Hubert L. Dreyfus

Harper Colophon Books

New York

1979

Reviewed by: Jacob Richardson

August, 1998

In What Computers Can’t Do, Hubert Dreyfus argues that computers do not now possess any of the basic elements of intelligence and that the basic structure of a computer will prevent it from becoming intelligent anytime in the future. The book was originally published in 1967. The edition I am reviewing contains a revised introduction by the author written in 1979 in which he claims his arguments are still valid 12 years later. I believe Dreyfus’ arguments are still valid today. Dreyfus criticizes the biological, psychological, and epistemological assumptions of most leading AI researchers. He believes that the brain, for the most part, does not function like a computer. Therefore, attempts to give a computer human-like intelligence can only achieve very limited success because a computer cannot replicate in its own system most of the complex operations of the human brain.

Dreyfus argues that the human brain does not function like a digital computer. Early work in AI envisioned the brain as something like a large computer with many different pathways that carry bits of information. However, according to Dreyfus, recent work in neuropsychology has shown that the brain might receive information from the frequency of pulses passing through a neuron. If this is true, then the brain functions more like an analog computer. He says, " . . . the difference between the ‘strongly interactive’ nature of brain organization and the noninteractive character of machine organization suggests that insofar as arguments from biology are relevant, the evidence is against the possibility of using digital computers to produce intelligence"(162). It seems that many AI researchers assume that the brain functions like a digital computer and that for this reason it should be possible to build a computer sophisticated enough be considered intelligent.

Many AI researchers and psychologists argue that when a human being is behaving intelligently the human being is following heuristic rules similar to those that are used to program computers. The assumption is made that the human being is "an information-processing system functioning like a heuristically programmed digital computer"(164). It is true that the brain processes the inputs it receives and gives them meaning, but it does not follow that the brain is therefore following a rule-based sequence of discrete operations. Dreyfus points out that research in AI is based on the "a priori assumption that the mind must work like a heuristically programmed digital computer"(187). He argues that since there is no conclusive biological or empirical evidence to support this claim, the assumption is false. AI researchers believe it is only a matter of time before they create an intelligent machine because they have uncritically assumed that human intelligence is produced by a brain that functions like an advanced digital computer.

Dreyfus also examines the epistemological assumptions of AI research. AI scientists believe human knowledge can be formalized somehow so that it can be understood and manipulated by a machine. Dreyfus argues that "a timeless, contextless theory of competence cannot be used to reproduce the moment-to-moment involved behavior required for human performance"(190-91). Human beings do not always behave according to formal rules. In AI research, "It is assumed that, in principle at least, human behavior can be represented by a set of independent propositions describing the inputs to the organism, correlated with a set of propositions describing its outputs"(193), says Dreyfus. However, such an assumption ignores the fact that individual facts take on different meanings and levels of importance in different situations and contexts. A set of rules would have to be written to determine what value facts have in different contexts. Each context would have a near infinite number of variables that would have to be handled by the computer. Human beings don’t always process their perceptions as individual bits of data, but rather as information connected to multiple layers of contexts. Dreyfus believes it is not possible for a computer to reduce these multiple layers of meaning into ones and zeros so that the computer could manipulate them intelligently.

Dreyfus’ arguments are very convincing. He ignores some of the recent work in training computers to learn on their own so that they can deal with new and unexpected situations. However, new computer learning techniques do not change the fact that computers must reduce the information they receive to finite length bit-strings. Within restricted domains Dreyfus would agree that computers can be useful tools for human beings and even process some questions posed in natural language. But human intelligence is beyond the grasp of digital computers in his view. Most readers could understand the technical discussions in this book. Dreyfus discusses many advanced philosophical concepts that might be beyond the grasp of the uninitiated. Anyone interested in the philosophical basis of AI research and theory would enjoy this book.