What Computers Still Can't Do: A Critique of Artificial Reason
Source: https://mitpress.mit.edu/9780262540674/what-computers-still-cant-do/ ↗
The 1992 update to Dreyfus's 1972 original, written twenty years later with the critique deepened rather than softened.
Dreyfus adds new introductions addressing connectionism, neural networks, and the failures of expert systems, while the core phenomenological argument remains intact: human intelligence is embodied, situated, and fundamentally unlike formal symbol manipulation.
The book documents how AI's grand promises collapsed into the first winter — and how the field responded by shifting vocabulary rather than assumptions.
With large language models reopening the same questions about machine understanding, the sequel is as relevant as the first edition.
Dreyfus's philosophical precision makes it an indispensable counterweight to cycles of hype.
Central argument
Dreyfus argues that classical AI fails because it assumes human intelligence is reducible to formal symbol manipulation — a set of explicit rules that can be encoded and executed by a machine. Drawing on Heidegger and Merleau-Ponty, he contends that human cognition is irreducibly embodied and situated: we navigate the world through practical, context-sensitive engagement that cannot be decomposed into discrete representations. The 1992 update extends this argument to connectionism and neural networks, showing that switching from rule-based to statistical architectures didn't resolve the deeper philosophical problem — that meaning is not a property of symbols or weights, but of beings embedded in a world.
Critique
Dreyfus's phenomenological framework is powerful as a critique but underdeveloped as a research program — he demonstrates what AI cannot do without offering a tractable account of what an embodied, situated machine would actually look like. More pointedly, the rapid capability gains of large language models since 2020 create genuine pressure on the claim that formal systems cannot approximate contextual, open-ended language use, even if the philosophical question of understanding versus performance remains unresolved. A thoughtful reader might accept the argument's premises and still conclude that functional adequacy at scale matters more for practical purposes than whether the underlying mechanism resembles human cognition.
Why it matters for product
Product leaders routinely over-invest in AI features on the assumption that more data and better models will eventually close the gap with human judgment — Dreyfus's argument gives a principled reason to treat that assumption with suspicion rather than as a roadmap. Concretely, it supports designing workflows where machine outputs are inputs to human deliberation rather than substitutes for it, which has direct consequences for how discovery teams structure AI-assisted research or how product orgs define the scope of automated decision-making in high-stakes contexts. It also sharpens the diagnostic lens when evaluating vendor claims: the pattern Dreyfus identified — shifting vocabulary without changing underlying assumptions — is exactly what happens in enterprise AI sales cycles today.