Definition: Dreyfus's critique of AI is the argument, first made by the philosopher Hubert Dreyfus in 1965 and in his 1972 book What Computers Can't Do, that human intelligence cannot be written down as explicit rules. Drawing on Heidegger and Merleau-Ponty, he said experts act from trained skill and a sense of what matters, not from a list of facts.
TL;DR: Hubert Dreyfus told the AI field in the 1960s that rules and stored facts would never add up to common sense, because people decide what is relevant through skill, not lookup. Symbolic AI stalled where he said it would. Neural networks then learned from data instead, a route he doubted would scale. His frame problem and his five-stage skill model still apply to AI agents. Build an AI agent free →
Ask a good nurse how she knew a patient was getting worse before the monitors showed it, and she may struggle to answer. She did not run through a checklist. She saw it. A beginner nurse would need the checklist, and might still miss the signs. Dreyfus built his critique on that gap between the beginner who follows rules and the expert who sees what the situation calls for. His claim was that early AI tried to build the beginner and call it the expert.
Why Dreyfus Matters in 2026
Dreyfus matters in 2026 because the problem he named, knowing which facts are relevant right now, is still the hard part of building AI agents. Modern models do not run on hand-written rules, so his attack on symbolic AI no longer hits its original target. But his deeper point about relevance, skill, and context shows up every time an agent has too much context, the wrong context, or no sense of what matters.
He was also the first serious critic of the field from outside it. In his 2007 paper Why Heideggerian AI failed and how fixing it would require making it more Heideggerian (Artificial Intelligence 171(18):1137-1160, free preprint), Dreyfus recalls that MIT AI Lab students came to his Heidegger course in the early 1960s and told him that philosophers had failed and the AI Lab would now succeed. He quotes Marvin Minsky in 1968: "Within a generation we will have intelligent computers like HAL in the film, 2001." Those predictions did not come true on that schedule. The critique Dreyfus wrote in response became one of the most argued-over texts in the history of AI.
A Short History of the Critique
The critique grew over four decades, from a RAND report to a paper on why "Heideggerian" AI also failed.
| Year | Work | What it said |
|---|---|---|
| 1963 | RAND invites Dreyfus to evaluate Newell and Simon | Their claim that minds and computers are both physical symbol systems |
| 1964 to 1965 | "Alchemy and Artificial Intelligence", RAND paper P-3244 | Digital programs "systematically exclude" three basic human forms of information processing |
| 1972 | What Computers Can't Do | The four assumptions of AI, and the case that common sense resists formalization |
| 1986 | Mind over Machine, with Stuart Dreyfus | The five-stage model of skill, from novice to expert |
| 1992 | What Computers Still Can't Do | A revised edition with a new look at neural networks |
| 2007 | "Why Heideggerian AI failed" | Even robots built on Heidegger's ideas missed changing relevance |
| 2024 | Reynolds, "Framing the predictive mind" | Applies the critique to predictive processing theories |
The 1964 or 1965 question. Philip Brey's chapter Hubert Dreyfus: Humans versus Computers (in American Philosophy of Technology: The Empirical Turn, Indiana University Press, 2001) settles it: the report was "written in 1964", it "was vigorously criticized by Newell and Simon, but was released by the RAND corporation the next year over their objections." So both dates are right. RAND's own catalog page lists the title and summary but shows no date. Brey also calls it "the first detailed critique of AI to be published."
The Four Assumptions He Attacked
Dreyfus argued that early AI rested on four assumptions that its researchers treated as facts. The Wikipedia article on Dreyfus's views summarizes his claim that each is "one hypothesis among others" that must be tested, not taken as given.
| Assumption | What AI researchers assumed | Dreyfus's objection | How it aged |
|---|---|---|---|
| Biological | The brain processes information in discrete, on/off steps | Neurons have analog features | Most researchers had dropped it by 1972 |
| Psychological | The mind applies formal rules to bits of information | Understanding rests on unconscious background know-how | Daniel Crevier wrote that by 1993 AI researchers "no longer made the psychological assumption" |
| Epistemological | All knowledge can be formalized | Much human knowledge is not symbolic | Learned models replaced most hand-written knowledge |
| Ontological | The world is a set of independent facts | Things show up as relevant to what we are doing | Still the most debated, because it is about relevance |
The ontological assumption is the one that matters most today. It is the claim that the world can be broken into context-free facts that a system then picks from. Dreyfus thought the world shows up already organized by what we care about. A model that stores facts still has to decide which ones matter for this request, and that decision is where AI systems still fail most visibly.
What He Got Right
Dreyfus got three things right: symbolic AI stalled, the frame problem was real, and common sense was the wall.
The frame problem. Dreyfus wrote that "the deep problem wasn't storing millions of facts; it was knowing which facts were relevant in any given situation." When Minsky proposed frames, structured descriptions of typical situations such as a birthday party, Dreyfus saw a regress. A system needs frames to recognize which frame applies, and frames to recognize those frames.
The regress Dreyfus saw in frame-based AI
situation: people exchanging wrapped boxes
|
v
which frame applies? --> need a frame for recognizing birthday parties
| |
v v
which frame applies to --> need a frame for recognizing
recognizing frames? frame-recognition situations
|
v
... no bottom, unless something other than rules
picks out what is relevant
Micro-worlds. In the same paper he notes that programs at MIT in the 1960s and early 1970s worked in "micro-worlds", artificial settings where the few relevant features were fixed in advance. Terry Winograd's SHRDLU, which moved virtual blocks on command, was "the parade case of a micro-world program that really worked", but only in its micro-world. The promised generalization did not follow.
Common sense. The Wikipedia article reports Crevier's verdict that "time has proven the accuracy and perceptiveness of some of Dreyfus's comments", and links his distinction between intuitive and deliberate thinking to later work by Kahneman and Tversky.
He also paid a price for being right early. The same article records that Seymour Papert dismissed part of the paper as "gossip", that Herbert Simon accused him of using the RAND name for publicity, and that Papert arranged a chess match against the program Mac Hack, which Dreyfus lost. One bulletin headline read "A Ten Year Old Can Beat the Machine", followed by "But the Machine Can Beat Dreyfus."
What He Could Not Foresee
Dreyfus could not foresee how far learning from data at scale would go. He did see that neural networks fit his picture better than symbolic AI. Brey writes that Dreyfus "has asserted that the basic assumptions of neural network AI are compatible with his own vision of intelligence", because knowledge in a network is not a set of explicit representations.
But Dreyfus was pessimistic about where networks would end up. In Brey's summary, the problem was "the incredible complexity of human intelligence": a network learns from the situations it happened to meet, and to generalize to new ones it would need to know which past experience is relevant. Brey concludes that, on this view, a network that generalizes like a person would need human-scale complexity and an artificial body that goes through a human-like development. "So far the creation of such an android form of life belongs to science fiction."
Large language models are the strongest test of that prediction. They learned from a vast record of human text rather than from a body, and they generalize across far more situations than any 1990s network. Critics of Dreyfus read this as a refutation. His defenders reply that models still fail on exactly the relevance and context problems he named, from hallucinations to context rot, and that their competence borrows from the bodily experience of the people who wrote the text. The symbol grounding and embodied cognition articles cover that reply. The fairest summary is that the route changed, and the argument about the destination is still open.
Dreyfus Meets the LLM
Recent philosophy applies Dreyfus to the theories that sit closest to how LLMs work. Jack Reynolds' Framing the predictive mind: why we should think again about Dreyfus (Phenomenology and the Cognitive Sciences, published online May 2024) takes the critique to predictive processing, the view that brains minimize prediction error (see predictive coding). Reynolds argues that predictive processing faces an unresolved tension: it aims to explain all cognition with its formal tools, yet it must also explain how living systems establish relevance and switch flexibly between contexts. Those are the abilities Dreyfus tied to embodied skill, habit, and judgment.
Dreyfus himself rejected the first wave of "Heideggerian" AI. In the 2007 paper he gives Rodney Brooks, author of Intelligence without representation (Artificial Intelligence 47, 1991), credit for robots that respond to the world through sensors instead of an internal model. But he says Brooks' robots "respond only to fixed isolable features of the environment, not to context or changing significance", and so they "finesse rather than solve the frame problem." He held that view of Phil Agre's work too, and pointed instead to Walter Freeman's neurodynamics as a more promising model.
| What the philosopher said | What an LLM does | Open question |
|---|---|---|
| Relevance cannot be looked up from stored facts | Weighs every token in context through attention | Does learned attention solve relevance, or hide the problem? |
| Experts see the situation, novices follow rules | Answers fluently without stated rules | Is fluency expertise, or a very good imitation of it? |
| Skill comes from a body in a world | Learns from text that people wrote about their world | Can borrowed experience stand in for lived experience? |
| Micro-worlds do not generalize | Generalizes across many domains | Where are the edges, and how do we find them before users do? |
Novice to Expert: The Dreyfus Skill Model
The five-stage skill model is the most practical thing Dreyfus left behind. The Wikipedia article on the model records that Stuart E. Dreyfus and Hubert L. Dreyfus first proposed it in February 1980, in a report for the U.S. Air Force Office of Scientific Research produced at UC Berkeley's Operations Research Center. Mind over Machine (1986) developed it further.
| Stage | How the person works | What they rely on |
|---|---|---|
| 1. Novice | Follows context-free rules step by step | Rules and instructions |
| 2. Advanced beginner | Notices situation-specific details | Rules of thumb from experience |
| 3. Competence | Chooses a goal and a plan, and owns the result | Deliberate planning |
| 4. Proficiency | Sees what the situation calls for, then decides how to respond | Intuition plus rules |
| 5. Expertise | Sees and acts in one move, without deliberation | Trained perception |
The model has critics. Fernand Gobet and Philippe Chassy argued in 2008 and 2009 that "there is no empirical evidence for the presence of stages in the development of expertise", and that experts often use slow, analytic thinking, such as look-ahead search in chess. Defenders reply that the critics overlook what Dreyfus called "deliberative rationality". Treat the stages as a useful description, not a measured law.
How Dreyfus's Critique Shows Up in Taskade
Dreyfus's critique shows up in Taskade as a way to stage an agent, not as a feature. His novice-to-expert ladder maps onto how much room you give a Taskade AI agent: start it on written rules and a checklist, review its work, then widen its scope as the record shows that it is ready for more.
| Skill stage | How to set up a Taskade agent | What the person does |
|---|---|---|
| Novice | Write explicit steps and a checklist into the agent's instructions (prompt engineering) | Review every output |
| Advanced beginner | Attach example projects that show good and bad cases as knowledge | Review most outputs |
| Competence | Let the agent plan a multi-step task and report the plan first | Approve the plan, spot-check the result |
| Proficiency and beyond | Let the agent run inside an automation with clear limits | Keep a human-in-the-loop step for high-stakes actions |
This is a staging method for people who manage agents. It does not claim that an agent gains skill the way a person does. The model under an agent does not retrain between runs. What improves is the material it reads: clearer instructions, better examples, and a written record of past mistakes. The agent evaluation page covers how to measure whether an agent is ready for more room. Taskade Genesis builds an app with its projects, agents, and automations from one prompt. To start from a working example, clone the Onboarding Guide Portal kit, an app kit with projects, an AI agent, and automations you can clone. Its written steps are the novice stage in practice. The Free plan covers 3 apps and 1 agent and starts with 0 AI credits.
Describe yours and build it free →
Related Concepts
- Cognitive Science: Mind, Brain and AI: the hub for this category
- Phenomenology and AI: the Husserl, Heidegger, and Merleau-Ponty ideas Dreyfus built on
- Embodied Cognition and Enactivism: the research program that took up his point about the body
- Symbol Grounding Problem: how symbols get meaning, and whether LLMs solve it
- Intentionality and AI: the aboutness of thought
- Context Window: how much a model can hold at once, which is not the same as knowing what matters
- Context Engineering: the practical craft of giving a model the relevant context
- Jagged Intelligence: why models are strong at some tasks and weak at nearby ones
- Predictive Coding: the brain theory Reynolds tests against Dreyfus
- AI Meets Psychology: how psychology and AI shaped each other
- What Is Intelligence?: from neurons to AI agents, the longer story
- The Hard Problem of Consciousness: why experience resists a functional explanation
Frequently Asked Questions About Dreyfus's Critique of AI
Who was Hubert Dreyfus?
Hubert Dreyfus was an American philosopher who taught a Heidegger course at MIT in the early 1960s that AI Lab students attended. He specialized in Heidegger and Merleau-Ponty and used their ideas to argue that early AI could not reach human intelligence. His best-known books are What Computers Can't Do (1972) and Mind over Machine (1986, with Stuart Dreyfus).
What was Dreyfus's main argument against AI?
He argued that human intelligence rests on background skill and a sense of what is relevant, which cannot be written down as explicit rules and facts. Symbolic AI assumed it could be. He predicted that this approach would stall on common sense, and it did.
When was "Alchemy and Artificial Intelligence" written?
Dreyfus wrote the RAND report in 1964, and RAND released it in 1965 over the objections of Allen Newell and Herbert Simon, according to Philip Brey's 2001 account. It is often dated 1965 for that reason. Brey calls it the first detailed critique of AI to be published.
What is the frame problem in Dreyfus's sense?
It is the problem of knowing which facts are relevant in a given situation, and which can be assumed unchanged when something happens. Dreyfus argued that adding more rules only moves the problem, because a system needs rules to decide which rules apply. Relevance, in his view, comes from skill, not lookup.
Did large language models prove Dreyfus wrong?
Partly, and the question is open. LLMs learned from data instead of hand-written rules, and they handle far more situations than the systems Dreyfus attacked. His defenders reply that models still fail on relevance and context, and that their competence borrows from human experience recorded in text.
What are the five stages of the Dreyfus skill model?
Novice, advanced beginner, competence, proficiency, and expertise. A novice follows rules, a competent performer plans, and an expert sees what to do without deliberating. Stuart and Hubert Dreyfus proposed the model in 1980, and critics question whether the stages are real.
How can the Dreyfus model help with AI agents?
It gives a way to stage trust. Start an agent on explicit instructions and a checklist, review its work closely, and widen its scope only as its results earn it. Keep a person on the decisions where an error would cost the most.