For several decades, a sometimes heated debate has taken place among the experts concerning the role and ultimate usefulness of artificial intelligence (AI) systems. In the early years, sweeping predictions were made of the promise of AI and the likelihood that AI systems might replace and outperform human intelligence in many spheres of activity. In fact, impressive expert systems have been designed for many applications areas, and automatic systems have excelled in formal environments, such as many game-playing situations. Unfortunately, it has not been possible up to now to find generalizing principles that would make any of the special-purpose AI solutions applicable in the open-ended, unrestricted environments in which ordinary human beings operate so effectively.

Over the years, the number of believers in general-purpose AI has become smaller, but the acrimony of the debate has not diminished. When Maurice Wilkes suggested in a recent opinion piece [1] that “perhaps the time has come to face the possibility that [machine intelligence in the sense that Turing had envisioned] will never be realized,” he found himself called “lamentably ignorant” [2] and told to take some courses in AI [3].

Dreyfus, who is now submitting the third edition of his well-known What computers can’t do, has been a vocal critic of AI activities for nearly 30 years. He has never wavered in his view that the usefulness of AI would be confined to special-purpose tasks in restricted environments where the context and operations lend themselves to formalized treatment. In introducing this third edition, he states:

The book now offers not a controversial position in an angry debate, but a view of a bygone period of history. For now that the 20th century is drawing to a close, it is becoming clear that one of the great dreams of the century is ending too.… After 50 years of effort, it is now clear to all but a few diehards that [the] attempt to produce general intelligence has failed (p. ix).

When the first edition of What computers can’t do [4] appeared in 1972, Eric Weiss said that it was a “brilliant, illuminating, stimulating, provocative, valuable, and important” effort (see <CR> 13, 7, July 1972, Rev. 23,463). Predictably, this opinion was not shared within the artificial intelligence community. In another review, Bruce Buchanan argued that Dreyfus’s work was nonscientific and permeated by hostility toward computer scientists. Buchanan complained in particular that Dreyfus’s critique was based on early AI developments conducted prior to 1967, and claimed that the analysis would have been much different if more recent efforts had been considered:

It is dishonest to entitle the book a “critique” of AI when it dwells on the failure of early language translation programs (based primarily on syntactic analysis) without analyzing the recent work on understanding natural language (based on syntax, semantics, and context). (See <CR>, 14, 1, Jan. 1973, Rev. 24,357.)

Dreyfus provided the requested updated analysis in a second edition of his book, published in 1979 [5]. That edition left the basic text substantially intact but added a new 60-page introduction to discuss more recent efforts under the titles “Manipulating Micro-worlds (1967-1972)” and “Facing the Problems of Knowledge Representation (1972-1979).” Contrary to Buchanan’s expectation, the updated analysis did not lead to new conclusions. In fact, general-purpose language understanding programs are as far removed from practical realization in 1993 as they were in 1967, precisely because no viable methods have been found for supplying the needed semantics and context in open-ended situations. Dreyfus states in the 1979 introduction, “AI research has then passed from stagnation to crisis in the decade since I conducted my research for this book” ([5], p. 3).

The current edition once again leaves the original 1972 text basically unchanged, but provides another new introduction of some 40 pages, plus a slightly revised title. In this most recent introduction, Dreyfus once again summarizes the arguments why the representational theories of knowledge used in “good old-fashioned AI” (GOFAI) will not lead to general solutions. In GOFAI, the environment is represented by a limited set of facts and relationships, plus rules designed to manipulate the available knowledge. The successes that derive from such an approach are ad hoc, however, and based on special-purpose heuristics that are not generalizable. To succeed with a GOFAI approach, Dreyfus argues that a symbolic representation would have to be created that encompassed all of human experience and background. A computer might then experience what is relevant at each stage, much the way human beings do. He feels, however, that the task of “making an inarticulate, preconceptual background understanding explicit in a symbolic representation [is] hopeless” (p. xii). His general conclusion so far as fact-and rule-based AI is concerned is stated in the following terms: “It seems highly likely that the rationalist dream of representationalist AI will be over by the end of the century” (p. xxx).

In the new introduction, Dreyfus does consider an alternative to the conventional AI approach, namely the recent connectionist efforts, where a network of cells is used to transform particular inputs into desired outputs. He finds such an approach attractive, because the preconstructed input knowledge needed in conventional approaches can be replaced by training sessions that teach the network to respond in suitable ways when appropriate training data are submitted as input.

The connectionist approach raises problems of its own, however, ranging from the difficulty of providing meaningful training data to the problem of producing human-like, rather than uninteresting, responses and generalizations. Overall, the author’s assessment of the neural-network approach is not optimistic. “Neural-network researchers with their occasional ad-hoc success but no principled way to generalize seem to be at the stage of GOFAI researchers when I wrote about them in the 1960s” (p. xxxvii). Thus, in 1992, the author stands by his conclusions of 25 years ago--that the dream of general-purpose AI is just about over, and that this conclusion is likely to stand the test of time for the foreseeable future.

Readers who know the 1972 [4] or 1979 [5] versions of this book need not acquire the new edition. For all others, this text remains highly stimulating and provocative, and it continues to be recommended to anyone concerned with the larger problem of artificial intelligence.