Viertel, New York: Pantheon, 1979.

Harry M. Collins, Artificial Experts: Social Knowledge and Intelligent Machines, Cambridge: MIT Press, 1990.

Daniel Dennett, Why the law of effect will not go away, in Brainstorms: Philosophical Essays on Mind and Psychology, Montgomery, VT: Bradford Books, 1978.

Rene Descartes, The Philosophical Works of Rene Descartes, translated by Elizabeth S. Haldane and G. R. T. Ross, volume 1, Cambridge: Cambridge University Press, 1972.

Hubert L. Dreyfus, What Computers Can't Do: A Critique of Artificial Reason, New York: Harper and Row, 1972.

Hubert L. Dreyfus and Stuart Dreyfus. Making a mind vs. modeling the brain: AI back at a branchpoint, Daedalus 117(1), 1988, pages 15-43.

Paul Edwards, The Closed World: Computers and the Politics of Discourse in Cold War America, Cambridge: MIT Press, 1996.

Richard E. Fikes and Nils J. Nilsson, STRIPS: A new approach to the application of theorem proving to problem solving, Artificial Intelligence 2(3), 1971, pages 189-208.

Richard E. Fikes, Peter E. Hart, and Nils J. Nilsson, Learning and executing generalized robot plans, Artificial Intelligence 3(4), 1972, pages 251-288.

Diane E. Forsythe, Engineering knowledge: The construction of knowledge in artificial intelligence, Social Studies of Science 23(3), 1993, pages 445-477.

C. R. Gallistel, The Organization of Action: A New Synthesis, Hillsdale, NJ: Erlbaum, 1980.

Charles Goodwin, Conversational Organization: Interaction Between Speakers and Hearers, New York: Academic Press, 1981.

Martin Heidegger, Being and Time, translated by John Macquarrie and Edward Robinson, Harper and Row, 1961. Originally published in German in 1927.

Ludger Holscher, The Reality of the Mind: Augustine's Philosophical Arguments for the the Human Soul as a Spiritual Substance, London: Routledge, 1986.

Karl S. Lashley, The problem of serial order in behavior, in Lloyd A. Jeffress, ed, Cerebral Mechanisms in Behavior: The Hixon Symposium, New York: Wiley, 1951.

Emily Martin, The Woman in the Body: A Cultural Analysis of Reproduction, Boston: Beacon Press, 1987.

David McAllester and David Rosenblitt, Systematic nonlinear planning, Proceedings of the National Conference on Artificial Intelligence, Los Altos, CA: Morgan Kaufmann Publishers, 1991, pages 634-639.

Paul McReynolds, The clock metaphor in the history of psychology, in Thomas Nickles, ed, Scientific Discovery: Case Studies, Dordrecht: Reidel, 1980.

George A. Miller, Information and memory, Scientific American 195(2), 1956, pages 42-46.

George A. Miller, Eugene Galanter, and Karl H. Pribram, Plans and the Structure of Behavior, New York: Henry Holt and Company, 1960.

Philip Mirowski, More Heat Than Light: Economics as Social Physics, Physics as Nature's Economics, Cambridge: Cambridge University Press, 1989.

Allen Newell and Herbert A. Simon, GPS: A program that simulates human thought, in Edward A. Feigenbaum and Julian Feldman, eds, Computers and Thought, New York: McGraw-Hill, 1963.

Elizabeth F. Preston, Heidegger and artificial intelligence, Philosophy and Phenomenological Research 53(1), 1993, pages 43-69.

Paul Ricoeur, The model of the text: Meaningful action considered as a text, Social Research 38, 1971, pages 529-562.

Donald A. Schon, The Reflective Practitioner: How Professionals Think in Action, New York: Basic Books, 1983.

Brian C. Smith, Prologue to Reflection and Semantics in a Procedural Language, in Ronald J. Brachman and Hector J. Levesque, eds, Readings in Knowledge Representation, Los Altos, CA: Morgan Kaufmann, 1985.

Judson Chambers Webb, Mechanism, Mentalism, and Mathematics: An Essay on Finitism, Dordrecht: Reidel, 1980.

Alan R. White, ed, The Philosophy of Action, London: Oxford University Press, 1968.

Ludwig Wittgenstein, Philosophical Investigations, third edition, translated by G. E. M. Anscombe, New York: Macmillan, 1968. Originally published in 1953.