Computing Machinery and Intelligence

In 1950 Alan Turing replaced an unanswerable question — “can machines think?” — with a game. Seventy-five years on, his paper still frames the argument.

In October 1950, Alan Turing opened the philosophy journal Mind with a sentence that reads more like a dare than an essay: “I propose to consider the question, ‘Can machines think?’” By the second paragraph he had abandoned it. The words “machine” and “think” were too loaded, too Wittgensteinian, too much like a Gallup poll. So he replaced the question with one he could actually attack — and in doing so created the founding text of artificial intelligence.

The paper is short, sly, and deeply weird. It is also, page for page, one of the most influential things ever written about minds.

The imitation game

Turing's substitution is famous now as the Turing test, but it began as a Victorian parlour game. Three rooms, three participants. In Room A, a man. In Room B, a woman. In Room C, an interrogator who can only communicate with the others by typed messages. The man's job is to pretend to be a woman; the woman's job is to convince the interrogator she is the woman. The interrogator must decide which is which.

Now, Turing said, replace the man with a digital computer. Same setup. Same typed messages. Can the computer do at least as well as the man at fooling the interrogator?

Notice what the substitution buys him. He has not asked whether the machine has an inner life, a soul, qualia, or a stream of consciousness. He has asked whether its observable conversational behaviour, sustained across any topic the interrogator chooses, is indistinguishable from a human's. The metaphysics is set aside. What remains is an empirical, decidable challenge.

Critics have been chipping at the test ever since. But the strategy of the move — replace an essentially contested question with an operational test, and see whether the test captures enough of what we care about — is the move that gave AI a research programme.

Nine objections, answered in advance

Turing then does something striking: he lists nine objections he expects to be raised, and answers them. The list is a snapshot of mid-twentieth-century anxieties about thinking machines, and most of those anxieties are still in circulation today.

The theological objection holds that thinking is a function of the immortal soul God gave to humans, not machines. Turing answers, drily, that an omnipotent God could presumably bestow a soul on a machine if He wished — denying this would limit His power.

The “heads in the sand” objection: it would be too dreadful if machines could think, so let's not consider the possibility. Turing dispatches this in a paragraph. It is not an argument, it is a wish.

The mathematical objection (from Gödel and Turing himself) points out that there are statements no consistent formal system can prove, so no machine can prove them either, while a human mathematician somehow can. Turing's reply is essentially: we don't actually know that no human is subject to similar limits; and in any case, the inability to answer one specific question does not preclude general intelligence.

The argument from consciousness, which Turing identifies as the deepest, says that a machine pushing symbols around does not feel anything, and until we know it does we cannot say it thinks. To this Turing offers an elegant rejoinder: by the same standard, I cannot know that you think either — I only have your behaviour to go on. Solipsism is the price of the objection, and few are willing to pay it.

The Lady Lovelace objection — “the Analytical Engine has no pretensions to originate anything; it can only do whatever we know how to order it to perform” — Turing concedes historically and then upends. Machines can in fact surprise their programmers, very frequently, and it is precisely the inability to predict the consequences of one's premises that distinguishes interesting reasoning from rote calculation.

The remaining objections — informality of behaviour, continuity of the nervous system, various disabilities, even ESP — he treats with varying patience. (He takes ESP unusually seriously, given that the parapsychological evidence of his time, he thought, was strong. This part has not aged well.)

The child machine

Turing closes with one of the loveliest ideas in the paper. Programming an adult human's worth of experience, opinion, and skill directly into a machine is hopeless. There is too much in there. So why not, he asks, build something simpler — a child machine, with the structure of an unformed mind — and then teach it?

He sketches a kind of evolutionary education — random mutations to the structure, rewards for correct behaviour, punishments for incorrect. He guesses that around 10⁹ binary digits of storage might be enough. He guesses, with extraordinary boldness for 1950, that by the year 2000 a machine should be able to hold the imitation game well enough that the average interrogator does no better than seventy percent after five minutes.

That last prediction is roughly when chatbots first started fooling casual users — about half a century late, but in the right neighbourhood for the idea. The architecture of the bet — train, don't program; learn from data, don't enumerate rules; treat the system as an organism, not a clockwork — is the bet that modern machine learning eventually went on to win.

Why the paper still matters

“The original question, ‘Can machines think?’, I believe to be too meaningless to deserve discussion. Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.”

Reading “Computing Machinery and Intelligence” today, after seventy-five years of large-scale neural networks, transformers, and language models that hold conversations of a kind Turing would have recognised, it is hard not to feel the strangeness of his foresight. He wrote before transistors were widespread. The most powerful computer in the world had a few kilobytes of memory.

What he saw, that almost nobody else then saw, is that intelligence might not be a property of any particular substrate. If symbol-manipulation is what brains essentially do, and if Turing machines can in principle perform any well-defined symbol-manipulation, then minds and machines are not categorically different things; they are differently realised cases of the same kind of thing. The imitation game is the operational consequence of that hypothesis. The child machine is the engineering consequence.

He was right about the words. We do speak that way now. Whether we are also right to is the question the paper bequeathed us — and it is still open.