What Computers Can't Do: The Limits of Artificial Intelligence
Source: https://archive.org/details/whatcomputerscan00drey ↗
The earliest and most philosophically rigorous critique of symbolic AI — written when the AI community was making promises remarkably similar to today's.
Dreyfus draws on phenomenology (Heidegger, Merleau-Ponty) to argue that human expertise is fundamentally embodied and situational, not rule-based.
The book was mocked at the time and vindicated by the first AI winter.
With LLMs reopening the same questions about the nature and limits of machine intelligence, it is more relevant than ever.
For product people navigating the current AI hype cycle, Dreyfus offers the intellectual tools to distinguish genuine capability from projected expectation.
Central argument
Dreyfus argues that symbolic AI — systems that represent knowledge as explicit rules and logical propositions — cannot replicate human intelligence because genuine expertise is not rule-based but embodied and situational. Drawing on Heidegger and Merleau-Ponty, he contends that human beings understand the world through lived, bodily engagement with context, not through the manipulation of discrete symbols, and that this difference is not a technical gap to be closed but a categorical one. The core thesis is that any system built on the assumption that intelligence can be fully formalized will systematically fail at the open-ended, context-sensitive tasks that define real-world competence.
Critique
Dreyfus's phenomenological framework, while philosophically powerful against classical symbolic AI, sits uneasily with the empirical success of subsymbolic systems — particularly deep learning and LLMs — that do not rely on explicit rule encoding yet exhibit surprisingly fluent contextual behavior. He may have been right about the wrong target: his argument defeats GOFAI but does not obviously apply to statistical systems that learn latent structure from massive datasets without anyone hand-coding representations. A thoughtful reader must ask whether his critique requires updating rather than wholesale vindication, since the question is no longer whether machines can follow rules but whether pattern interpolation at scale constitutes something meaningfully different from — or uncomfortably close to — understanding.
Why it matters for product
For a CPO navigating AI integration decisions, Dreyfus provides a precise diagnostic tool: when a vendor or engineering team claims their model 'understands' user intent, the phenomenological lens asks whether the system can handle genuinely novel, edge-case situations that fall outside its training distribution — which is exactly where product failures tend to cluster. This maps directly to discovery and scoping decisions: features that require contextual judgment in ambiguous, high-stakes moments (medical triage tools, complex onboarding, exception handling) carry a different risk profile than those operating in well-defined, high-frequency domains where statistical approximation is sufficient. Dreyfus also implicitly argues for keeping human expertise in the loop not as a compliance gesture but as a structural necessity, which has concrete implications for how product teams design handoff points between automated systems and human agents.