Blog · Review Essay · Last reviewed June 25, 2026

Understanding Computers and Cognition and the Action Behind the Interface

Terry Winograd and Fernando Flores's Understanding Computers and Cognition: A New Foundation for Design is one of the most useful old books for the current AI moment because it refuses the cleanest fantasy in computing: that meaning can be solved by better internal representation alone. Its target is not merely 1980s symbolic AI. It is the broader habit of treating cognition as a formal structure inside a machine, detached from bodies, practices, commitments, institutions, and the situations in which people actually act.

For this review, situated action means that an output gains meaning inside a practice: who asked, what role they hold, what evidence is visible, what authority the system has, what commitment the language creates, and what repair path exists when the action breaks down.

The practical artifact is an action-setting file: define the speech act, user role, authority boundary, source record, permitted tool action, human acceptance point, appeal path, and breakdown drill before treating a generated output as work.

The Book

Understanding Computers and Cognition was published by Ablex in 1986 and issued as an Addison-Wesley paperback in 1987. Google Books lists the Addison-Wesley edition at 207 pages with ISBN 0201112973, while Winograd's own publication list records the Ablex edition at 220 pages and notes later translations in French, German, Italian, Spanish, and Japanese. His CV also says the book was named the best information science book of 1987 by the American Society for Information Science.

The book sits at a hinge in AI history. Winograd had built SHRDLU, one of the canonical early natural-language AI systems, then became a central figure in human-computer interaction and design. Flores brought cybernetics, organizational design, political experience, and a theory of action through language. Together they produced a book that is partly AI critique, partly philosophy of cognition, partly design theory, and partly a proposal for thinking about computers as participants in human work rather than as isolated artificial minds.

That mixture explains why the book can feel difficult and fresh at once. It draws on phenomenology, hermeneutics, biology of cognition, speech-act theory, system design, management, and cooperative work. It is not a manual for building today's AI systems. It is a way to notice what those systems leave out when they reduce human meaning to text, task state, user profile, objective function, database field, or prompt context.

The AI It Argues Against

The book's main opponent is the rationalist image of cognition: the idea that intelligence consists mainly of representing facts about the world, manipulating those representations by rules, and producing correct inferences. That image was not invented by AI, but AI made it operational. If language is a mapping from sentences to facts, and thought is formal manipulation, then a computer looks like the natural home of intelligence.

Winograd and Flores do not deny that formal systems are powerful. They deny that formalization is the whole story of understanding. People do not encounter the world as detached proposition processors. They are already involved in projects, habits, tools, roles, histories, expectations, and social practices. Meaning shows up inside use. A request, a promise, an apology, a diagnosis, a ticket, an alert, a legal notice, and a medical instruction are not just strings with semantic content. They change obligations, authorize action, create risk, and reorganize attention.

This is why the book remains a useful antidote to model-centered AI discourse. A model can manipulate language with extraordinary fluency and still leave the harder question open: what kind of action is this output entering, and who is committed by it? The danger is not that formal systems are useless. It is that institutions mistake a formal handle on a situation for the situation itself.

After the Blocks World

Winograd's 1980 essay "What Does It Mean to Understand Language?" is the best companion to the book because it makes the trajectory visible. SHRDLU worked in a simulated blocks world where language, objects, actions, and feedback were tightly constrained. It could answer questions, follow commands, track some conversational context, and manipulate toy objects on a screen. It was impressive because the world was narrow enough for representation and action to line up.

The lesson is not that narrow worlds are fake. All useful systems narrow the world. The lesson is that the narrowing must remain visible. SHRDLU did not discover the general structure of human understanding. It demonstrated how much understanding can appear when a domain has been carefully built so that symbols, operations, and consequences fit together.

Current AI systems are less brittle in language and vastly broader in surface competence, but they still depend on constructed worlds: training corpora, tool APIs, retrieval indexes, product policies, benchmark tasks, interface scaffolds, enterprise permissions, and hidden conventions about what counts as a good answer. The new blocks world is not a tabletop. It is the whole stack of captured text, institutional workflow, ranking systems, and tool permissions that makes an answer look situated.

Breakdown as the Design Moment

One of the book's strongest design ideas is breakdown. Most of the time, a tool recedes into activity. The keyboard, form, calendar, terminal, spreadsheet, dashboard, ticket queue, and chatbot window do not appear as objects of attention while they work. They appear when something goes wrong: the form has no field for the real case, the interface routes the user to the wrong department, the summary omits the decisive fact, the agent acts with the wrong permission, the model is fluent but off-target.

Breakdown is not merely failure. It is the moment when the hidden structure of a system becomes available for inspection. A good design makes breakdowns recoverable, accountable, and informative. A bad design hides them, smooths them over, or converts them into user blame.

This matters sharply for AI interfaces. The more fluent a system becomes, the easier it is to miss the point at which the user should stop and inspect. An AI assistant can make a bad assumption sound continuous with the user's intention. A coding agent can turn a mistaken model of the repository into a patch. A customer-service bot can turn a policy gap into a polished refusal. A medical scribe can turn an uncertain conversation into official record. The design question is not only how to prevent error. It is how to make the right breakdowns visible before they harden into action.

Language as Action

The language/action perspective is the book's most practical inheritance. In ordinary software thinking, language often appears as information: a message carries content from sender to receiver. Winograd and Flores push the reader toward a different view. Language is something people do. It requests, offers, promises, declines, commits, authorizes, delegates, escalates, cancels, records, and repairs.

That makes the book newly relevant to large language models. Generative AI systems do not merely emit text. They increasingly draft emails, open issues, summarize meetings, prepare reports, answer customers, file tickets, call tools, update records, and coordinate other systems. In those settings, words become institutional action. A generated message can create a commitment even if nobody intended the model itself to be responsible.

The key distinction is between text that describes and text that binds. A summary can bind a patient to a medical history. A Jira ticket can bind an engineer to a priority. A model-written denial can bind a consumer to a dispute process. A procurement note can bind an organization to a vendor interpretation. A chatbot apology can bind a company to a representation. The output matters because it enters a social machinery of commitments.

Institutions and Commitments

Flores's influence shows most clearly where the book turns from cognition to organizations. Work is not just individual problem solving. It is a network of commitments: who asks, who promises, who waits, who checks, who escalates, who closes the loop, and who is accountable when the promise fails. Computer systems can support those commitments, distort them, or make them disappear behind workflow automation.

This is the institutional lesson for AI agents. An agent is not simply a smart tool when it can read private context, call services, change files, contact people, or route decisions. It becomes part of a commitment structure. The important questions become administrative and political: whose authority does it borrow, which promises can it make, what can it change without asking, what evidence does it preserve, what escalation path exists, and who can contest the action afterward?

The book helps separate automation from accountability. A workflow can be efficient while making responsibility less legible. A summary can be useful while hiding who selected the sources. A request can be routed faster while making it harder to know who refused it. A dashboard can coordinate work while converting negotiation into status categories. The design task is to preserve the human commitments that make action answerable.

The Current AI Reading

Read in 2026, Understanding Computers and Cognition is not a prediction that neural networks would fail. It is a warning about a category error: do not confuse operational competence inside a designed environment with full participation in human understanding. Today's models are stronger than the systems the book criticized, but many deployments still rely on the same simplification. They treat enough context as the same thing as situation, enough language as the same thing as meaning, enough tool access as the same thing as agency, and enough user approval as the same thing as responsibility.

The strongest version of the book's critique is not "machines cannot understand." That claim can become sterile because it turns into metaphysics too quickly. The more useful claim is: understanding is not located in the model alone. It lives across people, tools, histories, bodies, records, incentives, norms, and institutions. A system that ignores those relations can still be useful. It should not be mistaken for the whole intelligence of the setting it has entered.

This changes how AI evaluation should work. It is not enough to ask whether the model answered correctly in isolation. Ask what practice the answer enters. Ask whether the user can detect breakdown. Ask whether the system preserves source, uncertainty, and authority boundaries. Ask what commitments the output creates. Ask whether the institution has become more capable of correction or merely faster at producing official-looking action.

As of June 25, 2026, the current standards and legal record pushes in the same direction. NIST's AI Risk Management Framework treats risk management as lifecycle work across design, deployment, use, and evaluation. NIST's AI Agent Standards Initiative frames agents as systems that need standards for identity, authentication, interoperability, security, and evaluation. EU AI Act Articles 13 and 14 require, for high-risk systems, transparency for deployers and effective human oversight during use. None of those sources proves that any product understands its setting. They show that governance is moving toward the action context the book says designers cannot ignore.

Governance and Safety

The governance lesson is that an AI interface should be audited as a commitment machine, not only as an answer machine. If a system drafts, routes, files, summarizes, escalates, calls tools, or updates records, the design must specify what speech acts it is allowed to perform, whose authority it borrows, which commitments require explicit human acceptance, and what evidence survives for appeal or repair.

Current governance language is catching up to that point. EU AI Act Article 14 requires high-risk AI systems to be designed for effective human oversight during use, including the ability to understand capacities and limitations, monitor operation, interpret output, decide not to use it, override it, intervene, or stop the system where appropriate. Article 13 ties high-risk use to transparency and instructions that let deployers interpret and use outputs appropriately. NIST's AI Risk Management Framework adds a lifecycle frame: govern, map, measure, and manage risks across design, development, deployment, use, and evaluation.

For Winograd and Flores, those controls are not after-the-fact compliance furniture. They are part of whether the system can participate responsibly in work. A useful deployment should make likely breakdowns visible, preserve source and uncertainty, log authority-changing actions, separate draft text from binding action, and give affected people a route to correction. A human reviewer with no time, context, or override power is not oversight; it is an interface decoration.

The minimum governance file should connect the AI system inventory, model or system card, audit trail, procurement record, and incident channel. It should state which outputs are drafts, which become records, which tool calls require approval, which users can contest the action, and who can roll back the workflow when the interface has made the wrong commitment easy.

For agents, add a speech-act permission table. The system may be allowed to suggest, summarize, draft, or classify without being allowed to promise, deny, approve, submit, purchase, delete, publish, or change permissions. That table is not bureaucratic excess. It is how the institution keeps language from silently turning into action.

The Commitment Ledger

The practical governance artifact this review takes from Winograd and Flores is a commitment ledger. For any workflow where generated language can become institutional action, the ledger should record the speech act, user role, source record, authority boundary, acceptance point, breakdown signal, repair path, and accountable owner.

The ledger should distinguish draft, advice, recommendation, request, instruction, submission, approval, denial, publication, purchase, deletion, and official record write. These labels matter because identical text can have different force in different settings. "Looks good" in a chat transcript is not the same as approving a benefits denial, closing a clinical task, accepting a contract revision, or instructing an agent to change production code.

For agents, the same ledger should capture agent identity, credential owner, tool scope, approval gate, external write, notification, rollback path, and incident owner. It belongs beside agent observability, tool-use controls, notice and appeal, tool permission, and incident review. A system that cannot produce this record is not merely undocumented; it is asking the institution to let language act without a recoverable social contract.

The safety gain is not paperwork for its own sake. The ledger makes breakdown visible early. It gives reviewers a reason to pause when a system converts uncertain evidence into a confident action, when an interface hides who is committed by an output, or when a tool call borrows human authority without giving that human time, context, and power to refuse.

Where the Book Needs Friction

The book has real limits. Its prose can be abstract, and some of its business-system design examples feel tied to the office-automation debates of its period. Readers looking for a technical treatment of contemporary machine learning will not find one here. Readers looking for a direct account of race, gender, class, disability, labor exploitation, surveillance capitalism, or platform governance will need other books beside it.

There is also a risk in overstating the anti-representational lesson. Formal representations, benchmarks, databases, ontologies, and models are not optional in complex institutions. The right conclusion is not that representation is false. It is that representation is partial, situated, and consequential. A form, model, prompt, schema, or knowledge graph is a social intervention, not a neutral window.

The language/action program also deserves scrutiny. Treating organizational work as structured commitments can clarify accountability, but it can also over-formalize human cooperation. Not every valuable exchange is a request, promise, or closure. Care, hesitation, tacit skill, refusal, ambiguity, and informal repair can be damaged by systems that make every relation explicit enough to manage.

What This Changes

The book changes the center of AI criticism. The important question is not whether a machine has an inner state sufficiently similar to a person's. The practical question is what kind of human situation the machine is joining, how the interface defines that situation, and what forms of action become easier after the machine arrives.

For builders, that means design begins before the prompt and continues after the output. Map the work practice. Name the commitments. Identify the likely breakdowns. Preserve context that cannot fit in the model. Make authority visible. Keep appeal paths human. Do not let a fluent response become the institutional record without a way to inspect how it was produced and what it leaves out.

Understanding Computers and Cognition remains valuable because it moves computation back into the world. It asks designers to stop treating meaning as something sealed inside a representation and start treating it as something that happens when people, machines, language, and institutions act together. That is still the harder problem. It is also the one that matters most when the interface no longer just displays information, but begins to answer, decide, and move.

Source Discipline

This review separates bibliographic evidence from interpretation and governance evidence. Google Books, Internet Archive, Stanford Profiles, Winograd's CV, Cambridge Core, and ACM establish publication details, author context, reception, and the book's link to language/action design. Winograd's 1980 essay supplies companion evidence for SHRDLU and the limits of formal language understanding. EU AI Act and NIST sources establish current oversight and risk-management context; they do not prove that any particular AI product is safe.

The bounded claim is not that language models cannot be useful, nor that representation is obsolete. The claim is narrower: once model output enters work, law, care, education, customer service, or software operations, the unit of analysis is the whole action setting. Treating the model alone as the bearer of understanding hides the institutional machinery that turns text into consequence.

Current book, author, scholarly, legal, standards, and retail claims were rechecked against publisher/catalog, author, university, standards-body, and regulator sources on June 25, 2026. This page makes no claim that any AI system is conscious, divine, or AGI; it treats AI systems as designed action settings that remain accountable to human institutions.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Search on Amazon Browse Books

Return to Blog · Return to Books