Understanding Computers and Cognition: A New Foundation for Design
Source: https://archive.org/details/understandingcom00wino ↗
Winograd built SHRDLU, one of the most celebrated early natural-language AI systems, and then wrote this book to explain why the entire approach was wrong.
Drawing on Heidegger's phenomenology, Maturana's biology of cognition, and Austin's speech-act theory, Winograd and Flores argued that computers cannot understand language because understanding is not computation — it is a form of being in the world.
The book directly influenced the design of Lotus Notes and the concept of workflow software as coordination tools rather than knowledge processors.
For product people working with AI today, the arguments remain disturbingly relevant: the gap between pattern matching and genuine understanding that Winograd identified in 1986 has not been closed, only papered over with more data.
This is the rare work where a technologist dismantles his own achievement to build something more honest.
Central argument
Winograd and Flores argue that the rationalist tradition underlying AI and software design fundamentally misrepresents what human understanding is. Drawing on Heidegger's concept of 'being-in-the-world', Maturana's biology of cognition, and Austin's speech-act theory, they contend that understanding is not the manipulation of symbolic representations but an embodied, pre-reflective engagement with a world that can never be fully formalised. The practical consequence is that software should be designed not to simulate comprehension but to structure human coordination — commitments, requests, and promises between people — which is why the book directly shaped workflow tools like Lotus Notes.
Critique
The book's dependence on Heideggerian phenomenology, while philosophically serious, creates a tension it never fully resolves: if human understanding is fundamentally pre-reflective and resistant to formalisation, it is unclear what design principles can legitimately follow from that claim without themselves becoming another rationalist reduction. Winograd and Flores propose speech-act theory as a constructive foundation, but critics have noted that Austin's framework, when embedded in software, risks reifying the very kind of explicit, rule-governed communication the authors argued was derivative of deeper, tacit engagement. The move from critique to design prescription may be less secure than the authors acknowledge.
Why it matters for product
For a CPO building AI-assisted products today, the book's core diagnosis reframes a concrete strategic risk: systems instrumented around output metrics — task completion rates, query resolution, engagement — will consistently be mistaken for understanding what users need, because those metrics measure the surface of interaction, not the breakdown moments where genuine coordination fails. Winograd and Flores' insight that software should structure commitments between people rather than model knowledge suggests a different frame for discovery: instead of asking what information users are seeking, ask what promises and requests between stakeholders are currently informal, ambiguous, or untracked — and design the coordination infrastructure around those. This is precisely the lens that distinguishes workflow and collaboration products that compound in value from AI features that plateau.