Introduction
Dreyfus's critique of AI is often summarized as the claim that computers lack common sense. That summary misses the argument's structure. The problem Dreyfus identifies isn't how much a system knows. It's how a system knows which of the things it knows matter. This reading works through his statement of that problem in three moves, connects it to Merleau-Ponty's intentional arc, flags a common misreading, and ends with a model paraphrase and a test case from current AI. This content reflects the field as of September 2026.
Locating the Passage
The passage is in the introduction Dreyfus wrote for the 1992 edition of What Computers Still Can't Do (Dreyfus 1992, Introduction to the MIT Press Edition). There he reviews the research since the book first appeared, including efforts to give programs a large store of everyday knowledge, and argues that the obstacle they met is relevance.
The book isn't freely available. His 1965 RAND paper is, and it states an early version of the argument (Dreyfus 1965). In chess, the paper argues, a strong player doesn't count out every possible move. The player's awareness of the whole board lets them zero in on the promising lines, which is something the programs of the time couldn't do. Key phrases from the 1992 passage are paraphrased below.
Walking Through the Argument
Step 1: The problem of relevant facts
Dreyfus begins with an observation about everyday knowledge. Any ordinary situation involves indefinitely many facts that could matter. Whether rain is relevant depends on whether you're walking to work, driving, or reading indoors.
Notice what this rules out. Facts don't come labeled as relevant. Their relevance depends on the situation. But a symbolic system can represent the situation only as further facts. So the system's first task, before it can use any of its knowledge, is to find which of it applies. Dreyfus's point is that this task isn't a detail to be handled later. It's the central problem.
Step 2: Why adding more rules restates the problem
The natural response is to add rules for relevance: if walking outdoors, weather is relevant. Dreyfus's second move shows why this doesn't help.
Each relevance rule has conditions, such as "walking outdoors." The system must recognize that the conditions hold, and doing so means finding the relevant facts about the current situation. That's the original problem. A rule for recognizing the conditions would need conditions of its own, and so on. The shift here is from a shortage of knowledge to a regress. More knowledge doesn't end the regress. It lengthens it.
The regress can stop only in something that isn't a further rule. For Dreyfus, it stops in a skill: a practical ability to see what's relevant without first deciding it.
Step 3: How skilled perception already "sees" what matters
The third move turns to human beings. We don't face the regress, Dreyfus argues, because we don't begin with context-free facts. A skilled agent is already in a situation, and the situation shows up already organized. Some things stand out. Most things don't show up at all.
An experienced driver doesn't check whether the weather matters. The wet road shows up as slippery, and the driver's foot eases off. Relevance is given in perception, through skill, before any judgment. That's why the relevance problem and the skill model belong together. The expert's intuitive response is the human solution to the problem symbolic AI can't solve.
Callout: Merleau-Ponty's Intentional Arc
Dreyfus finds the mechanism for this in Merleau-Ponty. The intentional arc is Merleau-Ponty's name for the pre-reflective projection, around us, of our past, our future, our human setting, and our situation, which unifies perception, action, and understanding (Merleau-Ponty [1945] 2012, pt. 1, ch. 3).
Dreyfus reads the arc as a feedback loop. What we've learned through skilled activity isn't stored as facts or rules. It's carried forward in how the world shows up to us next. Each new situation solicits responses that our history has prepared us to feel. Relevance, on this reading, isn't computed. It's the shape our past gives to the present.
Key Considerations
Terms and translations.
| Term | Source | What it names |
|---|
| Relevance problem | Dreyfus | How a system finds which facts matter in a situation |
| Frame problem | AI research | Narrowly, what changes after an action; broadly, relevance |
| Fringe consciousness | Dreyfus, 1965 | A peripheral awareness of the whole situation |
| Intentional arc (arc intentionnel) | Merleau-Ponty; the same phrase in Landes and Smith | The pre-reflective projection that unifies perception, action, and understanding |
A common misreading. The most common misreading takes Dreyfus to claim that machines can never do anything intelligent.
- He distinguished domains. He expected programs to succeed in domains that can be fully formalized, such as calculation, and to struggle where relevance depends on context.
- His target was an approach. He criticized symbolic AI and its assumptions, not the idea of an artificial system as such. In 2007 he sketched what a better approach would require.
- His claim was partly empirical. It predicted where a research program would stall. Predictions can be tested, and Dreyfus revised some of his own.
Summary
The passage moves from the problem of relevant facts, through the regress that relevance rules create, to skilled perception that already sees what matters. Here is a model paraphrase of the argument to check your reading against:
A symbolic system represents the world as facts and acts by rules. In any real situation, indefinitely many facts could be relevant, and which ones matter depends on the situation. Rules for deciding relevance don't help, because each rule must itself be applied, and recognizing when it applies is the same problem again. The regress stops only in a skill: an embodied capacity to see what matters without consulting a rule. People have this capacity because their past experience shapes how situations show up to them, as Merleau-Ponty's intentional arc describes. So symbolic AI's problem isn't that it knows too little. It's that it starts from facts, where human intelligence starts from situations (Dreyfus 1992).
The paraphrase gives you a test to apply to current AI. Take a chat assistant built on a large language model, and ask it whether to take an umbrella on a walk when rain is forecast. It will likely answer sensibly. Does it face the relevance problem?
| Question to ask | If it has answered the problem | If it has sidestepped it |
|---|
| Where does its sense of relevance come from? | Learned from examples, as an expert's comes from experience | Borrowed from human text that already encodes what people find relevant |
| What happens if the usual relevance is overturned? | It tracks the change: you want to get wet because you're testing a waterproof jacket | It gives the usual advice, missing that the situation has changed |
| Does it have a situation of its own? | Not needed, if relevance can be learned | Needed, on Dreyfus's view, and missing |
The test doesn't settle the question. It tells you what to look for.
References
- Dreyfus, Hubert L. 1965. Alchemy and Artificial Intelligence. RAND Paper P-3244. Santa Monica, CA: RAND Corporation.
- Dreyfus, Hubert L. 1992. What Computers Still Can't Do: A Critique of Artificial Reason. Cambridge, MA: MIT Press.
- Merleau-Ponty, Maurice. (1945) 2012. Phenomenology of Perception. Translated by Donald A. Landes. London: Routledge.