What Computers Can’t Do

The best-known philosophical critique of early artificial intelligence, which grew out of Dreyfus's 1965 RAND paper 'Alchemy and Artificial Intelligence'. Dreyfus surveys early AI research, whose first successes were followed by stagnation, and argues that the field's optimism rested on four mistaken assumptions. The biological assumption holds that the brain works like a digital computer; the psychological, that the mind processes information by following formal rules; the epistemological, that all knowledge can be formalised; and the ontological, that the world consists of independent, context-free facts. Drawing on Heidegger, Merleau-Ponty and Wittgenstein, he argues that human intelligence depends on an embodied, situated sense of what is relevant, on skilled know-how rather than explicit rules, and on a background of practices that cannot be fully spelled out. Later editions extended the critique, and the 1992 edition, retitled What Computers Still Can't Do, added a new introduction. Many AI researchers rejected it at the time, though several of its points later became widely accepted.

It remains the classic case that intelligence depends on bodies, skills and context rather than rules, and it is newly relevant to debates about what today's AI systems understand.