Blog · Review Essay · Last reviewed June 23, 2026

What Computers Still Can't Do and the Background of Intelligence

Hubert Dreyfus's What Computers Still Can't Do: A Critique of Artificial Reason is not useful because it settled the AI question. It did not. It is useful because it keeps returning attention to the parts of intelligence that become invisible when cognition is imagined as symbol manipulation, prediction, prompt response, or workflow automation: body, skill, situation, care, background, and the practical world that makes some possibilities matter more than others.

Background intelligence, in this review, means the situated field of habits, tools, norms, stakes, roles, and local knowledge that makes an action intelligible before it becomes a formal problem. Dreyfus matters in 2026 because many AI failures are not failures to produce language. They are failures to preserve the background that gives language its consequence.

This is a feedback problem, not a mystical one. Once a system becomes the place where work is described, routed, measured, and remembered, its missing background can return as policy, training data, workflow default, and institutional common sense. The practical question is where situated judgment survives after the interface has made the work machine-readable.

The Book

What Computers Still Can't Do was published by MIT Press in 1992 as a revised version of Dreyfus's earlier What Computers Can't Do. MIT Press lists the paperback as 408 pages, published October 30, 1992, with ISBN 9780262540674. Internet Archive's catalog record identifies the book as a 1992 MIT Press title on artificial intelligence, revised from the 1979 edition, with bibliographical references and an index.

The book sits inside a longer argument. RAND's record for Dreyfus's 1965 paper Alchemy and Artificial Intelligence describes it as an examination of the difficulty of simulating cognitive processes on computers, focused on forms of human information processing that resist translation into digital computer language. Berkeley's obituary for Dreyfus describes him as a scholar of Heidegger and European philosophy who taught at UC Berkeley for nearly fifty years and challenged expectations for artificial intelligence as early as the 1960s.

That background matters because Dreyfus was not merely making a technical forecast. He was attacking a picture of mind. Early AI often treated intelligence as the manipulation of explicit symbols according to formal rules. Dreyfus argued that human intelligence depends on background know-how that is not first represented as detached facts. We cope with situations before we describe them. We notice relevance before we search all possibilities. We learn practices through bodies, tools, environments, correction, and social life.

Artificial Reason

The title can mislead if it is read as a scoreboard. Modern systems can now do many things that would have looked startling to earlier AI critics: translate, summarize, draft code, recognize images, play strategic games, operate in multimodal interfaces, and produce fluent language at scale. The question is not whether computers can perform impressive cognitive tasks. They plainly can.

Dreyfus's stronger question is what kind of intelligence is being modeled when a system succeeds. Symbolic AI treated the world as if it could be represented in well-formed facts and rules, then searched or reasoned over those representations. The problem was not only scale. It was relevance. In ordinary life, people do not examine every possible fact and rule. They inhabit a situation in which some things show up as important, urgent, boring, dangerous, polite, plausible, or absurd.

That is why the book belongs beside Computer Power and Human Reason, Understanding Computers and Cognition, The AI Mirror, and The Experience Machine. It is a book about the gap between operating on representations and being answerable to a world.

The Background Problem

Dreyfus's most useful concept is the background. A person does not enter a task as a blank problem solver facing a database of possible facts. A person enters with bodily habits, learned equipment, social expectations, purposes, emotions, memories, institutional roles, and a sense of what normally matters. Much of that cannot be converted into a short rule without being changed.

The point is not that background is ineffable. It is that background is distributed. It lives in a clinician's trained eye, a classroom's rhythm, a shop floor's worn tools, a courtroom's procedure, a family's history, a codebase's local conventions, a public office's unwritten repair practices, and the time pressure under which a decision is made. A system can capture pieces of that setting, but the capture is always a selection.

A useful distinction is between prompt context and practical background. Prompt context is what can be supplied to a model: instructions, retrieved documents, logs, examples, screenshots, user preferences, or tool descriptions. Practical background is the larger setting that tells people why an action matters, when a normal answer is unsafe, which exception is morally or legally decisive, and who must be able to object. More context can help, but it is not the same as participation in the practice.

This is not mystical. It is practical. A nurse, driver, teacher, carpenter, moderator, lawyer, parent, organizer, editor, or programmer often knows that something is off before they can state the rule that was violated. Expertise is not just more facts. It is a better grip on the situation. It includes timing, salience, tact, caution, and the ability to revise the frame when the obvious interpretation fails.

AI systems force a governance question around that background. If a model drafts the medical note, where did the clinician's situated judgment enter? If an answer engine compresses a dispute into a confident paragraph, where did uncertainty remain visible? If an agent chooses the next step in a workflow, who defined what counts as relevant? If a risk score reorganizes public service delivery, whose background knowledge was stripped away as noise?

The danger is not that machines lack souls. The danger is that institutions may treat a machine's output as if it had already done the background work. A generated answer can look complete while depending on missing context. A dashboard can look objective while excluding local knowledge. A workflow can look efficient because it has moved ambiguity onto the person with the least power to object.

The Background Test

Dreyfus is most useful when turned into a deployment test. Before a system is trusted in a workflow, ask what background the system must assume: normal practice, local exceptions, tacit warning signs, professional role, legal duty, social stakes, bodily risk, and the difference between a plausible output and a responsible action.

The test is practical. What does the system see, and what is outside its frame? Which human practice supplied the examples, labels, demonstrations, documents, or corrections? Which exceptions cannot be captured by the input fields? Who can say "not in this case" and have the authority to change the outcome? What evidence survives when the system's judgment is later disputed?

The artifact should be a background register. It names the domain assumptions, missing context, local exceptions, irreversible actions, affected people, escalation triggers, source records, override authority, appeal path, and retirement conditions. It also states which parts of the background are deliberately outside the model because they belong to professional judgment, legal process, worker testimony, patient consent, community knowledge, or physical inspection.

A system passes only when the missing background has a place to go: visible uncertainty, source trails, local amendment, human override, incident review, appeal, and a way to retire the tool when the domain changes. Background intelligence is not a decorative philosophical idea. It is the difference between assistance that enlarges situated judgment and automation that hides judgment where no one can inspect it.

After Symbolic AI

The book's 1992 edition already responded to a changing AI field, including connectionism and neural networks. A 2007 Dreyfus article in Artificial Intelligence revisited "Heideggerian AI" after robotics, embodied approaches, and other attempts to move beyond old symbolic representation. A 2024 AAAI Spring Symposium paper revisits Dreyfus again in light of deep learning, large language models, hybrid AI, hallucination, common sense, and the continuing question of whether systems can move from performance to human-like understanding.

That history prevents two lazy readings. One says Dreyfus was simply right because AI still hallucinates, lacks common sense, and depends on human training data. The other says he was simply wrong because neural networks and LLMs blew past old capability limits. Both flatten the problem.

Modern AI weakens some of Dreyfus's old claims and strengthens others. It weakens the idea that explicit rules are the only route to broad machine performance. Pattern-learning systems can produce astonishing results without hand-coded common-sense rules. But those systems also make his background problem more urgent. They can simulate situated speech without sharing the situation. They can generate plausible reasons without being practically committed to the consequences. They can appear context-aware while depending on an interface, data pipeline, retrieval layer, prompt, evaluator, or user to supply the missing world.

That is why LLMs feel so uncanny. They often do not fail like old brittle symbolic systems. They fail by making the wrong kind of sense. They give an answer that fits the conversational surface while missing the practical, legal, emotional, institutional, or bodily situation in which the answer will be used.

Current Context

As of June 23, 2026, Dreyfus has to be read against systems that are no longer only symbolic programs or text generators. The active frontier includes multimodal models, retrieval-augmented generation, coding and browser agents, enterprise connectors, and embodied AI systems that connect models to tools, accounts, sensors, or physical action.

This does not make the book a prophecy or a refutation. It changes the location of the background problem. Modern systems often outsource background to scaffolding around the model: retrieval corpora, prompts, tool descriptions, memory stores, policy layers, human raters, logs, vendor contracts, and product defaults. When those scaffolds are weak, the model may look more situated than it is.

NIST's 2026 AI Agent Standards Initiative and NCCoE project on software and AI agent identity make this concrete. Agentic systems raise identity, authentication, authorization, interoperability, logging, and security questions because model output can now become action. OWASP's 2026 agentic-application risks point in the same direction: goal hijacking, tool misuse, privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and misplaced human trust are background failures in operational form.

That means "context engineering" cannot be treated as a synonym for understanding. A larger context window, a richer retrieval index, or a persistent memory store may improve performance, but the governance question remains: who decides what enters the context, what remains outside it, what authority the model may exercise from it, and what record proves that the system respected the boundary?

The Institutional Reading

Read in 2026, the book is less a metaphysical verdict on machine minds than an audit tool for delegated judgment. The practical question is not "Can this system think?" It is "What background has been removed from the task, and who is expected to repair the loss?"

In a private notebook, a model's missing background may be a manageable inconvenience. In a school, hospital, court, welfare office, newsroom, workplace, police department, or military system, missing background becomes institutional risk. A model summary can become the record. A recommendation can become the default. A confidence score can become a supervisor's reason. A generated explanation can become the public account of a decision that no person fully understands.

This is where Dreyfus connects to Tools for Thought. A good cognitive tool expands the user's contact with the world: more sources, more alternatives, more inspectable uncertainty, more situated correction, more ability to revise. A bad cognitive tool narrows the world while making the narrowed version feel fluent. It replaces friction with closure.

The same test applies to agents. An AI agent does not only answer. It routes attention, selects tools, writes to systems, triggers workflows, remembers context, and produces an action trail. That makes the background problem operational. If the agent does not know what matters, someone else must encode, constrain, monitor, or correct it. If no one can do that, the interface is not augmenting judgment. It is laundering underspecified judgment through automation.

Governance and Safety

Dreyfus becomes practical when an AI system enters a consequential workflow. The governance task is to locate the missing background before the output becomes action: domain assumptions, local exceptions, sources, uncertainty, affected people, human authority, appeal paths, and the cost of being wrong. The question is not whether the model has an inner experience. The question is whether the institution has preserved enough situated judgment for the decision to remain answerable.

Current public frameworks move in that direction, but they do not all have the same legal force. NIST's AI Risk Management Framework organizes risk work around govern, map, measure, and manage, and treats governance as lifecycle work rather than a one-time test. The EU AI Act is in phased implementation: Article 113 sets a general application date of August 2, 2026 with exceptions, while the European Commission's current implementation page says the May 7, 2026 political agreement on the AI omnibus sets high-risk application dates of December 2, 2027 for certain high-risk areas and August 2, 2028 for product-embedded high-risk systems. The operational lesson is still immediate: Article 12 requires logging for high-risk systems, Article 14 requires effective human oversight, and Article 27 requires certain deployers to assess affected groups, risks, oversight measures, and complaint mechanisms before use. U.S. OMB Memorandum M-25-21 is limited to federal agencies, but its high-impact AI controls are a useful operational template: pre-deployment testing, impact assessment, ongoing monitoring, human training, oversight, appeals, feedback, traceability, and discontinuing non-compliant use.

The safety checklist follows from Dreyfus's critique. A high-stakes AI system needs a background register tied to the AI system inventory, model or system card, audit trail, and recourse process. It should state what context the system sees, what it cannot see, which local expertise can amend the record, what uncertainty must remain visible, who owns the decision, what records prove meaningful review, and how an affected person can challenge or repair the result. For AI agents, add tool permissions, approval gates for irreversible actions, scoped credentials, prompt-injection testing, action traces, rollback plans, identity controls, and a stop condition.

This is why human oversight must be designed as authority, not symbolism. A reviewer who cannot inspect sources, slow the workflow, override the output, consult local knowledge, or trigger an appeal is not supplying the background Dreyfus worried about. They are being used as a decorative bridge between machine fluency and institutional responsibility.

Where the Book Needs Friction

What Computers Still Can't Do is brilliant and exasperating. Its polemical force helped make it famous, but the force can overrun the distinctions a reader needs now. Dreyfus was strongest when criticizing simplified pictures of mind and weakest when the word "can't" made a moving technical frontier sound fixed.

The book also risks romanticizing human cognition. Humans have background understanding, but we also rationalize, stereotype, forget, conform, overfit, hallucinate socially, and turn institutional habit into false common sense. A human decision is not automatically richer because it is embodied. A machine decision is not automatically empty because it is computational. The question is the arrangement: what evidence, feedback, responsibility, appeal, skill, and care surround the decision?

The same caution applies to governance. "Human judgment" should not mean undocumented discretion, expert mystique, or local custom protected from scrutiny. The point is to preserve situated evidence and accountable repair, not to sanctify whatever a person or institution already does.

Dreyfus does not give an AI governance program. He gives a philosophical warning. Readers still need other books for procurement, algorithmic accountability, race and gender bias, data labor, surveillance, model evaluation, safety cases, law, and platform power. The value of the warning is that it keeps those policy questions anchored in a deeper one: what kind of world must be assumed before the system can act?

What This Changes

The book changes the evaluation question. Do not ask only whether the system can produce the right answer in a benchmark setting. Ask what the system needs the world to become so that its answer counts as right.

For AI products, the test is concrete. Does the interface preserve uncertainty? Does it expose sources and alternatives? Does it let users bring local knowledge back into the loop? Does it support correction after action, not just before deployment? Can a user refuse the generated frame? Does the workflow teach skill, or does it make the user dependent on fluent prompts and hidden defaults?

For institutions, the test is sharper. Never let a model's fluency stand in for situated responsibility. If a decision affects rights, care, work, money, reputation, safety, or public memory, the institution must know where background judgment lives. It must be inspectable, contestable, and connected to people with authority to change the process.

Dreyfus's old critique survives because it is not finally about what computers can output. It is about the world behind the output. Intelligence is not only a result on a screen. It is a relation among bodies, tools, purposes, histories, institutions, and consequences. Any AI system that enters that relation should be judged by how honestly it preserves the parts it cannot contain.

Source Discipline

This review separates four layers of evidence. MIT Press, Internet Archive, RAND, Berkeley, PhilPapers, DBLP, and AAAI establish book history, Dreyfus's publication record, reception, and later scholarly reassessment. NIST, NCCoE, OWASP, the European Commission's AI Act Service Desk, and OMB establish current governance and agent-security language. Internal links supply continuity with the site's concepts, not external validation.

The distinction matters because Dreyfus is easy to misuse. A benchmark result does not settle the philosophical question. A product demo does not prove situated understanding. A hallucination example does not prove that all machine assistance is empty. A regulatory requirement does not prove that a deployment is safe. Strong claims about AI in practice need the task, domain, data source, interface, human role, evaluation date, failure modes, appeal path, and accountable owner.

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