A body, a world, a situation: what rules do not say

Dreyfus's critique of AI

Hubert Dreyfus 1929 · 2017Human intelligence comes from a body engaged in a world, not from a set of rules: that is why programs that try to formalize everything fail.

See profile

What it claims

Symbolic AI fails to reproduce human intelligence wherever that intelligence rests on this know-how, this background and this situation. As early as 1965, Dreyfus sees that analyzing intelligent behavior in computer language "systematically excludes" these human forms of information processing.

Why

Human intelligence rests on bodily know-how and common sense that cannot be reduced to explicit rules.

The same idea through history

Martin Heidegger1927In Being and Time, the human being is first of all "being-in-the-world": it copes with things before representing them. This is the source Dreyfus draws on.

Maurice Merleau-Ponty1945Phenomenology of Perception describes the lived body as that through which we understand the world. Dreyfus mentions it as early as 1965 as a possible path.

Ludwig Wittgenstein1953Wittgenstein remarks that we cannot draw the exact boundaries of our concepts, and that following a rule is not applying one more rule. Dreyfus cites him as early as 1965.

Hubert Dreyfus1972The book What Computers Can't Do develops the critique: the body, the situation and human needs.

In short

"A body, a world, a situation: what rules do not say" is an argument by Hubert Dreyfus (Alchemy and Artificial Intelligence, 1965; What Computers Can't Do, 1972, 1972) on the question "Can a machine think?" (answer: it takes a living body). Its claim: Symbolic AI fails to reproduce human intelligence wherever that intelligence rests on this know-how, this background and this situation. As early as 1965, Dreyfus sees that analyzing intelligent behavior in computer language "systematically excludes" these human forms of information processing. According to AskPhilosophy's logic check (Z3 solver), the argument is valid only if one grants a premise the author leaves unstated (P4). AskPhilosophy presents 3 objections, 3 of them with a documented reply, and the original text behind each premise, checked word for word.

Alchemy and Artificial Intelligence, 1965; What Computers Can't Do, 1972

Valid with an implicit premise

In the early 1960s the first programs play chess and prove theorems, and their makers announce that machines will soon think like us. Hubert Dreyfus, a philosopher trained in phenomenology, answers that this is a dead end. When a doctor, a driver or a speaker does well, they do not follow a list of rules. They rely on know-how acquired by a living body, on a background of common sense that nobody ever wrote down, and on a way of grasping what matters in the situation. An AI that turns everything into explicit rules therefore misses the essential. An important point: Dreyfus does not say that no machine will ever think. He targets a precise kind of AI, the one that manipulates symbols by rules.

click a premise to reject it

If you accept every premise, logic forces you to accept the conclusion. So the debate is about the premises.

Checked by a logic solver: the premises cannot all be true while the conclusion is false.
Without the implicit premise: Without P4, the computer finds a world where an ability irreducible to rules could still be reproduced by a program that has only rules.

Valid with P4. This hidden premise is almost a definition: a symbolic AI program is made of explicit rules. The point of attack is elsewhere, in P2: showing that know-how cannot be reduced to rules is very hard, and that is where the objections bite. Dreyfus himself limits the scope: in his 1965 summary, of four types of intelligent activity, only two can be programmed without these human forms of information processing.

Analogy. Learning to ride a bicycle from a manual: "if you lean left, turn the handlebars left by a proportional angle...". Nobody rides by applying these rules. The expert has know-how in the body, and the manual, even a perfect one, gives only its shadow. The image comes from commentators on Dreyfus, not from him.

Objections

Each objection is a conversation: click to see who replies and where the debate stands.

John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon 1955 · attacks premise 2

The founding project of AI starts from the opposite conjecture: every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can simulate it. A know-how we cannot yet write down is a problem to solve, not a limit.

a defender replies

Hubert Dreyfus defender of the argument · 1965

Dreyfus answers by examining the existing programs (games, machine translation, pattern recognition). The difficulties are not due to the machines' power: analyzing intelligent behavior in computer language systematically leaves out three human forms of information processing (fringe consciousness, the distinction between essential and accidental, tolerance of ambiguity).

Marvin Minsky 1974

Summary: Minsky proposes "frames", knowledge structures that describe a typical situation (a room, a birthday party) with its usual expectations. It is a way of writing down common sense and context, which Dreyfus said could not be formalized.

See the rest of the debate (1)

Hubert Dreyfus 1992

Summary: in the introduction to What Computers Still Can't Do, Dreyfus judges that frames move the problem without solving it: one still has to decide which frame fits the present situation, and that choice depends on what matters to someone in a world, which no frame contains.

Where the debate stands (our summary)The basic disagreement is about whether common sense can be written down. AI researchers tried for decades (frames, large knowledge bases) with limited results, which Dreyfus reads as a confirmation. Their defenders reply that the task is huge but not impossible.

David Rumelhart, James McClelland and the PDP group 1986 · attacks premise 4

Neural networks (connectionism) have no written rules: they learn from examples and adjust thousands of connections. What they know is not a list of explicit rules, and yet it is a machine. So the critique of symbolic AI does not touch machines in general.

a defender replies

Hubert Dreyfus defender of the argument · 1992

Summary: Dreyfus grants this point in part. Networks escape the critique of rules, so they are a more promising path than symbolic AI. But he doubts they can reach human know-how as long as they have no body, no needs, no world where things matter to them: without that, they generalize without knowing what is relevant.

Where the debate stands (our summary)Both camps agree on one point: a neural network does not apply explicit rules. The disagreement is about what follows: can large-scale learning recover, without a body or a life, the sensitivity to the situation that Dreyfus holds proper to living beings? Recent deep learning successes have reopened this question without settling it.

Rodney Brooks 1991 · attacks the conclusion

Brooks agrees that symbolic representations get in the way: for simple intelligence, it is better to use the world as its own model. But he draws a different conclusion: one can build robots with a body that act in real offices without supervision. So the body is not an obstacle for machines, it is a solution.

a defender replies

Hubert Dreyfus defender of the argument · 2007

Summary: in "Why Heideggerian AI Failed", Dreyfus acknowledges that Brooks goes in the right direction (a body, no central model), but notes that these robots did not go beyond the level of insects. In his view they still lack what makes things relevant for a living being: needs of their own and a body that adapts to what matters to it.

Where the debate stands (our summary)Dreyfus and Brooks agree in rejecting the central model of representations. They differ on what to add: for Brooks, a robot body in a real world is enough to progress step by step; for Dreyfus, a body with needs and a world where things matter are required. Open question.