ArticlePDF Available

Artificial intelligence: The very idea: J. Haugeland, (MIT Press, Cambridge, MA, 1985); 287 pp.

Book Review
J. Haugeland,
Artificial Intelligence: The Very Idea
(MIT Press, Cambridge,
MA, 1985); 287 pp.
Reviewed by: Andr6 Vellino
Advanced Computational Methods Center,
University of Georgia, Athens, GA 30602, U.S.A.
John Haugeland's new book is a splendid layperson's introduction to the
foundational problems of artificial intelligence. While its scope is very broad,
ranging from an overview of the history of Western philosophy to an analysis of
Turing's thesis, it nevertheless has a unifying theme: to assess the idea that
underpins the enterprise of AI--the proposition that
Thinking is computing.
True to a philosopher's professional ethics, Haugeland presents his analysis
of this foundational idea as impartially as he can. He vows not to take sides on
whether artificial intelligence is bound to succeed or doomed to failure. On the
contrary, his central thesis, first stated in the introduction and reiterated
throughout the book, is that the foundational proposition of AI is
The possibility of artificial intelligence, the possibility that "machines can
replicate human thinking," is neither a priori absurd nor a priori obvious: it is a
matter of experience. Moreover, there is at present insufficient evidence to
balance the scale one way or the other.
Haugeland does well to frame the AI enterprise in its broad philosophical
context, a task too often ignored in the AI literature. He begins in Chapter 1
with an overview of the history of Western philosophy as it pertains to the idea
that "thinking (intellection) essentially is rational manipulation of mental
symbols (viz., ideas)" (p. 4). Thomas Hobbes, in this account, is singled out as
the grandfather of AI. In Hobbes' philosophy, everything, including human
reasoning, could be explained in terms of matter in motion. Indeed Hobbes
had hoped to do for the science of the mind--reduce it to the science of
motion--what Galileo had done for the science of motion--reduce it to the
science of geometry. The cost of this enterprise, according to Haugeland, was
having to give up on a theory of how thoughts can be meaningful: "he cannot
tell" Haugeland writes "between minds and books" (p. 25). If human beings
are merely pieces of matter in motion then indeed there is no essential
distinction between a mental thought and the written word. But, Haugeland
suggests, there is a distinction which Hobbesian materialism fails to draw,
Artificial Intelligence
29 (1986) 349-353
0004-3702/86/$3.50 © 1986, Elsevier Science Publishers B.V. (North-Holland)
namely that thoughts
meanings whereas books
meanings. The
materialism underlying AI science, on the other hand, appears to account for
the concept of meaning to Haugeland's satisfaction, though the reasons for this
are never terribly clear.
Any account of the rationalist heritage of AI would be incomplete without
reference to (and reverence for) Descartes. Haugeland applauds the French
philosopher's discovery of the distinction between the symbol and what it
symbolizes--a discovery owed to the observation that algebraic relations
among physical properties are expressible in purely geometric terms. In the
same vein, thoughts are, to use Haugeland's anachronistic phase, "symbolic
representations" of objects and their relations (be they physical or mathemati-
cal). Seen in that light Descartes was responsible for discovering the found-
ational concepts of AI.
But Descartes was also influential for inventing the mind/body problem. If,
as he believed, the mind is distinct from the body (indeed, for him they were
different "substances"), an explanation is required as to how they interact--
that when I will my arm to move, it moves. AI's answer to Descartes' dualist
philosophy, says Haugeland, is materialism: the theory that all phenomena
(including mental phenomena) are "made up" of matter. It is worth noting that
Haugeland is not taking sides on the materialism/dualism debate. He simply
makes the case that AI makes sense of the idea that the mind is a machine.
There is, however a difficulty with the naively materialist (or Hobbesian)
reply to Descartes that Haugeland refers to as "the Paradox of Mechanical
Reason"--the problem of how it is possible for a mechanical object to
Mechanized reasoning, as Haugeland explains in detail
in Chapters 2 and 3, is the manipulation of meaningful symbols using a system
of rules. Now if the rule manipulator (i.e. the machine) attaches no meaning to
the symbols it manipulates then we can't really say it is reasoning. On the other
hand if it performed its manipulations according to the meaning of the rules
and symbols it manipulates, it wouldn't be mechanical since "meanings (what-
ever exactly they are) don't exert physical forces" (p. 39). Moreover, even if
meanings are conferred to rules by the machine, the question arises about how
those meanings are themselves "understood" by the machine. We would have
to imagine "homunculi" inside the machine that understand the meanings of
the rules, thereby leading to an infinite regress of circular explanations.
Although the "Paradox of Mechanical Reason" appears to be something of a
straw man, Haugeland's demonstration of how the paradigm of AI can resolve
it merits some attention. In particular, he uses the thesis that "a computer is an
interpreted formal system"
to explain how "meanings" can be embodied in a
machine and how reasoning can be mechanized. A formal system is loosely
defined as a system of rules that define the possible configurations of the tokens
(or "individuals") in the formal system. Games, like chess and tic-tac-toe, for
example, are formal systems that provide rules for manipulating chess pieces or
"O" and "x" symbols respectively.
There are three essential features of formal games: that they "manipulate
tokens" (pieces, symbols), that they are "digital" and that they are "finitely
playable." Games that manipulate tokens are commonplace and Haugeland's
explanation of token-manipulating rules is straightforward: usually the purpose
of a formal game is to arrive at a recognizable configuration of tokens
(checkmate, three "O" or "x"s in a row, etc). The "digital" character of
formal games is less intuitive. A digital operation is usually thought of as
choice function on the discrete states of a system. But Haugeland generalizes
the notion of a digital operation to be "a set of positive and reliable techniques
for producing ('writing') and reidentifying ('reading') tokens or configurations
of tokens from a prespecified set of [token] types." (p. 53) This implies, for
example, that a Shakespeare sonnet is a digital system because its tokens are
alphabetic characters that can be recognized (read) or reproduced (written) in
a positive (unambiguous) and reliable way. A Rembrandt painting, on the
other hand, does not have "digital" properties because of its colors and
textures. Haugeland's reasoning is a bit weak here for he does not explain why
a painting cannot be "digitized," even in principle. This is particularly puzzling
in view of the flexibility of his definition of a "positive" technique for
recognizing tokens. Presumably, the ability to reflect rays of light of exactly the
right frequency at precisely the right intensities should count as a "positive"
techinque for identifying graphics patterns.
The concept of "finite playability" and its relation to algorithms is clearly
explained while avoiding all technicalities. A formal game is finitely playable if
any player can determine, in a finite number of steps, whether a given move is
legal and if the player can produce a legal move when such exists. This
discussion leads to a natural definition of the essence of computer programs--
the algorithm. An algorithm is simply a recipe for playing a finitely playable
formal game. This is followed by a discussion of schedules, conditional branch-
ing, nondeterministic algorithms, and heuristics, all of which are explained with
great ease and intuitive examples.
Haugeland complements this account of automatic formal systems with an
entertaining chapter on the semantics of formal systems, thus returning once
again to the problem of meaning. He draws the usual distinctions between
syntax and semantics, successfully unravels the concept of "interpretation" as it
applies to the axioms of arithmetic and logic, and all the while avoids the use of
any mathematical or logical symbolism. It is also in this chapter that we are
shown how "GOFAI" (good old fashioned artificial intelligence--that is, the
AI enterprise conceived in the Hobbesian spirit) succeeds in solving the
"Paradox of Mechanical Reason" mentioned earlier. The paradox arose, says
Haugeland, because we were not looking at the problematic nesting of rule
manipulators (leading to infinite regress) from the right point of view. Re-
garded as purely syntactic devices, the problem of how each "homumculus"
responsible for rule manipulations understands the "meaning" of rules simply
disappears. Each rule manipulator can be thought of as being made up of
subrule manipulators, each of which has a subsemantics until eventually the
only rule manipulations are "mechanical" (adding binary digits).
While this account of "meaning" is appealing, there is little evidence to
indicate that it can be extended to intentional objects generally. He does not
make the case, for example, that desires or beliefs can be treated in the same
way as rule governed linguistic behavior. One gets the distinct impression that
the problem of meaning has been dismissed rather than solved.
Chapter 4, Computer architecture, is historically informative. Haugeland
covers some familiar material starting with Babbage's analytical engine, and
continuing with Turing machines and universal Turing machines, von Neumann
machines, LISP machines, and production systems. Conspicuously absent from
the discussion, however, arc parallel architectures (vector processors, multiple
processors systems, bypercubes, and the like). Haugeland's apologies in the
acknowledgements do not quite make up for this lacuna, particularly since he
explicitly observes in this chapter that one of the essential differences between
a von Neumann computer and the brain is the parallelism inherent in cerebral
neuron connections.
Chapter 5, Real machines, could have been entitled, What machines can't
do. Here, Haugeland exposes the limitations inherent in some of the ap-
proaches that have been taken in the history of AI research, all the while
reminding the reader that the failure of these approaches do not condemn the
project of AI (or even refute the fundamental proposition that thinking is
computing). The discovery that machine translation, cybernetics, automatic
theorem proving (Newell's general problem solver), micro-worlds (Winograd's
SHRDLU), and expert systems all fail to deliver the promise of AI is not a
sufficient reason to damn the project as unrealistic or even impossible. If
anything, it has shown, according to Haugeland, that specialized systems
implementing brute force methods have limited applicability.
therefore, needs to synthesize the intelligence in Real people, the topic of the
last chapter.
The discussion of Real people is not intended to be an indictment of
GOFAI: on the contrary, it points to many areas that need to be sorted out
before Real AI is possible. Real people, he points out, understand
sational implicatures
(context-dependent presuppositions), they can disam-
biguate according to context, they can manipulate mental images in their
minds, and they have a sense of personal identity. Of these capabilities one
may well ask which is the most important for an AI system. The answer is not
definite, but Haugeland thinks that personal identity or "ego involvement"
ranks among the most important. "Ego involvement" he writes "may be an
integral part of understanding." Having a sense of humor, he suggests, requires
some kind of ego involvement, or at least the capacity to imagine what it is like
to be in someone else's shoes. At any rate, there are global aspects of
intelligent behavior that require the replication of human capacities and that
GOFAI hasn't yet succeeded in even mimicing.
The main thesis of the book is undeniably sound. The very idea of AI may or
may not be successful: we can only find out by trying. Haugeland's explanation
of why this is so is usually informative, even if sometimes cursory: there are
many details which beg for further analysis. But the net result is convincing:
the case is clearly made that AI is possible but it is equally clear that the proof
of the pudding is in the eating.
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.