Stanford Encyclopedia of Philosophy

A | B | C | D | E | F | G | H | I | J | K | L | M | N | O | P | Q | R | S | T | U | V | W | X | Y | Z

Artificial Intelligence

Artificial intelligence (AI) is the field devoted to building artificial animals (or at least artificial creatures that -- in suitable contexts -- appear to be animals) and, for many, artificial persons (or at least artificial creatures that -- in suitable contexts -- appear to be persons). Such goals immediately ensure that AI is a discipline of considerable interest to many philosophers, and this has been confirmed (e.g.) by the energetic attempt, on the part of numerous philosophers, to show that these goals are in fact un/attainable. On the constructive side, many of the core formalisms and techniques used in AI come out of, and are indeed still much used and refined in, philosophy: first-order logic, intensional logics suitable for the modeling of doxastic attitudes and deontic reasoning, inductive logic, probability theory and probabilistic reasoning, practical reasoning and planning, and so on. In light of this, some philosophers conduct AI research and development as philosophy.

In the present entry, the history of AI is briefly recounted, proposed definitions of the field are discussed, and an overview of the field is provided. In addition, both philosophical AI (AI pursued as and out of philosophy) and philosophy of AI are discussed, via examples of both. The entry ends with some speculative commentary regarding the future of AI.

The History of AI

The field of artificial intelligence (AI) officially started in 1956, launched by a small but now-famous DARPA-sponsored summer conference at Dartmouth College, in Hanover, New Hampshire. (The 50-year celebration of this conference, AI@50, was held in July 2006 at Dartmouth, with five of the original participants making it back. What happened at this historic conference figures in the final section of this entry.) Ten thinkers attended, including John McCarthy (who was working at Dartmouth in 1956), Claude Shannon, Marvin Minsky, Arthur Samuel, Trenchard Moore (apparently the lone note-taker at the original conference), Ray Solomonoff, Oliver Selfridge, Allen Newell, and Herbert Simon. From where we stand now, at the start of the new millennium, the Dartmouth conference is memorable for many reasons, including this pair: one, the term ‘artificial intelligence’ was coined there (and has long been firmly entrenched, despite being disliked by some of the attendees, e.g., Moore); two, Newell and Simon revealed a program -- Logic Theorst (LT) -- agreed by the attendees (and, indeed, by nearly all those who learned of and about it soon after the conference) to be a remarkable achievement. LT was capable of proving elementary theorems in the propositional calculus.[1]

Though the term ‘artificial intelligence’ made its advent at the 1956 conference, certainly the field of AI was in operation well before 1956. For example, in a famous Mind paper of 1950, Alan Turing argues that the question “Can a machine think?” (and here Turing is talking about standard computing machines: machines capable of computing only functions from the natural numbers (or pairs, triples, ... thereof) to the natural numbers that a Turing machine or equivalent can handle) should be replaced with the question “Can a machine be linguistically indistinguishable from a human?.” Specifically, he proposes a test, the “Turing Test” (TT) as it's now known. In the TT, a woman and a computer are sequestered in sealed rooms, and a human judge, in the dark as to which of the two rooms contains which contestant, asks questions by email (actually, by teletype, to use the original term) of the two. If, on the strength of returned answers, the judge can do no better than 50/50 when delivering a verdict as to which room houses which player, we say that the computer in question has passed the TT. Passing in this sense operationalizes linguistic indistinguishability. Later, we shall discuss the role that TT has played, and indeed coninues to play, in attempts to define AI. At the moment, though, the point is that in his paper, Turing explicitly lays down the call for building machines that would provide an existence proof of an affirmative answer to his question. The call even includes a suggestion for how such construction should proceed. (He suggests that “child machines” be built, and that these machines could then gradually grow up on their own to learn to communicate in natural language at the level of adult humans. This suggestion has arguably been followed by Rodney Brooks and the philosopher Daniel Dennett in the Cog Project: (Dennett 1994). In addition, the Spielberg/Kubrick movie A.I. is at least in part a cinematic exploration of Turing's suggestion.) The TT continues to be at the heart of AI and discussions of its foundations, as confirmed by the appearance of (Moor 2003). In fact, the TT continues to be used to define the field, as in Nilsson's (1998) position, expressed in his textbook for the field, that AI simply is the field devoted to building an artifact able to negotiate this test.

Returning to the issue of the historical record, even if one bolsters the claim that AI started at the 1956 conference by adding the proviso that ‘artificial intelligence’ refers to a nuts-and-bolts engineering pursuit (in which case Turing's philosphical discussion, despite calls for a child machine, wouldn’t exactly count as AI per se), one must confront the fact that Turing, and indeed many predecessors, did attempt to build intelligent artifacts. In Turing's case, such building was surprisingly well-understood before the advent of programmable computers: Turing wrote a program for playing chess before there were computers to run such programs on, by slavishly following the code himself. He did this well before 1950, and long before Newell (1973) gave thought in print to the possibility of a sustained, serious attempt at building a good chess-playing computer.[2]

From the standpoint of philosophy, neither the 1956 conference, nor Turing's Mind paper, come close to marking the start of AI. This is easy enough to see. For example, Descartes proposed TT (not the TT by name, of course) long before Turing was born.[3] Here's the relevant passage:

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 recognise 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 fall short in others, by which means we may discover that they did not act from knowledge, but only for 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. (Descartes 1911, p. 116)

At the moment, Descartes is certainly carrying the day.[4] Turing predicted that his test would be passed by 2000, but the fireworks-across-the-globe start of the new millennium has long since died down, and the most articulate of computers still can't meaningfully debate a sharp toddler. Moreover, while in certain focussed areas machines out-perform minds (IBM's famous Deep Blue prevailed in chess over Gary Kasparov, e.g.), minds have a (Cartesian) capacity for cultivating their expertise in virtually any sphere. (If it were announced to Deep Blue, or any current successor, that chess was no longer to be the game of choice, but rather a heretofore unplayed variant of chess, the machine would be trounced by human children of average intelligence having no chess expertise.) AI simply hasn't managed to create general intelligence; it hasn't even managed to produce an artifact indicating that eventually it will create such a thing.

But what if we consider the history of AI not from the standpoint of philosophy, but rather from the standpoint of the field with which, today, it is most closely connected? The reference here is to computer science. From this standpoint, does AI run back to well before Turing? Interestingly enough, the results are the same: we find that AI runs deep into the past, and has always had philosophy in its veins. This is true for the simple reason that computer science grew out of logic and probability theory, which in turn grew out of (and is still intertwined with) philosophy. Computer science, today, is shot through and through with logic; the two fields cannot be separated. This phenomenon has become an object of study unto itself (Halpern et al. 2001). The situation is no different when we are talking not about traditional logic, but rather about probabilistic formalisms, also a significant component of modern-day AI: These formalisms also grew out of philosophy, as nicely chronicled, in part, by Glymour (1992). For example, in the one mind of Pascal was born a method of rigorously calculating probabilities, conditional probability that plays a large role in AI to this day, and such fertile philosophico-probabilistic arguments as Pascal's wager, according to which it is irrational not to become a Christian.

That modern-day AI has its roots in philosophy, and in fact that these historical roots are temporally deeper than even Descartes’ distant day, can be seen by looking to the clever, revealing cover of the comprehensive textbook Artificial Intelligence: A Modern Approach (known in the AI community as simply AIMA for (Russell & Norvig 2002)).

What you see there is an eclectic collection of memorabilia that might be on and around the desk of some imaginary AI researcher. For example, if you look carefully, you will specifically see: a picture of Turing, a view of Big Ben through a window (perhaps R&N are aware of the fact that Turing famously held at one point that a physical machine with the power of a universal Turing machine is physically impossible: he quipped that it would have to be the size of Big Ben), a planning algorithm described in Aristotle's De Motu Animalium, Frege's fascinating notation for first-order logic, a glimpse of Lewis Carroll’s (1958) pictorial representation of syllogistic reasoning, Ramon Lull’s concept-generating wheel from his 13th-century Ars Magna, and a number of other pregnant items (including, in a clever, recursive, and bordering-on-self-congratulatory touch, a copy of AIMA itself). Though there is insufficient space here to make all the historical connections, we can safely infer from the appearance of these items that AI is indeed very, very old. Even those who insist that AI is at least in part an artifact-building enterprise must concede that, in light of these objects, AI is ancient, for it isn’t just theorizing from the perspective that intelligence is at bottom computational that runs back into the remote past of human history: Lull’s wheel, for example, marks an attempt to capture intelligence not only in computation, but in a physical artifact that embodies that computation.

One final point about the history of AI seems worth making.

It is generally assumed that the birth of modern-day AI in the 1950’s came in large part because of and through the advent of the modern high-speed digital computer. This assumption accords with common-sense. After all, AI (and, for that matter, to some degree its cousin, cognitive science, particularly computational cognitive modeling, the sub-field of cognitive science devoted to producing computational simulations of human cognition) is aimed at implementing intelligence in a computer, and it stands to reason that such a goal would be inseparably linked with the advent of such devices. However, this is only part of the story: the part that reaches back but to Turing and others (e.g., von Neuman) responsible for the first electronic computers. The other part is that, as already mentioned, AI has a particularly strong tie, historically speaking, to reasoning (logic-based and, in the need to deal with uncertainty, probabilistic reasoning). In this story, nicely told by Glymour (1992), a search for an answer to the question “What is a proof?” eventually led to an answer based on Frege’s version of first-order logic (FOL): a mathematical proof consists in a series of step-by-step inferences from one formula of first-order logic to the next. The obvious extension of this answer (and it isn’t a complete answer, given that lots of classical mathematics, despite conventional wisdom, clearly can’t be expressed in FOL; even the Peano Axioms require SOL) is to say that not only mathematical thinking, but thinking, period, can be expressed in FOL. (This extension was entertained by many logicians long before the start of information-processing psychology and cognitive science -- a fact some cognitive psychologists and cognitive scientists often seem to forget.) Today, logic-based AI is only part of AI, but the point is that this part still lives (with help from logics much more powerful, but much more complicated, than FOL), and it can be traced all the way back to Aristotle's theory of the syllogism. In the case of uncertain reasoning, the question isn’t “What is a proof?”, but rather questions such as “What is it rational to believe, in light of certain observations and probabilities?” This is a question posed and tackled before the arrival of digital computers.

What Exactly is AI?

So far we have been proceeding as if we have a firm grasp of AI. But what exactly is AI? Philosophers arguably know better than anyone that defining disciplines can be well nigh impossible. What is physics? What is biology? What, for that matter, is philosophy? These are remarkably difficult, maybe even eternally unanswerable, questions. Perhaps the most we can manage here under obvious space constraints is to present in encapsulated form some proposed definitions of AI. We do include a glimpse of recent attempts to define AI in detailed, rigorous fashion.

Russell and Norvig (1995, 2002), in their aforementioned AIMA text, provide a set of possible answers to the “What is AI?” question that has considerable currency in the field itself. These answers all assume that AI should be defined in terms of its goals: a candidate definition thus has the form “AI is the field that aims at building ...” The answers all fall under a quartet of types placed along two dimensions. One dimension is whether the goal is to match human performance, or, instead, ideal rationality. The other dimension is whether the goal is to build systems that reason/think, or rather systems that act. The situation is summed up in this table:

Four Possible Goals for AI According to AIMA
Human-Based Ideal Rationality
Reasoning-Based: Systems that think like humans. Systems that think rationally.
Behavior-Based: Systems that act like humans. Systems that act rationally.

Please note that this quartet of possibilities does reflect (at least a significant portion of) the relevant literature. For example, philosopher John Haugeland (1985) falls into the Human/Reasoning quadrant when he says that AI is “The exciting new effort to make computers think ... machines with minds, in the full and literal sense.” Luger and Stubblefield (1993) seem to fall into the Ideal/Act quadrant when they write: “The branch of computer science that is concerned with the automation of intelligent behavior.” The Human/Act position is occupied most prominently by Turing, whose test is passed only by those systems able to act sufficiently like a human. The “thinking rationally” position is defended (e.g.) by Winston (1992).

It’s important to know that the contrast between the focus on systems that think/reason versus systems that act, while found, as we have seen, at the heart of AIMA, and at the heart of AI itself, should not be interpreted as implying that AI researchers view their work as falling all and only within one of these two compartments. Researchers who focus more or less exclusively on knowledge representation and reasoning, are also quite prepared to acknowledge that they are working on (what they take to be) a central component or capability within any one of a family of larger systems spanning the reason/act distinction. The clearest case may come from the work on planning -- an AI area traditionally making central use of representation and reasoning. For good or ill, much of this research is done in abstraction (in vitro, as opposed to in vivo), but the researchers involved certainly intend or at least hope that the results of their work can be embedded into systems that actually do things, such as, for example, execute the plans.

What about Russell and Norvig themselves? What is their answer to the What is AI? question? They are firmly in the the “acting rationally” camp. In fact, it’s safe to say both that they are the chief proponents of this answer, and that they have been remarkably successful evangelists. Their extremely influential AIMA can be viewed as a book-length defense and specification of the Ideal/Act category. We will look a bit later at how Russell and Norvig lay out all of AI in terms of intelligent agents, which are systems that act in accordance with various ideal standards for rationality. But first let’s look a bit closer at the view of intelligence underlying the AIMA text. We can do so by turning to (Russell 1997). Here Russell recasts the “What is AI?” question as the question “What is intelligence?” (presumably under the assumption that we have a good grasp of what an artifact is), and then he identifies intelligence with rationality. More specifically, Russell sees AI as the field devoted to building intelligent agents, which are functions taking as input tuples of percepts from the external environment, and producing behavior (actions) on the basis of these percepts. Russell’s overall picture is this one:

The Basic Picture Underlying Russell’s Account of Intelligence/Rationality

Let’s unpack this diagram a bit, and take a look, first, at the account of perfect rationality that can be derived from it. The behavior of the agent in the environment E (from a class E of environments) produces a sequence of states or snapshots of that environment. A performance measure U evaluates this sequence; notice the utility box in the previous figure. We let V(f,E,U) denote the expected utility according to U of the agent function f operating on E. Now we identify a perfectly rational agent with the agent function

fopt = argmaxf V(f,E,U)

Of course, as Russell points out, it’s usually not possible to actually build perfectly rational agents. For example, though it’s easy enough to specify an algorithm for playing invincible chess, it’s not feasible to implement this algorithm. What traditionally happens in AI is that programs that are -- to use Russell’s apt terminology -- calculatively rational are constructed instead: these are programs that, if executed infinitely fast, would result in perfectly rational behavior. In the case of chess, this would mean that we strive to write a program that runs an algorithm capable, in principle, of finding a flawless move, but we add features that truncate the search for this move in order to play within intervals of digestible duration.

Russell himself champions a new brand of intelligence/rationality for AI; he calls this brand bounded optimality. To understand Russell’s view, first we follow him in introducing a distinction: we say that agents have two components: a program, and a machine upon which the program runs. We write Agent(P,M) to denote the agent function implemented by program P running on machine M. Now, let (M) denote the set of all programs P that can run on machine M. The bounded optimal program Popt then is:

Popt = argmaxP V(Agent(P,M),E,U)

You can understand this equation in terms of any of the mathematical idealizations for standard computation. For example, machines can be identified with Turing machines minus instructions (i.e., TMs are here viewed architecturally only: as having tapes divided into squares upon which symbols can be written, read/write heads capable of moving up and down the tape to write and erase, and control units which are in one of a finite number of states at any time), and programs can be identified with instructions in the Turing machine model (telling the machine to write and erase symbols, depending upon what state the machine is in). So, if you are told that you must “program” within the constraints of a 22-state Turing machine, you could search for the “best” program given those constraints. In other words, you could strive to find the optimal program within the bounds of the 22-state architecture. Russell’s (1997) view is thus that AI is the field devoted to creating optimal programs for intelligent agents, under time and space constraints on the machines implementing these programs.[5]

It should be mentioned that there is a different, much more straightforward answer to the “What is AI?” question. This answer, which goes back to the days of the original Dartmouth conference, was expressed by, among others, Newell (1973), one of the grandfathers of modern-day AI (recall that he attended the 1956 conference); it is:

AI is the field devoted to building artifacts that are intelligent, where ‘intelligent’ is operationalized through intelligence tests (such as the Wechsler Adult Intelligence Scale), and other tests of mental ability (including, e.g., tests of mechanical ability, creativity, and so on).

Though few are aware of this now, this answer was taken quite seriously for a while, and in fact underlied one of the most famous programs in the history of AI: the ANALOGY program of Evans (1968), which solved geometric analogy problems of a type seen in many intelligence tests. An attempt to rigorously define this forgotten form of AI (as what they dub Psychometric AI), and to resurrect it from the days of Newell and Evans, is provided by Bringsjord and Schimanski (2003). Recently, a sizable private investment has been made in the ongoing attempt, known as Project Halo, to build a “digital Aristotle”, in the form of a machine able to excel on standardized tests such at the AP exams tackled by US high school students (Friedland et al. 2004). In addition, researchers at Northwestern have forged a connection between AI and tests of mechanical ability (Klenk et al. 2005).

In the end, as is the case with any discipline, to really know precisely what that discipline is requires you to, at least to some degree, dive in and do, or at least dive in and read. Two decades ago such a dive was quite manageable. Today, because the content that has come to constitute AI has mushroomed, the dive (or at least the swim after it) is a bit more demanding. Before looking in more detail at the content that composes AI, we take a quick look at the explosive growth of AI.

The Explosive Growth of AI

First, a point of clarification. The growth of which we speak is not a shallow sort correlated with amount of funding provided for a given sub-field of AI. That kind of thing happens all the time in all fields, and can be triggered by entirely political and financial changes designed to grow certain areas, and diminish others. Rather, we are speaking of an explosion of deep content: new material which someone intending to be conversant with the field needs to know. Relative to other fields, the size of the explosion may or may not be unprecedented. (Though it should perhaps be noted that an analogous increase in philosophy would be marked by the development of entirely new formalisms for reasoning, reflected in the fact that, say, longstanding philosophy textbooks like Copi’s (2004) Introduction to Logic are dramatically rewritten and enlarged to include these formalisms, rather than remaining anchored to essentially immutable core formalisms, with incremental refinement around the edges through the years.) But it certainly appears to be quite remarkable, and is worth taking note of here, if for no other reason than that AI’s near-future will revolve in significant part around whether or not the new content in question forms a foundation for new long-lived research and development that would not otherwise obtain.

Were you to have begun formal coursework in AI in 1985, your textbook would likely have been Eugene Charniak's comprehensive-at-the-time Introduction to Artificial Intelligence (Charniak & McDermott 1985). This book gives a strikingly unified presentation of AI -- as of the early 1980’s. This unification is achieved via first-order logic (FOL), which runs throughout the book and binds things together. For example: In the chapter on computer vision (3), everyday objects like bowling balls are represented in FOL. In the chapter on parsing language (4), the meaning of words, phrases, and sentences are identified with corresponding formulae in FOL (e.g., they reduce “the red block” to FOL on page 229). In Chapter 6, “Logic and Deduction”, everything revolves around FOL and proofs therein (with an advanced section on nonmonotonic reasoning couched in FOL as well). And Chapter 8 is devoted to abduction and uncertainty, where once again FOL, not probability theory, is the foundation. It’s clear that FOL renders (Charniak & McDermott 1985) esemplastic. Today, due to the explosion of content in AI, this kind of unification is no longer possible.

Though there is no need to get carried away in trying to quantify the explosion of AI content, it isn't hard to begin to do so for the inevitable skeptics. (Charniak & McDermott 1985) has 710 pages. The first edition of AIMA, published ten years later in 1995, has 932 pages, each with about 20% more words per page than C&M's book. The second edition of AIMA weighs in at a backpack-straining 1023 pages, with new chapters on probabilistic language processing, and uncertain temporal reasoning.

The explosion of AI content can also be seen topically. C&M cover nine highest-level topics, each in some way tied firmly to FOL implemented in (a dialect of) the programming language Lisp, and each (with the exception of Deduction, whose additional space testifies further to the centrality of FOL) covered in one chapter:

  1. FOL for Internal Representation
  2. Vision
  3. Language Parsing
  4. Language Understanding
  5. Search Techniques
  6. Deduction (two chapters)
  7. Abduction and Expert Systems
  8. Planning
  9. Learning

In AIMA the expansion is obvious. For example, Search is given three full chapters, and Learning is given four chapters. AIMA also includes coverage of topics not present in C&M's book; one example is robotics, which is given its own chapter in AIMA. In the second edition, as mentioned, there are two new chapters: one on constraint satisfaction that constitutes a lead-in to logic, and one on uncertain temporal reasoning that covers hidden Markov models, Kalman filters, and dynamic Bayesian networks. A lot of other additional material appears in new sections introduced into chapters seen in the first edition. For example, the second edition includes coverage of propositional logic as a bona fide framework for building significant intelligent agents. In the first edition, such logic is introduced mainly to facilitate the reader's understanding of full FOL.

The Resurgence of Neurocomputational Techniques

One of the remarkable aspects of (Charniak & McDermott 1985) is this: The authors say the central dogma of AI is that “What the brain does may be thought of at some level as a kind of computation” (p. 6). And yet nowhere in the book is brain-like computation discussed. In fact, you will search the index in vain for the term ‘neural’ and its variants. Please note that the authors are not to blame for this. A large part of AI’s growth has come from formalisms, tools, and techniques that are, in some sense, brain-based, not logic-based. A recent paper that conveys the importance and maturity of neurocomputation is (Litt et al. 2006). (Growth has also come from a return of probabilistic techniques that had withered by the mid-70’s and 80’s. More about that momentarily, in the next “resurgence” section.)

One very prominent class of non-logicist formalism does make an explicit nod in the direction of the brain: viz., artificial neural networks (or as they are often simply called, neural networks, or even just neural nets). (The structure of neural networks is discussed below). Because Minsky and Pappert's (1969) Perceptrons led many (including, specifically, many sponsors of AI research and development) to conclude that neural networks didn't have sufficient information-processing power to model human cognition, the formalism was pretty much universally dropped from AI. However, Minsky and Pappert had only considered very limited neural networks. Connectionism, the view that intelligence consists not in symbolic processing, but rather non-symbolic processing at least somewhat like what we find in the brain (at least at the cellular level), approximated specifically by artificial neural networks, came roaring back in the early 1980’s on the strength of more sophisticated forms of such networks, and soon the situation was (to use a metaphor introduced by John McCarthy) that of two horses in a race toward building truly intelligent agents.

If one had to pick a year at which connectionism was resurrected, it would certainly be 1986, the year Parallel Distributed Processing (Rumelhart & McClelland 1986) appeared in print. The rebirth of connectionism was specifically fueled by the back-propagation algorithm over neural networks, nicely covered in Chatper 20 of AIMA. The symbolicist/connectionist race led to a spate of lively debate in the literature (e.g., Smolensky 1988, Bringsjord 1991), and some AI engineers have explicitly championed a methodology marked by a rejection of knowledge representation and reasoning. For example, Rodney Brooks was such an engineer; he wrote the well-known “Intelligence Without Representation” (1991), and his Cog Project, to which we referred above, is arguably an incarnation of the premeditatedly non-logicist approach. Increasingly, however, those in the business of building sophisticated systems find that both logicist and more neurocomputational techniques are required (Wermter & Sun 2001).[6] In addition, the neurocomputational paradigm today includes connectionism only as a proper part, in light of the fact that some of those working on building intelligent systems strive to do so by engineering brain-based computation outside the neural network-based approach (e.g., Granger 2004a, 2004b).

The Resurgence of Probabilistic Techniques

There is a second dimension to the explosive growth of AI: the explosion in popularity of probabilistic methods that aren’t neurocomputational in nature, in order to formalize and mechanize a form of non-logicist reasoning in the face of uncertainty. Interestingly enough, it is Eugene Charniak himself who can be safely considered one of the leading proponents of an explicit, premeditated turn away from logic to statistical techniques. His area of specialization is natural language processing, and whereas his introductory textbook of 1985 gave an accurate sense of his approach to parsing at the time (as we have seen, write computer programs that, given English text as input, ultimately infer meaning expressed in FOL), this approach was abandoned in favor of purely statistical approaches (Charniak 1993). At the recent AI@50 conference, Charniak boldly proclaimed, in a talk tellingly entitled “Why Natural Language Processing is Now Statistical Natural Language Processing,” that logicist AI is moribund, and that the statistical approach is the only promising game in town -- for the next 50 years.[7] The chief source of energy and debate at the conference flowed from the clash between Charniak's probabilistic orientation, and the original logicist orientation, upheld at the conference in question by John McCarthy and others.

AI's use of probability theory grows out of the standard form of this theory, which grew directly out of technical philosophy and logic. This form will be familiar to many philosophers, but let's review it quickly now, in order to set a firm stage for making points about the new probabilistic techniques that have energized AI.

Just as in the case of FOL, in probability theory we are concerned with declarative statements, or propositions, to which degrees of belief are applied; we can thus say that both logicist and probabilistic approaches are symbolic in nature. More specifically, the fundamental proposition in probability theory is a random variable, which can be conceived of as an aspect of the world whose status is initially unknown. We usually capitalize the names of random variables, though we reserve p, q, r, ... as such names as well. In a particular murder investigation centered on whether or not Mr. Black committed the crime, the random variable Guilty might be of concern. The detective may be interested as well in whether or not the murder weapon -- a particular knife, let us assume -- belongs to Black. In light of this, we might say that Weapon = true if it does, and Weapon = false if it doesn't. As a notational convenience, we can write weapon and weapon for these two cases, respectively; and we can use this convention for other variables of this type.

The kind of variables we have described so far are Boolean, because their domain is simply {true, false}. But we can generalize and allow discrete random variables, whose values are from any countable domain. For example, PriceTChina might be a variable for the price of (a particular, presumably) tea in China, and its domain might be {1, 2, 3, 4, 5}, where each number here is in US dollars. A third type of variable is continuous; its domain is either the reals, or some subset thereof.

We say that an atomic event is an assignment of particular values from the appropriate domains to all the variables composing the (idealized) world. For example, in the simple murder investigation world introduced just above, we have two Boolean variables, Guilty and Weapon, and there are just four atomic events. Note that atomic events have some obvious properties. For example, they are mutually exclusive, exhaustive, and logically entail the truth or falsity of every proposition. Usually not obvious to beginning students is a fourth property, namely, any proposition is logically equivalent to the disjunction of all atomic events that entail that proposition.

Prior probabilities correspond to a degree of belief accorded a proposition in the complete absence of any other information. For example, if the prior probability of Black's guilt is .2, we write

P(Guilty = true) = .2

or simply P(guilty) = .2. It is often convenient to have a notation allowing one to refer economically to the probabilities of all the possible values for a random variable. For example, we can write


as an abbreviation for the five equations listing all the possible prices for tea in China. We can also write

P(PriceTChina) = <1, 2, 3, 4, 5>

In addition, as further convenient notation, we can write P(Guilty, Weapon) to denote the probabilities of all combinations of values of the relevant set of random variables. This is referred to as the joint probability distribution of Guilty and Weapon. The full joint probability distribution covers the distribution for all the random variables used to describe a world. Given our simple murder world, we have 20 atomic events summed up in the equation

P(Guilty, Weapon, PriceTChina)

The final piece of the basic language of probability theory corresponds to conditional probabilities. Where p and q are any propositions, the relevant expression is P(p|q), which can be interpreted as “the probability of p, given that all we know is q.” For example,

P(guilty|weapon) = .7

says that if the murder weapon belongs to Black, and no other information is available, the probability that Black is guilty is .7.

Andrei Kolmogorov showed how to construct probability theory from three axioms that make use of the machinery now introduced, viz.,

  1. All probabilities fall between 0 and 1. I.e., ∀ p 0 ≤ P(p) ≤ 1.
  2. Valid (in the traditional logicist sense) propositions have a probability of 1; unsatisfiable (in the traditional logicist sense) propositions have a probability of 0.
  3. P(p ∨ q) = P(p) + P(q) - P(p ∧ q),
These axioms are clearly at bottom logicist. The remainder of probability theory can be erected from this foundation (conditional probabilities are easily defined in terms of prior probabilities). We can thus say that logic is in some fundamental sense sense still being used to characterize the set of beliefs that a rational agent can have. But where does probabilistic inference enter the picture on this account, since traditional deduction is not used for inference in probability theory?

Probabilistic inference consists in computing, from observed evidence expressed in terms of probability theory, posterior probabilities of propositions of interest. For a good long while, there have been algorithms for carrying out such computation. These algorithms precede the resurgence of probabilistic techniques in the 1990’s. (Chapter 13 of AIMA presents a number of them.) For example, given the Kolmogorov axioms, here is a straightforward way of computing the probability of any propostion, using the full joint distribution giving the probabilities of all atomic events: Where p is some proposition, let α(p) be the disjunction of all atomic events in which p holds. Since the probability of a proposition (i.e., P(p)) is equal to the sum of the probabilities of the atomic events in which it holds, we have an equation that provides a method for computing the probability of any proposition p, viz.,

P(p) = Σei ∈ α (p) P(ei)

Unfortunately, there were two serious problems infecting this original probabilistic approach: One, the processing in question needed to take place over paralyzingly large amounts of information (enumeration over the entire distribution is required). And two, the expressivity of the approach was merely propositional. (It was by the way the philosopher Hilary Putnam (1963) who pointed out that there was a price to pay in moving to the first-order level. The issue is not discussed herein.) Everything changed with the advent of a new formalism that marks the marriage of probabilism and graph theory: Bayesian networks (also called belief nets). The pivotal text was (Pearl 1988).

To explain Bayesian networks, and to provide a contrast between Bayesian probabilistic inference, and argument-based approaches that are likely to be attractive to classically trained philosophers, let us build upon the example of Black introduced above. Suppose that we want to compute the posterior probability of the guilt of our murder suspect, Mr. Black, from observed evidence. We have three Boolean variables in play: Guilty, Weapon, and Intuition. Weapon is true or false based on whether or not a murder weapon (the knife, recall) belonging to Black is found at the scene of the bloody crime. The variable Intuition is true provided that the very experienced detective in charge of the case, Watson, has an intuition, without examining any physical evidence in the case, that Black is guilty; intuition holds just in case Watson has no intuition either way. Here is a table that holds all the (eight) atomic events in the scenario so far:

weapon weapon
intuition intuition intuition intuition
guilty 0.208 0.016 0.072 0.008
guilty 0.011 0.063 0.134 0.486

Were we to add the aforeintroduced discrete random variable PriceTChina, we would of course have 40 events, corresponding in tabular form to the preceding table associated with each of the five possible values of PriceTChina. That is, there are 40 events in

P(Guilty, Weapon, Intuition, PriceTChina)

Bayesian networks provide a economical way to represent the situation. Such networks are directed, acyclic graphs in which nodes correspond to random variables. When there is a directed link from node Ni to node Nj, we say that Ni is the parent of Nj. With each node Ni there is a corresponding conditional probability distribution


where, of course, Parents(Ni) denotes the parents of Ni. The following figure shows such a network for the case we have been considering. The specific probability information is omitted; readers should at this point be able to readily calculate it using the machinery provided above.

A Simple Bayesian Net

Notice the economy of the network, in striking contrast to the prospect, visited above, of listing all 40 possibilities. The price of tea in China is presumed to have no connection to the murder, and hence the relevant node is isolated. In addition, only some l probability info is included, corresponding to the relevant tables shown in the figure (typically termed a conditional probability table). And yet from a Bayesian network, every entry in the full joint distribution can be easily calculated, as follows. First, for each node/variable Ni we write Ni = ni to indicate an assignment to that node/variable. The conjunction of the specific assignments to every variable in the full joint probability distribution can then be written as

P(N1 = n1 ... Nn = nn)

and abbreviated as P(n1, ..., nn). Where parents(Ni) denotes the specific assignments to the variables in the set of all parents of Ni, we can use a Bayesian net to produce that value of any entry via this equation:
Πni = 1 P(ni|parents(Ni))

Earlier, we observed that the full joint distribution can be used to infer an answer to queries about the domain. Given this, it follows immediately that Bayesian networks have the same power. But in addition, there are much much efficient methods over such networks for answering queries. These methods, and increasing the expressivity of networks toward the first-order case, are outside the scope of the present entry. Readers are directed to AIMA, or any of the other textbooks affirmed in this entry (see note 8).

Before concluding this section, it is probably worth noting that, from the standpoint of philosophy, a situation such as the murder investigation we have exploited above would often be analyzed into arguments, and strength factors, not into numbers to be crunched by purely arithmetical procedures. For example, in the epistemology of Roderick Chisholm, as presented his Theory of Knowledge (Chisholm 1966, 1977), Detective Watson might classify a proposition like Black committed the murder. as counterbalanced if he was unable to take a find a compelling argument either way, or perhaps probable if the murder weapon turned out to belong to Black. Such categories cannot be found on a continuum from 0 to 1, and they are used in articulating arguments for or against Black's guilt. Argument-based approaches to uncertain and defeasible reasoning are virtually non-existent in AI. One exception is Pollock's approach, covered below. This approach is Chisholmian in nature.

Approaches to AI

There are a number of ways of “carving up” AI. By far the most prudent and productive way to summarize the field is to turn yet again to the AIMA text, by any metric a masterful, comprehensive overview of the field.[8]

The Intelligent Agent Continuum

As Russell and Norvig (2002) tell us in the Preface of AIMA:

[Our] main unifying theme is the idea of an intelligent agent. We define AI as the study of agents that receive percepts from the environment and perform actions. Each such agent implements a function that maps percept sequences to actions, and we cover different ways to represent these functions... (Russell & Norvig 2002, vii)
The basic picture is thus summed up in this figure:

Impressionistic Overview of an Intelligent Agent

The content of AIMA derives, essentially, from fleshing out this picture; that is, corresponds to the different ways of representing the overall function that intelligent agents implement. And there is a progression from the least powerful agents up to the more powerful ones. The following figure gives a high-level view of a simple kind of agent discussed early in the book. (Though simple, this sort of agent corresponds to the architecture of representation-free agents designed and implemented by Rodney Brooks 1991.)

A Simple Reflex Agent

As the book progresses, agents get increasingly sophisticated, and the implementation of the function they represent thus draws from more and more of what AI can currently muster. The following figure gives an overview of an agent that is a bit smarter than the simple reflex agent. This smarter agent has the ability to internally model the outside world, and is therefore not simply at the mercy of what can at the moment be directly sensed.

A More Sophisticated Reflex Agent

There are eight parts to AIMA. As the reader passes through these parts, she is introduced to agents that take on the powers discussed in each part. Part I is an introduction to the agent-based view. Part II is concerned with giving an intelligent agent the capacity to think ahead a few steps in clearly defined environtments. Examples here include agents able to successfully play games of perfect information, such as chess. Part III deals with agents that have declarative knowledge and can reason in ways that will be quite familiar to most philosophers and logicians (e.g., knowledge-based agents deduce what actions should be taken to secure their goals). Part IV of the book outfits agents with the power to handle uncertainty by reasoning in probabilistic fashion. In Part VI, agents are given a capacity to learn. The following figure shows the overall structure of a learning agent.

A Learning Agent

The final set of powers agents are given allow them to communicate. These powers are covered in Part VII.

Philosophers who patiently travel the entire progression of increasingly smart agents will no doubt ask, when reaching the end of Part VII, if anything is missing. Are we given enough, in general, to build an artificial person, or is there enough only to build a mere animal? This question is implicit in the following from Charniak and McDermott (1985):

The ultimate goal of AI (which we are very far from achieving) is to build a person, or, more humbly, an animal. (Charniak & McDermott 1985, 7)

To their credit, Russell & Norvig, in AIMA's Chapter 27, “AI: Present and Future,” consider this question, at least to some degree. They do so by considering some challenges to AI that have hitherto not been met. One of these challenges is described by R&N as follows:

[M]achine learning has made very little progress on the important problem of constructing new representations at levels of abstraction higher than the input vocaulary. For example, how can an autonomous robot generate useful predicates such as Office and Cafe if they are not supplied to it be humans? Similar considerations apply to learning behavior--HavingACupOfTea is an important high-level action, but how does it get into an action library that initially contains much simpler actions such as RaiseArm and Swallow? Unless we understand such issues, we are faced with the daunting task of constructing large commonsense knowledge bases by hand. (Russell & Norvig 2002, p. 970)

This specific challenge is actually merely the foothill before a dizzyingly high mountain that AI must eventually somehow manage to climb. That mountain, put simply, is reading. Despite the fact that, as noted, Part IV of AIMA is devoted to machine learning, AI, as it stands, offers next to nothing in the way of a mechanization of learning by reading. Yet when you think about it, reading is probably the dominant way you learn at this stage in your life. Consider what you're doing at this very moment. It’s a good bet that you are reading this sentence because, earlier, you set yourself the goal of learning about the field of AI. Yet the formal models of learning provided in AIMA's Part IV (which are all and only the models at play in AI) cannot be applied to learning by reading.[9] These models all start with a function-based view of learning. According to this view, to learn is almost invariably to produce an underlying function f on the basis of a restricted set of pairs (a1, f(a1)), (a2, f(a2)), ..., (an, f(an)). For example, consider receiving inputs consisting of 1, 2, 3, 4, and 5, and corresponding range values of 1, 4, 9, 16, and 25; the goal is to “learn” the underlying mapping from natural numbers to natural numbers. In this case, assume that the underlying function is n2, and that you do “learn“ it. While this narrow model of learning can be productively applied to a number of processes, the process of reading isn’t one of them. Learning by reading cannot (at least for the foreseeable future) be modeled as divining a function that produces argument-value pairs. Instead, your reading about AI can pay dividends only if your knowledge has increased in the right way, and if that knowledge leaves you poised to be able to produce behavior taken to confirm sufficient mastery of the subject area in question. This behavior can range from correctly answering and justifying test questions regarding AI, to producing a robust, compelling presentation or paper that signals your achievement.

Two points deserve to be made about machine reading. First, it may not be clear to all readers that reading is an ability that is central to intelligence. The centrality derives from the fact that intelligence requires vast knowledge. We have no other means of getting systematic knowledge into a system than to get it in from text, whether text on the web, text in libraries, newspapers, and so on. You might even say that the big problem with AI has been that machines really don't know much compared to humans. That can only be because of the fact that humans read (or hear: illiterate people can listen to text being uttered and learn that way). Either machines gain knowledge by humans manually encoding and inserting knowledge, or by reading and listening. These are brute facts. (We leave aside supernatural techniques, of course. Oddly enough, Turing didn't: he seemed to think ESP should be discussed in connection with the powers of minds and machines. See (Turing 1950.))

Now for the second point. Humans able to read have invariably also learned a language, and learning languages has been modeled in conformity to the function-based approach adumbrated just above (Osherson et al. 1986). However, this doesn't entail that an artificial agent able to read, at least to a significant degree, must have really and truly learned a natural language. AI is first and foremost concerned with engineering computational artifacts that measure up to some test (where, yes, sometimes that test is from the human sphere), not with whether these artifacts process information in ways that match those present in the human case. It may or may not be necessary, when engineering a machine that can read, to imbue that machine with human-level linguistic competence. The issue is empirical, and as time unfolds, and the engineering is pursued, we shall no doubt see the issue settled.

It would seem that the greatest challenges facing AI are ones the field apparently hasn't even come to grips with yet. Ssome mental phenomena of paramount importance to many philosohers of mind and neuroscience are simply missing from AIMA. Two examples are subjective consciousness and creativity. The former is only mentioned in passing in AIMA, but subjective consciousness is the most important thing in our lives -- indeed we only desire to go on living because we wish to go on enjoying subjective states of certain types. Moreover, if human minds are the product of evolution, then presumably phenomenal consciousness has great survival value, and would be of tremendous help to a robot intended to have at least the behavioral repertoire of the first creatures with brains that match our own (hunter-gatherers; see Pinker 1997). Of course, subjective consciousness is largely missing from the sister fields of cognitive psychology and computational cognitive modeling as well.[10]

To some readers, it might seem in the very least tendentious to point to subjective consciousness as a major challenge to AI that it has yet to address. These readers might be of the view that pointing to this problem is to look at AI through a distinctively philosophical prism, and indeed a controversial philosophical standpoint.

But as its literature makes clear, AI measures itself by looking to animals and humans and picking out in them remarkable mental powers, and by then seeing if these powers can be mechanized. Arguably the power most important to humans (the capacity to experience) is nowhere to be found on the target list of most AI researchers. There may be a good reason for this (no formalism is at hand, perhaps), but there is no denying the state of affairs in question obtains, and that, in light of how AI measures itself, that it’s worrisome.

As to creativity, it's quite remarkable that the power we most praise in human minds is nowhere to be found in AIMA. Just as in (Charniak & McDermott 1985) one cannot find ‘neural’ in the index, ‘creativity’ can't be found in the index of AIMA. This is particularly odd because many AI researchers have in fact worked on creativity (especially those coming out of philosophy; e.g., Boden 1994, Bringsjord & Ferrucci 2000).

Although the focus has been on AIMA, any of its counterparts could have been used. As an example, consider Artificial Intelligence: A New Synthesis, by Nils Nilsson. (A synopsis and TOC are available at Http:// As in the case of AIMA, everything here revolves around a gradual progression from the simplest of agents (in Nilsson's case, reactive agents), to ones having more and more of those powers that distinguish persons. Energetic readers can verify that there is a striking parallel between the main sections of Nilsson's book and AIMA. In addition, Nilsson, like Russell and Norvig, ignores phenomenal consciousness, reading, and creativity. None of the three are even mentioned.

A final point to wrap up this section. It seems quite plausible to hold that there is a certain inevitability to the structure of an AI textbook, and the apparent reason is perhaps rather interesting. In personal conversation, Jim Hendler, a well-known AI researcher who is one of the main innovators behind Semantic Web (Berners-Lee, Hendler, Lassila 2001), an under-development “AI-ready” version of the World Wide Web, has said that this inevitability can be rather easily displayed when teaching Introduction to AI; here's how. Begin by asking students what they think AI is. Invariably, many students will volunteer that AI is the field devoted to building artificial creatures that are intelligent. Next, ask for examples of intelligent creatures. Students always respond by giving examples across a continuum: simple multi-celluar organisms, insects, rodents, lower mammals, higher mammals (culminating in the great apes), and finally human persons. When students are asked to describe the differences between the creatures they have cited, they end up essentially describing the progression from simple agents to ones having our (e.g.) communicative powers. This progression gives the skeleton of every comprehensive AI textbook. Why does this happen? The answer seems clear: it happens because we can’t resist conceiving of AI in terms of the powers of extant creatures with which we are familiar. At least at present, persons, and the creatures who enjoy only bits and pieces of personhood, are -- to repeat -- the measure of AI.

Logic-Based AI: Some Surgical Points

SEP already contains a separate entry entitled Logic and Artificial Intelligence, written by Thomason. This entry is focused on non-monotonic reasoning, and reasoning about time and change; the entry also provides a history of the early days of logic-based AI, making clear the contributions of those who founded the tradition (e.g., John McCarthy and Pat Hayes; see their seminal 1969 paper). Reasoning based on classical deductive logic is monotonic; that is, if , then for all , {} . Commonsense reasoning is not monotonic. While you may currently believe on the basis of reasoning that your house is still standing, if while at work you see on your computer screen that a vast tornado is moving through the location of your house, you will drop this belief. The addition of new information causes previous inferences to fail. In the simpler example that has become an AI staple, if I tell you that Tweety is a bird, you will infer that Tweety can fly, but if I then inform you that Tweety is a penguin, the inference evaporates, as well it should. Non-monotonic (or defeasible) logic includes formalisms designed to capture the mechanisms underlying these kinds of examples.

The formalisms and techniques discussed in Logic and Artificial Intelligence have now reached, as of 2006, a level of impressive maturity -- so much so that in various academic and corporate laboratories, implementations of these formalisms and techniques can be used to engineer robust, real-world software. It is strongly recommend that readers who have assimilated Thomason's entry and have an interest to learn where AI stands in these areas consult (Mueller 2006), which provides, in one volume, integrated coverage of non-monotonic reasoning (in the form, specifically, of circumscription, introduced by Thomason), and reasoning about time and change in the situation and event calculi. (The former calculus is also introduced by Thomason. In the second, timepoints are included, among other things.) The other nice thing about (Mueller 2006) is that the logic used is multi-sorted first-order logic (MSL), which has unificatory power that will be known to and appreciated by many technical philosophers and logicians (Manzano 1996).

In the present entry, three topics of importance in AI not covered in Logic and Artificial Intelligence are mentioned. They are:

  1. The overarching scheme of logicist AI, in the context of the attempt to build intelligent artificial agents.
  2. Common Logic and the intensifying quest for interoperability.
  3. A technique that can be called encoding down, which can allow machines to reason efficiently over knowledge that, were it not encoded down, would, when reasoned over, lead to paralyzing inefficiency.
This trio is covered in order, beginning with the first.

Detailed accounts of logicist AI that fall under the agent-based scheme can be found in (Nilsson 1991, Bringsjord & Ferrucci 1998).[11]. The core idea is that an intelligent agent receives percepts from the external world in the form of formulae in some logical system (e.g., first-order logic), and infers, on the basis of these percepts and its knowledge base, what actions should be performed to secure the agent's goals. (This is of course a barbaric simplification. Information from the external world is encoded in formulae, and transducers to accomplish this feat may be components of the agent.)

To clarify things a bit, we consider, briefly, the logicist view in connection with arbitrary logical systems X.[12] We obtain a particular logical system by setting X in the appropriate way. Some examples: If X = I, then we have a system at the level of FOL [following the standard notation from model theory; see e.g. (Ebbinghaus et al. 1984)]. II is second-order logic, and 1 is a “small system” of infinitary logic (countably infinite conjunctions and disjunctions are permitted). These logical systems are all extensional, but there are intensional ones as well. For example, we can have logical systems corresponding to those seen in standard propositional modal logic (Chellas 1980). One possibility, familiar to many philosophers, would be propositional KT45, or KT45.[13] In each case, the system in question includes a relevant alphabet from which well-formed formulae are constructed by way of a formal grammar, a reasoning (or proof) theory, a formal semantics, and at least some meta-theoretical results (soundness, completeness, etc.). Taking off from standard notation, we can thus say that a set of formulas in some particular logical system X, X, can be used, in conjunction with some reasoning theory, to infer some particular formula X. (The reasoning may be deductive, inductive, abductive, and so on. Logicist AI isn't in the least restricted to any particular mode of reasoning.) To say that such a sitution holds, we write


When the logical system referred to is clear from context, or when we don't care about which logical system is involved, we can simply write

Each logical system, in its formal semantics, will include objects designed to represent ways the world pointed to by formulae in this system can be. Let these ways be denoted by WiX. When we aren't concerned with which logical system is involved, we can simply wrte Wi. To say that such a way models a formula we write


We extend this to a set of formulas in the natural way: Wi means that all the elements of are true on Wi. Now, using the simple machinery we’ve established, we can describe, in broad strokes, the life of an intelligent agent that conforms to the logicist point of view. This life conforms to the basic cycle that undergirds intelligent agents in the AIMA2e sense.

To begin, we assume that the human designer, after studying the world, uses the language of a particular logical system to give to our agent an initial set of beliefs Δ0 about what this world is like. In doing so, the designer works with a formal model of this world, W, and ensures that W Δ0. Following tradition, we refer to Δ0 as the agent's (starting) knowledge base. (This terminology, given that we are talking about the agent's beliefs, is known to be peculiar, but it persists.) Next, the agent ADJUSTS its knowlege base to produce a new one, Δ1. We say that adjustment is carried out by way of an operation ; so 0] = Δ1. How does the adjustment process, , work? There are many possibilities. Unfortunately, many believe that the simplest possibility (viz., i] equals the set of all formulas that can be deduced in some elementary manner from Δi) exhausts all the possibilities. The reality is that adjustment, as indicated above, can come by way of any mode of reasoning -- induction, abduction, and yes, various forms of deduction corresponding to the logical system in play. For present purposes, it’s not important that we carefully enumerate all the options.

The cycle continues when the agent ACTS on the environment, in an attempt to secure its goals. Acting, of course, can cause changes to the environment. At this point, the agent SENSES the environment, and this new information Γ1 factors into the process of adjustment, so that 1 ∪ Γ1] = Δ2. The cycle of SENSES ADJUSTS ACTS continues to produce the life Δ0, Δ1, Δ2, Δ3, ... of our agent.

It may strike you as preposterous that logicist AI be touted as an approach taken to replicate all of cognition. Reasoning over formulae in some logical system might be appropriate for computationally capturing high-level tasks like trying to solve a math problem (or devising an outline for an entry in the Stanford Encyclopedia of Philosophy), but how could such reasoning apply to tasks like those a hawk tackles when swooping down to capture scurrying prey? In the human sphere, the task successfully negotiated by athletes would seem to be in the same category. Surely, some will declare, an outfielder chasing down a fly ball doesn’t prove theorems to figure out how to pull off a diving catch to save the game!

Needless to say, such a declaration has been carefully considered by logicists. For example, Rosenschein and Kaelbling (1986) describe a method in which logic is used to specify finite state machines. These machines are used at “run time” for rapid, reactive processing. In this approach, though the finite state machines contain no logic in the traditional sense, they are produced by logic and inference. Recently, real robot control via first-order theorem proving has been demonstrated by Amir and Maynard-Reid (1999, 2000, 2001). In fact, you can download version 2.0 of the software that makes this approach real for a Nomad 200 mobile robot in an office environment. Of course, negotiating an office environment is a far cry from the rapid adjustments an outfielder for the Yankees routinely puts on display, but certainly it’s an open question as to whether future machines will be able to mimic such feats through rapid reasoning. The question is open if for no other reason than that all must concede that the constant increase in reasoning speed of first-order theorem provers is breathtaking. (For up-to-date news on this increase, visit and monitor the TPTP site.) There is no known reason why the software engineering in question cannot continue to produce speed gains that would eventually allow an artificial creature to catch a fly ball by processing information in purely logicist fashion.

Now we come to the second topic related to logicist AI that warrants mention herein: common logic and the intensifying quest for interoperability between logic-based systems using different logics. Only a few brief comments are offered. Readers wanting more can explore the links provided in the course of the summary.

To begin, please understand that AI has always been very much much at the mercy of the vicissitudes of funding provided to researchers in the field by the United States Department of Defense (DoD). (The inaugural 1956 workshop was funded by DARPA, and many representatives from this organization attended AI@50.) It’s this fundamental fact that causally contributed to the temporary hibernation of AI carried out on the basis of artificial neural networks: When Minsky and Pappert (1959) bemoaned the limitations of neural networks, it was the funding agencies that held back money for research based upon them. Since the late 1950's it's safe to say that the DoD has sponsored the development of many logics intended to advance AI and lead to helpful applications. Recently, it has occurred to many in the DoD that this sponsorship has led to a plethora of logics between which no translation can occur. In short, the situation is a mess, and now real money is being spent to try to fix it, through standardization and machine translation (between logical, not natural, languages).

The standardization is coming chiefly through what is known as Common Logic (CL), and variants thereof. (CL is soon to be an ISO standard. ISO is the International Standards Organization.) Philosophers interested in logic, and of course logicians, will find CL to be quite fascinating. (From an historical perspective, the advent of CL is interesting in no small part because the person spearheading it is none other than Pat Hayes, the same Hayes who, as we have seen, worked with McCarthy to establish logicist AI in the 1960’s. Though Hayes was not at the original 1956 Dartmouth conference, he certainly must be regarded as one of the founders of contemporary AI.) One of the interesting things about CL, at least as I see it, is that it signifies a trend toward the marriage of logics, and programming languages and environments. Another system that is a logic/programming hybrid is Athena, which can be used as a programming language, and is at the same time a form of MSL. Athena is known as a denotational proof language (Arkoudas 2000).

How is interoperability between two systems to be enabled by CL? Suppose one of these systems is based on logic L, and the other on L'. (To ease exposition, assume that both logics are first-order.) The idea is that a theory L, that is, a set of formulae in L, can be translated into CL, producing CL, and then this theory can be translated into L'. CL thus becomes an inter lingua. Note that what counts as a well-formed formula in L can be different than what counts as one in L'. The two logics might also have different proof theories. For example, inference in L might be based on resolution, while inference in L' is of the natural deduction variety. Finally, the symbol sets will be different. Despite these differences, courtesy of the translations, desired behavior can be produced across the translation. That, at any rate, is the hope. The technical challenges here are immense, but federal monies are increasingly available for attacks on the problem of interoperability.

Now for the third topic in this section: what can be called encoding down. The technique is easy to understand. Suppose that we have on hand a set of first-order axioms. As is well-known, the problem of deciding, for arbitrary formula , whether or not it's deducible from is Turing-undecidable: there is no Turing machine or equivalent that can correctly return Yes or No in the general case. However, if the domain in question is finite, we can encode this problem down to the propositional calculus. An assertion that all things have F is of course equivalent to the assertion that Fa, Fb, Fc, as long as the domain contains only these three objects. So here a first-order quantified formula becomes a conjunction in the propositional calculus. Determining whether such conjunctions are provable from axioms themselves expressed in the propositional calculus is Turing-decidable, and in addition, in certain clusters of cases, the check can be done very quickly in the propositional case; very quickly. Readers interested in encdoing down to the propositional calculus should consult recent DARPA-sponsored work by Bart Selman. Please note that the target of encoding down doesn't need to be the propositional calculus. Because it's generally harder for machines to find proofs in an intensional logic than in straight first-order logic, it is often expedient to encode down the former to the latter. For example, propositional modal logic can be encoded in multi-sorted logic (a variant of FOL); see (Arkoudas & Bringsjord 2005).

Non-Logicist AI: A Summary

It’s tempting to define non-logicist AI by negation: an approach to building intelligent agents that rejects the distinguishing features of logicist AI. Such a shortcut would imply that the agents engineered by non-logicist AI researchers and developers, whatever the virtues of such agents might be, cannot be said to know that -- for the simple reason that, by negation, the non-logicist paradigm would have not even a single declarative proposition that is a candidate for . However, this isn't a particularly enlightening way to define non-symbolic AI. A more productive approach is to say that non-symbolic AI is AI carried out on the basis of particular formalisms other than logical systems, and to then enumerate those formalisms. It will turn out, of course, that these formalisms fail to include knowledge in the normal sense. (In philosophy, as is well-known, the normal sense is one according to which if p is known, p is a declarative statement.)

From the standpoint of formalisms other than logical systems, non-logicist AI can be partitioned into symbolic but non-logicist approaches, and connectionist/neurocomputational approaches. (AI carried out on the basis of symbolic, declarative structures that, for readability and ease of use, are not treated directly by researchers as elements of formal logics, does not count. In this category fall traditional semantic networks, Schank's (1972) conceptual dependency scheme, and other schemes.) The former approaches, today, are probabilistic, and are based on the formalisms (Bayesian networks) covered above. The latter approaches are based, as we have noted, on formalisms that can be broadly termed “neurocomputational.” Given our space constraints, only one of the formalisms in this category is described here (and briefly at that): the aforementioned artificial neural networks.[14]

Neural nets are composed of units or nodes designed to represent neurons, which are connected by links designed to represent dendrites, each of which has a numeric weight.

A “Neuron” Within an Artificial Neural Network

It is usually assumed that some of the units work in symbiosis with the external environment; these units form the sets of input and output units. Each unit has a current activation level, which is its output, and can compute, based on its inputs and weights on those inputs, its activation level at the next moment in time. This computation is entirely local: a unit takes account of but its neighbors in the net. This local computation is calculated in two stages. First, the input function, ini, gives the weighted sum of the unit's input values, that is, the sum of the input activations multiplied by their weights:
ini = ∑jWjiaj.

In the second stage, the activation function, g, takes the input from the first stage as argument and generates the output, or activation level, ai:
ai = g(ini) = g ( ∑jWjiaj ).

One common (and confessedly elementary) choice for the activation function (which usually governs all units in a given net) is the step function, which usually has a threshold t that sees to it that a 1 is output when the input is greater than t, and that 0 is output otherwise. This is supposed to be “brain-like” to some degree, given that 1 represents the firing of a pulse from a neuron through an axon, and 0 represents no firing. A simple three-layer neural net is shown in the following picture.

A Simple Three-Layer Artificial Neural Network

As you might imagine, there are many different kinds of neural networks. The main distinction is between feed-forward and recurrent networks. In feed-forward networks like the one pictured immediately above, as their name suggests, links move information in one direction, and there are no cycles; recurrent networks allow for cycling back, and can become rather complicated. In general, though, it now seems safe to say that neural networks are fundamentally plagued by the fact that while they are simple, efficient learning algorithms are possible, but when they are multi-layered and thus sufficiently expressive to represent non-linear functions, they are very hard to train.

Perhaps the best technique for teaching students about neural networks in the context of other statistical learning formalisms and methods is to focus on a specific problem, preferably one that seems unnatural to tackle using logicist techniques. The task is then to seek to engineer a solution to the problem, using any and all techniques available. One nice problem is handwriting recognition (which also happens to have a rich philosophical dimension; see e.g. Hofstadter & McGraw 1995). For example, consider the problem of assigning, given as input a handwritten digit d, the correct digit, 0 through 9. Because there is a database of 60,000 labeled digits available to researchers (from the National Institute of Science and Technology), this problem has evolved into a benchmark problem for comparing learning algorithms. It turns out that kernel machines currently reign as the best approach to the problem -- despite the fact that, unlike neural networks, they require hardly any prior iteration. A nice summary of fairly recent results in this competition can be found in Chapter 20 of AIMA.

Readers interested in AI (and computational cognitive science) pursued from an overtly brain-based orientation are encouraged to explore the work of Rick Granger (2004a, 2004b) and researchers in his Brain Engineering Laboratory and W.H. Neukom Institute for Computational Sciences. The contrast between the “dry”, logicist AI started at the original 1956 conference, and the approach taken here by Granger and associates (in which brain circuitry is directly modeled) is remarkable.

AI Beyond the Clash of Paradigms

At this point the reader has been exposed to the chief formalisms in AI, and may wonder about heterogeneous approaches that bridge them. Is there such research and development in AI? Yes. But coverage of this work is beyond the scope of the present entry. From an engineering standpoint, such work makes irresistably good sense. There is now an understanding that, in order to build applications that get the job done, one should choose from a toolbox that includes logicist, probabilistic/Bayesian, and neurocomputational techniques. Given that the original top-down logicist paradigm is alive and thriving (e.g., see Brachman & Levesque 2004, Mueller 2006, and ***), and that, as noted, a resurgence of Bayesian and neurocomputational approaches has placed these two paradigms on solid, fertile footing as well, AI now moves forward, armed with this fundamental triad, and it is a virtual certainty that applications (e.g., robots) will be engineered by drawing from elements of all three.

What, though, about deep, theoretical integration of the main paradigms in AI? Such integration is at present only a possibility for the future, but readers are directed to the research of some striving for such integration. For example: Sun (1994, 2002) has been working to demonstrate that human cognition that is on its face symbolic in nature (e.g., professional philosophizing in the analytic tradition, which deals explicitly with arguments and definitions carefully symbolized) can arise from cognition that is neurocomputational in nature. Koller (1997) has investigated the marriage between probability theory and logic. And, in general, the very recent arrival of so-called human-level AI is being led by theorists seeking to genuinely integrate the three paradigms set out above (e.g., Cassimatis 2006).

Philosophical AI

Notice that the heading for this section isn't Philosophy of AI. We’ll get to that category momentarily. Philosophical AI is AI, not philosophy; but it’s AI rooted in and flowing from, philosophy. Before we ostensively characterize Philosophical AI courtesy of a particular research program, let us consider the view that AI is in fact simply philosophy, or a part thereof.

Is AI a Branch of Philosophy?

Daniel Dennett (1979) has famously claimed not just that there are parts of AI intimately bound up with philosophy, but that AI is philosophy (and psychology, at least of the cognitive sort). (He has made a parallel claim about Artificial Life (Dennett 1998).) This view will turn out to be incorrect, but the reasons why it’s wrong will prove illuminating, and our discussion will pave the way for a discussion of Philosophical AI.

What does Dennett say, exactly? This:

I want to claim that AI is better viewed as sharing with traditional epistemology the status of being a most general, most abstract asking of the top-down question: how is knowledge possible? (Dennett 1979, 60)

Elsewhere he says his view is that AI should be viewed “as a most abstract inquiry into the possibility of intelligence or knowledge” (Dennett 1979, 64).

In short, Dennett holds that AI is the attempt to explain intelligence, not by studying the brain in the hopes of identifying components to which cognition can be reduced, and not by engineering small information-processing units from which one can build in bottom-up fashion to high-level cognitive processes, but rather by -- and this is why he says the approach is top-down -- designing and implementing abstract algorithms that capture cognition. Leaving aside the fact that, at least starting in the early 1980's, AI includes an approach that is in some sense bottom-up (see the neurocomputational paradigm discussed above, in Non-Logicist AI: A Summary; and see, specifically, Granger's (2004a, 2004b) work, hyperlinked in text immediately above, a specific counterexample), a fatal flaw infects Dennett's view. Dennett sees the potential flaw, as reflected in:

It has seemed to some philosophers that AI cannot plausibly be so construed because it takes on an additional burden: it restricts itself to mechanistic solutions, and hence its domain is not the Kantian domain of all possible modes of intelligence, but just all possible mechanistically realizable modes of intelligence. This, it is claimed, would beg the question against vitalists, dualists, and other anti-mechanists. (Dennett 1979, 61)

Dennett has a ready answer to this objection. He writes:

But ... the mechanism requirement of AI is not an additional constraint of any moment, for if psychology is possible at all, and if Church's thesis is true, the constraint of mechanism is no more severe than the constraint against begging the question in psychology, and who would wish to evade that? (Dennett 1979, 61)

Unfortunately, this is acutely problematic; and examination of the problems throws light on the nature of AI.

First, insofar as philosophy and psychology are concerned with the nature of mind, they aren't in the least trammeled by the presupposition that mentation consists in computation. AI, at least of the “Strong” variety (we'll discuss “Strong” versus “Weak” AI below) is indeed an attempt to substantiate, through engineering certain impressive artifacts, the thesis that intelligence is at bottom computational (at the level of Turing machines and their equivalents, e.g., Register machines). So there's a philosophical claim, for sure. But this doesn't make AI philosophy, any more than some of the deeper, more aggressive claims of some physicists (e.g., that the universe is ultimately digital in nature; see make their field philosophy. Philosophy of physics certainly entertains the proposition that the physical universe can be perfectly modeled in digital terms (in a series of cellular automata, e.g.), but of course philosophy of physics can't be identified with this doctrine.

Second, we now know well (and those familiar with the relevant formal terrain knew at the time of Dennett's writing) that information processing can exceed standard computation, that is, can exceed computation at and below the level of what a Turing machine can muster (Turing-computation, we shall say). (Such information processing is known as hypercomputation, a term coined by philosopher Jack Copeland, who has himself defined such machines (e.g., Copeland 1998). The first machines capable of hypercomputation were trial-and-error machines, introduced in the same famous issue of the Journal of Symbolic Logic (Gold 1965, Putnam 1965). A new hypercomputer is the infinite time Turing machine (Hamkins & Lewis 2000).) Dennett's appeal to Church's thesis thus flies in the face of the mathematical facts: some varieties of information processing exceed standard computation (or Turing-computation). Church's thesis, or more precisely, the Church-Turing thesis, is the view that a function f is effectively computable if and only if f is Turing-computable (i.e., some Turing machine can compute f). Thus, this thesis has nothing to say about information processing that is more demanding than what a Turing machine can achieve. (Put another way, there is no counter-example to CTT to be automatically found in an information-processing device capable of feats beyond the reach of TMs.) For all philosophy and psychology know, intelligence, even if tied to information processing, exceeds what is Turing-computational or Turing-mechanical.[15] This is especially true because philosophy and psychology, unlike AI, are in no way fundamentally charged with engineering artifacts, which makes the physical realizability of hypercomputation irrelevant from their perspectives. Therefore, contra Dennett, to consider AI as psychology or philosophy is to commit a serious error, precisely because so doing would box these fields into only a speck of the entire space of functions from the natural numbers (including tuples therefrom) to the natural numbers. (Only a tiny portion of the functions in this space are Turing-computable.) AI is without question much, much narrower than this pair of fields. Of course, it's possible that AI could be replaced by a field devoted not to building computational artifacts by writing computer programs and running them on embodied Turing machines. But this new field, by definition, would not be AI. Our exploration of AIMA and other textbooks provide direct empirical confirmation of this.

Third, most AI researchers and developers, in point of fact, are simply concerned with building useful, profitable artifacts, and don’t spend much time reflecting upon the kinds of abstract definitions of intelligence explored in this entry (e.g., What Exactly is AI?).

Though AI isn’t philosophy, there are certainly ways of doing real implementation-focussed AI of the highest caliber that are intimately bound up with philosophy. The best way to demonstrate this is to simply present such research and development, or at least a representative example thereof. The most prominent example in AI today is John Pollock's OSCAR project.

It’s important to note at this juncture that the OSCAR project, and the information processing that underlies it, are without question at once philosophy and technical AI. Given that the work in question has appeared in the pages of Artificial Intelligence, a first-rank journal devoted to that field, and not to philosophy, this is undeniable (see, e.g., Pollock 2001, 1992). This point is important because while it’s certainly appropriate, in the present venue, to emphasize connections between AI and philosophy, some readers may suspect that this emphasis is contrived: they may suspect that the truth of the matter is that page after page of AI journals are filled with narrow, technical content far from philosophy. Many such papers do exist. But we must distinguish between writings designed to present the nature of AI, and its core methods and goals, versus writings designed to present progress on specific technical issues.

Writings in the latter category are more often than not quite narrow, but, as the example of Pollock shows, sometimes these specific issues are inextricably linked to philosophy. And of course Pollock's work as a representative example. One could just as easily have selected work by folks who don't happen to also produce straight philosophy. For example, for an entire book written within the confines of AI and computer science, but which is epistemic logic in action in many ways, suitable for use in seminars on that topic, see (Fagin et al. 2004). (It is hard to find technical work that isn’t bound up with philosophy in some direct way. E.g., AI research on learning is all intimately bound up with philosophical treatments of induction, of how genuinely new concepts not simply defined in terms of prior ones can be learned. One possible partial answer offered by AI is inductive logic programming, discussed in Chapter 19 of AIMA.)

What of writings in the former category? Writings in this category, while by definition in AI venues, not philosophy ones, are nonetheless philosophical. Most textbooks include plenty of material that falls into this latter category, and hence they include discussion of the philosophical nature of AI (e.g., that AI is aimed at building artificial intelligences, and that’s why, after all, it’s called ‘AI’).

Now to OSCAR.

The OSCAR Project

OSCAR, according to Pollock, will eventually be not just an intelligent computer program, but an artificial person. (Lest it be thought that this is spinning Pollock's work in the direction of the stunningly ambitious, note that the subtitle of (Pollock 1995) is “A Blueprint for How to Build a Person,”, and that his prior book (1989) was How to Build a Person.) However, though persons have an array of perceptual powers (effectors that allow them to manipulate their environments, linguistic abilities, etc.) OSCAR, at least in the near term, will not have this breadth. OSCAR’s strong suit is the “intellectual” side of personhood. Pollock thus intends OSCAR to be an “artificial intellect”, or, to use his neologism, an artilect. An artilect is a rational agent; Pollock’s concern is thus with rationality. As to the roles of AI and philosophy addressing this concern, Pollock writes:

The implementability of a theory of rationality is a necessary condition for its correctness. This amounts to saying that philosophy needs AI just as much as AI needs philosophy. A partial test of the correctness of a theory of rationality is that it can form the basis of an autonomous rational agent, and to establish that conclusively, one must actually build an AI system implementing the theory. It behooves philosophers to keep this in mind when constructing their theories, because it takes little reflection to see that many kinds of otherwise popular theories are not implementable. (Pollock 1995: xii)

The distinguishing feature of OSCAR qua artilect, at least so far, is that the system is able to perform sophisticated defeasible reasoning.[16] The study of defeasible reasoning was started by Roderick Chisholm (1957, 1966, 1977) and Pollock (1965, 1967, 1974), long before AI took the project under a different name (nonmonotonic reasoning). Both Chisholm and Pollock, as we noted above, assume that reasoning proceeds by constructing arguments, and Pollock takes reasons to provide the atomic links in arguments. Conclusive reasons are reasons that aren't defeasible; conclusive reasons logically entail their conclusions. On the other hand, prima facie reasons provide support for their conclusions, but can be defeated. Defeaters overthrow or defeat prima facie reasons, and come in two forms: defeaters can provide a reason for denying the conclusion, and they can also attack the connection between the premises and the conclusion. As an example of the latter given by Pollock, consider: The proposition ‘a looks red to me’ is a prima facie reason for an agent to believe ‘a is red’. But if you know as well that a is illuminated by red lights, and that such lights can make things look red when they aren't, the connection is threatened. You don’t have a reason for thinking that it’s not the case that a is red, but the inference in question is shot down: it’s defeated.

We can bring a good deal of this to life, even within our space constraints, by considering how OSCAR supplies a solution to the lottery paradox, which arises as follows. Suppose you hold one ticket tk, for some k ≤ 1000000, in a fair lottery consisting of 1 million tickets, and suppose it is known that one and only one ticket will win. Since the probability is only .000001 of tk’s being drawn, it seems reasonable to believe that tk will not win. (Of course, to make this side of the apparent antinomy more potent, we can stipulate that the lottery has, say, a quadrillion tickets. In this case, it’s probably much more likely that you will be struck dead by a meteorite the next time you leave a building, than it is that you will win. And isn’t it true that you firmly believe, now, that when you walk outside tomorrow you won’t be struck dead in this way? If so, then surely you should believe, of your ticket, that it won’t win!) By the same reasoning it seems that you ought to believe that t1 will not win, that t2 will not win, ..., that t1000000 will not win (where you skip over k). Therefore it is reasonable to believe

¬∃ti (ti will win).
But on the other hand we know that
ti (ti will win).
We thus find ourselves caught in an outright contradiction.

What is Pollock's diagnosis of this paradox? In a nushell, it’s this: Since as rational beings we ought never to believe both p and ¬p, and since if we know anything we know that a certain ticket will win, we must conclude that it’s not the case that we ought to believe that tk will not win. We must replace this belief with a defeasible belief based on that fact that we have but a prima facie reason for believing that tk will not win.

Our situation can be described more carefully in Pollockian terms, which indicates that this situation is a case of collective defeat. Suppose that we are warranted in believing r and that we have equally good prima facie reasons for p1, ..., pn, where {p1, ..., pn} r is inconsistent but no proper subset of p1, ..., pn is inconsistent with r. Then, for every pi:

{r p1 ... pi-1 pi+1 ... pn} ¬pi

In this case we have equally strong support for each pi and each ¬pi, so they collectively defeat one another. Here is how Pollock at one point expresses the principle of collective defeat, operative in this case:

If we are warranted in believing r and we have equally good independent prima facie reasons for each member of a minimal set of propositions deductively inconsistent with r, and none of these prima facie reasons is defeated in any other way, then none of the propositions in the set is warranted on the basis of these prima facie reasons. (1995, 62)

Recall Pollock's insistence upon the implementability of theories of rationality. The neat thing is that OSCAR allows us to implement collective defeat -- indeed, though we will not go that far here, we can even implement in OSCAR the solution to the lottery paradox (and the paradox of the preface as well, as Pollock (1995) shows). These particular implementations are too detailed and technical to present in the present venue. But we can show here the use of OSCAR to solve some simple problems in deductive logic that philosophers give students in introductory philosophy and logic. Let’s start by giving OSCAR this problem:

{(p → q), (q s) → r} p → r

The reader will be spared the details concerning how this query is encoded and supplied to OSCAR, and so on. We move directly to what OSCAR instantly returns in response to the query:

This is an undefeated argument of strength 1.0 for: (p -> r)
which is of ultimate interest. 2. ((q v s) -> r) GIVEN 1. (p -> q) GIVEN 6. (q -> r) disj-antecedent-simp from { 2 } |---------------------------------------------------------- | Suppose: { p } |---------------------------------------------------------- | 3. p SUPPOSITION | 5. q modus-ponens1 from { 1 , 3 } | 8. r modus-ponens1 from { 6 , 5 } 9. (p -> r) CONDITIONALIZATION from { 8 } ===========================================================================

Notice how nice this output is: it conforms to the kind of natural deduction routinely taught to students in elementary philosophy and logic. For example, it would be easy enough to have OSCAR solve the bulk of the exercises supplied in Language, Proof, and Logic (Barwise & Etchemendy 1999), which teaches the system , so named because it's a Fitch-style natural deduction system. Of course, some of these exercises involve quantifiers. Here is a query that corresponds to one of the hardest problems in (Barwise & Etchemendy 1994), which teaches a natural deduction system very similar to :

∃x (B(x) → ∀y B(y))

Using quantifier shift, OSCAR produces the following as a solution, in less than a tenth of a second.

This is a deductive argument for:
      (some x)(( Bird x) -> (all y)( Bird y))
 which is of ultimate interest.

        | Suppose:  { ~(some x)(( Bird x) -> (all y)( Bird y)) }
        | 2.  ~(some x)(( Bird x) -> (all y)( Bird y))     REDUCTIO-SUPPOSITION
        | 5.  (all x)~(( Bird x) -> (all y)( Bird y))     neg-eg from { 2 }
        | 6.  ~(( Bird x3) -> (all y)( Bird y))     UI from { 5 }
        | 7.  ( Bird x3)     neg-condit from { 6 }
        | 8.  ~(all y)( Bird y)     neg-condit from { 6 }
        | 9.  (some y)~( Bird y)     neg-ug from { 8 }
        | 10.  ~( Bird @y5)     EI from { 9 }
 11.  (some x)(( Bird x) -> (all y)( Bird y))     REDUCTIO from { 10 , 7 }


How good is OSCAR, matched against the ambitious goal of literally building a person? Here only two points will be made; both should be uncontroversial.

First, certainly expressivity is a problem for OSCAR. Can OSCAR handle reasoning that seems to require intensional operators? There does not appear to be any such work with the system. Perhaps Pollock has such work in mind for the future, but at present, OSCAR is merely at the level of elementary extensional logic. (Of course, the technique of encoding down, encapsulated above, could be used in conjunction with OSCAR.)

A second, and not unrelated, concern, is that while Pollock’s method of finding rigorous innovation by striving to build a system capable of handling paradoxes is fruitful (and doubtless especially congenial to philosophers), the fact is that he has so far based his work on simple paradoxes and puzzles. Can OSCAR handle more difficult paradoxes? It would be nice, for example, if OSCAR could automatically find a solution to Newcomb's Paradox (NP) (Nozick 1970). As some readers will know, this paradox involves constructions (e.g., backtracking conditionals) quite beyond first-order logic. In addition, there are now infinitary paradoxes in the literature, and it's hard to see how OSCAR could be used to even represent the key parts of these paradoxes. Since some humans dissect and discuss NP and infinitary paradoxes (etc.) in connection with various more expressive logics, humans would appear to be functioning as artilects beyond the reach of at least the current version of OSCAR.

On the other hand, part of the reason for including coverage herein of OSCAR-based AI work is that such a direction, with roots in argument-based epistemology running back to the 1950’s, the same time modern AI started up (recall that the 1956 Dartmouth conference was held in 1956), promises to continue to provide a fruitful approach into the future. Evidence for this can be found in the form of Pollock's (2006) Thinking about Acting: Logical Foundations for Rational Decision Making, a philosophically sophisticated AI-relevant investigation of rational decision making for realistically resource-bounded agents.

Philosophy of Artificial Intelligence

“Strong” versus “Weak” AI

Recall that we earlier discussed proposed definitions of AI, and recall specifically that these proposals were couched in terms of the goals of the field. We can follow this pattern here: We can distinguish between “Strong” and “Weak” AI by taking note of the different goals that these two versions of AI strive to reach. “Strong” AI seeks to create artificial persons: machines that have all the mental powers we have, including phenomenal consciousness. “Weak” AI, on the other hand, seeks to build information-processing machines that appear to have the full mental repertoire of human persons (Searle 1997). “Weak” AI can also be defined as the form of AI that aims at a system able to pass not just the Turing Test (again, abbreviated as TT), but the Total Turing Test (Harnad 1991). In TTT, a machine must muster more than linguistic indistinguishability: it must pass for a human in all behaviors -- throwing a baseball, eating, teaching a class, etc.

It would certainly seem to be exceedingly difficult for philosophers to overthrow “Weak” AI (Bringsjord and Xiao 2000). After all, what philosophical reason stands in the way of AI producing artifacts that appear to be animals or even humans? However, some philosophers have aimed to do in “Strong” AI, and we turn now to the most prominent case in point.

The Chinese Room Argument Against “Strong AI”

Without question, the most famous argument in the philosophy of AI is John Searle's (1980) Chinese Room Argument (CRA), designed to overthrow “Strong” AI. In light of the existence of David Cole's SEP entry on CRA, a quick summary here is all that's needed -- that and a “report from the trenches” as to how AI practioners regard the argument. Readers wanting to further study CRA will find an excellent next step in (Bishop & Preston 2002).

CRA is based on a thought-experiment in which Searle himself stars. He is inside a room; outside the room are native Chinese speakers who don't know that Searle is inside it. Searle-in-the-box, like Searle-in-real-life, doesn't know any Chinese, but is fluent in English. The Chinese speakers send cards into the room through a slot; on these cards are written questions in Chinese. The box, courtesy of Searle's secret work therein, returns cards to the native Chinese speakers as output. Searle's output is produced by consulting a rulebook: this book is a lookup table that tells him what Chinese to produce based on what is sent in. To Searle, the Chinese is all just a bunch of -- to use Searle's language -- squiggle-squoggles. The following schematic picture sums up the situation. The labels should be obvious. O denotes the outside observers, in this case the Chinese speakers. Input is denoted by i and output by o. As you can see, there is an icon for the rulebook, and Searle himself is denoted by P.

The Chinese Room, Schematic View

Now, what is the argument based on this thought-experiment? Even if you've never heard of CRA before, you doubtless can see the basic idea: that Searle (in the box) is supposed to be everything a computer can be, and because he doesn't understand Chinese, no computer could have such understanding. Searle is mindlessly moving squiggle-squoggles around, and (according to the argument) that's all computers do, fundamentally.[17]

Where does CRA stand today? As I’ve already indicated, the argument would still seem to be alive and well; witness (Bishop & Preston 2002). However, there is little doubt that at least among AI practitioners, CRA is generally rejected. (This is of course thoroughly unsurprising.) Among these practicioners, the philosopher who has offered the most formidable response out of AI itself is Rapaport (1988), who argues that while AI systems are indeed syntactic, the right syntax can constitute semantics. It should be said that a common attitude among proponents of “Strong” AI is that CRA is not only unsound, but silly, based as it is on a fanciful story (CR) far removed from the practice of AI -- practice which is year by year moving ineluctably toward sophisticated robots that will once and for all silence CRA and its proponents. For example, John Pollock (as we've noted, philosopher and practitioner of AI) writes:

Once [my intelligent system] OSCAR is fully functional, the argument from analogy will lead us inexorably to attribute thoughts and feelings to OSCAR with precisely the same credentials with which we attribute them to human beings. Philosophical arguments to the contrary will be passé. (Pollock 1995)

Readers may wonder if there are philosophical debates that AI researchers engage in, in the course of working in their field (as opposed to when they might attend a philosophy conference). Surely, AI researchers have philosophical discussions amongst themselves, right?

Generally, one finds that AI researchers do discuss among themselves topics in philosophy of AI, and these topics are usually the very same ones that occupy philosophers of AI. However, the attitude reflected in the quote from Pollock immediately above is by far the dominant one. That is, in general, the attitude of AI researchers is that philosophizing is sometimes fun, but the upward march of AI engineering cannot be stopped, will not fail, and will eventually render such philosophizing otiose.

We will return to the issue of the future of AI in the final section of this entry.

The Gödelian Argument Against “Strong AI”

Four decades ago, J.R. Lucas (1964) argued that Gödel's first incompleteness theorem entails that no machine can ever reach human-level intelligence. His argument has not proved to be compelling, but Lucas initiated a debate that has produced more formidable arguments. One of Lucas' indefatigable defenders is the physicist Roger Penrose, whose first attempt to vindicate Lucas was a Gödelian attack on “Strong” AI articulated in his The Emperor's New Mind (1989). This first attempt fell short, and Penrose published a more elaborate and more fastidious Gödelian case, expressed in Chapters 2 and 3 of his Shadows of the Mind (1994).

In light of the fact that a separate entry on the Gödelian Argument is forthcoming in SEP, a full review here is not needed. Instead, readers will be given a decent sense of the argument by turning to an online paper in which Penrose, writing in response to critics (e.g., the philosopher David Chalmers, the logician Solomon Feferman, and the computer scientist Drew McDermott) of his Shadows of the Mind, distills the argument to a couple of paragraphs.[18] Indeed, in this paper Penrose gives what he takes to be the perfected version of the core Gödelian case given in SOTM. Here is this version, verbatim:

We try to suppose that the totality of methods of (unassailable) mathematical reasoning that are in principle humanly accessible can be encapsulated in some (not necessarily computational) sound formal system F. A human mathematician, if presented with F, could argue as follows (bearing in mind that the phrase “I am F” is merely a shorthand for “F encapsulates all the humanly accessible methods of mathematical proof”):
(A) “Though I don’t know that I necessarily am F, I conclude that if I were, then the system F would have to be sound and, more to the point, F′ would have to be sound, where F′ is F supplemented by the further assertion “I am F.” I perceive that it follows from the assumption that I am F that the Gödel statement G(F′) would have to be true and, furthermore, that it would not be a consequence of F'. But I have just perceived that “If I happened to be F, then G(F′) would have to be true,” and perceptions of this nature would be precisely what F′ is supposed to achieve. Since I am therefore capable of perceiving something beyond the powers of F′, I deduce that I cannot be F after all. Moreover, this applies to any other (Gödelizable) system, in place of F.” (Penrose 1995)
Does this argument succeed? A firm answer to this question is not appropriate to seek in the present entry. Interested readers are encouraged to consult four full-scale treatments of the argument (Hayes et. al 1998, Bringsjord and Xiao 2000, Shapiro 2003, Bowie 1982), and the entry in this encyclopedia when it is available.

Additional Topics and Readings in Philosophy of AI

In addition to the Gödelian and Searlean arguments covered briefly above, a third attack on “Strong” AI (of the symbolic variety) has been widely discussed, namely, one given by the philosopher Hubert Dreyfus (1972, 1992), some incarnations of which have been co-articulated with his brother, Stuart Dreyfus (1987), a computer scientist. Put crudely, the core idea in this attack is that human expertise is not based on the explicit, disembodied, mechanical manipulation of symbolic information (such as formulae in some logic, or probabilities in some Bayesian network), and that AI's efforts to build machines with such expertise are doomed if based on the symbolic paradigm. The genesis of the Dreyfusian attack was a belief that the critique of (if you will) symbol-based philosophy (e.g., philosophy in the logic-based, rationalist tradition, as opposed to what is called the Continental tradition) from such thinkers as Heidegger and Merleau-Ponty could be made against the rationalist tradition in AI. After further reading and study of Dreyfus' writings, readers much judge whether this critique is compelling, in an information-driven world increasingly managed by intelligent agents that carry out symbolic reasoning (albeit not even close to the human level).

For readers interested in exploring philosophy of AI beyond what Jim Moor (in a recent address -- “The Next Fifty Years of AI: Future Scientific Research vs. Past Philosophical Criticisms” -- as the 2006 Barwise Award winner at the annual eastern American Philosophical Association meeting) has called the “the big three” criticisms of AI, there is no shortage of additional material, much of it available on the Web. The last chapter of AIMA provides a compressed overview of some additional arguments against “Strong” AI, and is in general not a bad next step. Needless to say, Philosophy of AI today involves much more than the three well-known arguments discussed above, and, inevitably, Philosophy of AI tomorrow will include new debates and problems we can’t see now. Because machines, inevitably, will get smarter and smarter (regardless of just how smart they get), Philosophy of AI, pure and simple, is a growth industry. With every human activity that machines match, the “big” questions will only attract more attention.

The Future

If past predictions are any indication, the only thing we know today about tomorrow's science and technology is that it will be radically different than whatever we predict it will be like. Arguably, in the case of AI, we may also specifically know today that progress will be much slower than what most expect. After all, at the 1956 kickoff conference (discussed at the start of this entry), Herb Simon predicted that thinking machines able to match the human mind were “just around the corner” (for the relevant quotes and informative discussion, see the first chapter of AIMA). As it turned out, the new century would arrive without a single machine able to converse at even the toddler level. (Recall that when it comes to the building of machines capable of displaying human-level intelligence, Descartes, not Turing, seems today to be the better prophet.) Nonetheless, astonishing though it may be, people do continue today to issue incredibly optimistic predictions regarding the progress of AI. For example, Hans Moravec (1999), in his Robot: Mere Machine to Transcendent Mind, informs us that because the speed of computer hardware doubles every 18 months (in accordance with Moore's Law, which has apparently held in the past and shows no sign of failing), “fourth generation” robots will soon enough exceed humans in all respects, from running companies to writing novels. These robots, so the story goes, will evolve to such lofty cognitive heights that we will stand to them as single-cell organisms stand to us today.

Moravec is by no means singularly Pollyannaish: Many others in AI predict the same sensational future unfolding on about the same rapid schedule. In fact, at the aforementioned AI@50 conference, Jim Moor posed the question “Will human-level AI be achieved within the next 50 years?” to five thinkers who attended the original 1956 conference: John McCarthy, Marvin Minsky, Oliver Selfridge, Ray Solomonoff, and Trenchard Moore. McCarthy and Minsky gave firm, unhesitating affirmatives, and Solomonoff seemed to suggest that AI provided the one ray of hope in the face of fact that our species seems bent on destroying itself. (Selfridge's reply was a bit cryptic. Moore returned a firm, unambiguous negative, and declared that once his computer is smart enough to interact with him conversationally about mathematical problems, he might take this whole enterprise more seriously.) It is left to the reader to judge the accuracy of such risky predictions as have been given by Moravec, McCarthy, and Minsky.[19]

There are some things we can safely say about tomorrow. Certainly, barring some cataclysmic events (nuclear or biological warfare, global economic depression, a meteorite smashing into Earth, etc.), we now know that AI will succeed in producing artificial animals. Since even some natural animals (mules, e.g.) can be easily trained to work for humans, it stands to reason that artificial animals, designed from scratch with our purposes in mind, will be deployed to work for us. In fact, many jobs currently done by humans will certainly be done by appropriately programmed artificial animals. To pick an arbitrary example, it is impossible to believe that commercial drivers won't be artificial in the future. Other examples would include: cleaners, mail carriers, clerical workers, military scouts, surgeons, and pilots. (As to cleaners, probably a significant number of readers, at this very moment, have robots from iRobot cleaning the carpets in their homes.) It is hard to see how such jobs are inseparably bound up with the attributes often taken to be at the core of personhood -- attributes that would be the most difficult for AI to replicate.[20]

A specific technical prediction can be safely made regarding logicist AI, the form of AI started by the original 1956 conference[21]. The prediction is this: More and more investment will be made in the direction of marrying symbolic logic with diagrammatic representations. First-rate human reasoners get enormous leverage from pictorial representation schemes, but machines are still pretty much locked into formal languages that are linguistic through and through. Unfortunately, philosophy, logic, and AI lost Jon Barwise prematurely, but, along with his frequent collaborator, John Etchemendy, he laid down the foundation for future attempts at pulling off the marriage (Barwise & Etchemendy 1990, 1995).

As nearly all readers will know, logic, traditionally, is indeed linguistic: to specify a logic is to specify strings in some language, and to write down operations over those strings, where this meta-information is itself written in some language. Visual or diagrammatic logic allows for reasoning over representations that are pictorial, not linguistic. (Logics that present formulas in pictographic form, such as those that go back to Lewis Carroll and Charles Sanders Peirce, don't count. Visual logic would allow a house to be represented as a picture, not represented as a set of formulas which are represented by pictures.)

A key idea in visual logic is that of a -- to use the term preferred by Barwise & Etchemendy (1995) -- homomorphic representation; i.e., a representation that refers in virtue, at least in part, of its size, shape, texture, color, etc. In standard logic, one can use car-22 to refer to your car, but one could also use car22. There is no problem here. But if one has a diagram of your car, and one removes a piece of this diagram, that may cause the representation to fail to refer. For example, if one changes the color of the representation in the case of the visualization, your car may no longer be denoted. You may own a black Jeep SRT, while Jones may own a red one, and if the diagrammatic representation of your Jeep shows the color as red, reference fails. However, changing the font color of car22 to to the same string in a red font in a logic is guaranteed to have no impact at all on the meaning of this name. In sum, the meaning of a visualization or diagram is not preserved when one moves to a corresponding purely symbolic representation of it. This is a brute fact (nicely shown at length, e.g., in Barwise & Etchemendy 1995). Recently DARPA has invested in the development of logics that handle homomorphic representations; e.g., it has sponsored the creation and specification of the Vivid family of visual logics, described in full in a paper forthcoming in Artificial Intelligence journal.

For a non-technical prediction, it seems safe to say that the future Andy Clark (2003) sees is one that will come to pass: Humans will gradually become cyborgs, courtesy of artificial limbs and sense organs, and implants. The main driver of this trend will be that while standalone AIs are often desirable, they are hard to engineer when the desired level of intelligence is high. But to let humans “pilot” less intelligent machines is a good deal easier, and still very attractive for concrete reasons.

Let us conclude a discussion of the future and AI by bringing to your attention the fact that many rather smart and intelligent people see a particularly black future because of AI. The locus classicus here is without question a paper by Bill Joy (2000): “Why The Future Doesn't Need Us.” Joy believes that the human race is doomed, in no small part because it's busy building smart machines. He writes:

The 21st-century technologies -- genetics, nanotechnology, and robotics (GNR) -- are so powerful that they can spawn whole new classes of accidents and abuses. Most dangerously, for the first time, these accidents and abuses are widely within the reach of individuals or small groups. They will not require large facilities or rare raw materials. Knowledge alone will enable the use of them.

Thus we have the possibility not just of weapons of mass destruction but of knowledge-enabled mass destruction (KMD), this destructiveness hugely amplified by the power of self-replication.

I think it is no exaggeration to say we are on the cusp of the further perfection of extreme evil, an evil whose possibility spreads well beyond that which weapons of mass destruction bequeathed to the nation-states, on to a surprising and terrible empowerment of extreme individuals.[22]

Philosophers would be most interested in arguments for this view. What are Joy’s? Well, no small reason for the attention lavished on his paper is that, like Raymond Kurzweil (2000), Joy relies heavily on an argument given by none other than the Unabomber (Theodore Kaczynski):

First let us postulate that the computer scientists succeed in developing intelligent machines that can do all things better than human beings can do them. In that case presumably all work will be done by vast, highly organized systems of machines and no human effort will be necessary. Either of two cases might occur. The machines might be permitted to make all of their own decisions without human oversight, or else human control over the machines might be retained.

If the machines are permitted to make all their own decisions, we can’t make any conjectures as to the results, because it is impossible to guess how such machines might behave. We only point out that the fate of the human race would be at the mercy of the machines. It might be argued that the human race would never be foolish enough to hand over all the power to the machines. But we are suggesting neither that the human race would voluntarily turn power over to the machines nor that the machines would willfully seize power. What we do suggest is that the human race might easily permit itself to drift into a position of such dependence on the machines that it would have no practical choice but to accept all of the machines’ decisions. As society and the problems that face it become more and more complex and machines become more and more intelligent, people will let machines make more of their decisions for them, simply because machine-made decisions will bring better results than man-made ones. Eventually a stage may be reached at which the decisions necessary to keep the system running will be so complex that human beings will be incapable of making them intelligently. At that stage the machines will be in effective control. People won’t be able to just turn the machines off, because they will be so dependent on them that turning them off would amount to suicide.

On the other hand it is possible that human control over the machines may be retained. In that case the average man may have control over certain private machines of his own, such as his car or his personal computer, but control over large systems of machines will be in the hands of a tiny elite -- just as it is today, but with two differences. Due to improved techniques the elite will have greater control over the masses; and because human work will no longer be necessary the masses will be superfluous, a useless burden on the system. If the elite is ruthless they may simply decide to exterminate the mass of humanity. If they are humane they may use propaganda or other psychological or biological techniques to reduce the birth rate until the mass of humanity becomes extinct, leaving the world to the elite. Or, if the elite consists of soft-hearted liberals, they may decide to play the role of good shepherds to the rest of the human race. They will see to it that everyone’s physical needs are satisfied, that all children are raised under psychologically hygienic conditions, that everyone has a wholesome hobby to keep him busy, and that anyone who may become dissatisfied undergoes “treatment” to cure his “problem.” Of course, life will be so purposeless that people will have to be biologically or psychologically engineered either to remove their need for the power process or make them “sublimate” their drive for power into some harmless hobby. These engineered human beings may be happy in such a society, but they will most certainly not be free. They will have been reduced to the status of domestic animals.

This isn’t the place to assess this argument. (Having said that, the pattern pushed by the Unabomber and his supporters certainly appears to be invalid.[23]) Even if the argument is formally invalid, it leaves us with a question -- the basic question about AI and the future: Will AI produce artificial creatures that replicate and exceed human cognition? Put another way: Is what the Unabomber postulates true (as he and Kurzweil and Joy believe), or merely an interesting supposition?

This is a question not just for scientists and engineers; it is also a question for philosophers. This is so for two reasons. One, research and development designed to validate an affirmative answer must include philosophy -- for reasons rooted in earlier parts of this entry. (E.g., philosophy is the place to turn to for formalisms to model human propositional attitudes in machine terms.) Two, philosophers might be able to provide arguments that answer the question now, definitively. If a version of either of the three arguments against “Strong” AI alluded to above (Searle's CRA; the Gödelian attack; the Dreyfus argument) are sound, then of course AI will not manage to produce machines having the mental powers of persons. It is interesting to note that the genesis of Joy's paper was an informal conversation with John Searle and Raymond Kurzweil. According to Joy, Searle didn’t think there was much to worry about, since he he was (and is) quite confident that tomorrow’s robots can't be conscious.[24]



Thanks are due to Peter Norvig and Prentice-Hall for allowing figures from AIMA to be used in this entry. Thanks are due as well to the many first-rate (human) minds who have read earlier drafts of this entry, and provided helpful feedback.

Other Internet Resources

Related Entries

logic and artificial intelligence cognitive science | connectionism | computability and complexity | computing -- modern history of | language of thought hypothesis | logic and artificial intelligence | mind -- computational theory | entryn

Copyright © 2006 by
Selmer Bringsjord

A | B | C | D | E | F | G | H | I | J | K | L | M | N | O | P | Q | R | S | T | U | V | W | X | Y | Z
First published: (This is a pre-published version.)
Content last modified: 8.15.06