CS PAPERS DEEP-READ · PAPER 47
Alan Turing · journal Mind · 1950
In 1950 Alan Turing — the mathematician who invented the "Turing machine" and helped break wartime codes — wrote an article in a philosophy journal that opens with the question: "Can machines think?" But instead of arguing over what "think" means, he swapped that unresolvable question for a game you can actually play — what everyone now calls the Turing test. Ever since ChatGPT, people keep asking "does it really understand?" — and we're still using the framing Turing set down over seventy years ago.
"Can machines think?" sounds clear but can't be answered: nobody can pin down what "thinking" is, so the argument dissolves into everyone's private definitions. Turing's trick was to skip the definitions and look at performance: don't ask whether something is "going on inside" — just see whether what it does is distinguishable from a human. If you can't tell them apart, you've no grounds to say it isn't thinking.
Three players: a judge, a machine, and a real human. The machine and human are hidden and can only type on a screen to the judge (so voice and looks don't count — only how they "talk"). The judge asks anything, trying hard to catch which is the machine; the machine tries hard to pass as human. If the judge often can't tell, we say the machine passes. Turing even made a bold forecast: by the year 2000, a machine could fool an ordinary judge into guessing wrong about 30% of the time over five minutes.
Turing thought further ahead: directly programming a grown-up's clever mind is far too hard. Better to build a "child machine" and then educate it — like raising a kid, using rewards and punishments so it learns from experience on its own. Educated enough, it grows into an adult mind. This "make something that can learn, then teach it" idea is the seed of today's machine learning — named in a single stroke seventy years ago.
The "Turing test" became the cultural yardstick for machine intelligence; and "don't hand-code it, let it learn" became the main road of modern AI. A philosophy article without a single line of code laid the groundwork for a whole field.
One honest note: chatting well isn't the same as understanding. Later, a trivial template-matching program (ELIZA) fooled plenty of people, prompting the doubt: does the Turing test measure a machine's intelligence, or just how easily humans are fooled?
Turing swapped the unanswerable "can machines think?" for a game you can play: have a machine and a human type to a judge, and if the judge can't tell which is the machine, count it as thinking — turning a metaphysical puzzle into an observable behavioral test. He also foresaw that "build a child, then educate it" is the road to machine intelligence. One idea became AI's touchstone; the other, the herald of machine learning.
Want the full setup of the imitation game, Turing's replies to nine objections, and the later controversy? → switch to the deep read
Turing replaced the unverifiable metaphysical question "can machines think?" with an operational behavioral test — the imitation game, later called the Turing test: if a judge conversing by teleprinter with a machine and a human cannot reliably pick out the machine, we have no grounds to deny that the machine "thinks." He then rebutted nine objections one by one, and proposed a concrete route to machine intelligence — don't program an adult mind, build a "child machine" and educate it — now seen as a precursor to machine learning, reinforcement learning, and evolutionary computation.
The author is Alan Turing; the article appeared in October 1950 in the philosophy journal Mind (not a technical journal). By then, more than a decade had passed since his 1936 work founding "computability" and the universal machine, and since his codebreaking at Bletchley Park; the world's first programmable electronic computers had just appeared. It inherits Turing's own universal-machine ideas and launches toward the 1956 Dartmouth workshop (where the term "artificial intelligence" was coined) — this paper predates that by six years, and is the field's acknowledged founding text of philosophy and purpose.
Turing opens: "I propose to consider the question, can machines think?" But he immediately flags a flaw: to answer it you'd first have to define "machine" and "think," and those words can only be pinned down by surveying how ordinary people use them — which degenerates into "polling who votes that machines think," absurd and unresolvable.
The debate of the day was stuck right there: one camp was sure a machine is just a soulless, unconscious lump of metal, so talk of thinking is nonsense; the other was vague. Both were arguing over the essence of "thinking" — and essence can't be verified; you can't even prove that other people think, so why demand a higher bar of machines? What Turing does is cut this knot from an "argument about essence" into "a question you can actually decide."
Turing's first move is to swap in a different question — one "closely related to the original and expressed in relatively unambiguous words." He starts from a parlor game: a man (A) and a woman (B) sit hidden in another room, and a judge (C) asks questions by typed notes, trying to tell which is the man and which the woman; A tries to mislead the judge, B tries to help. Then Turing poses the key question: "What will happen when a machine takes the part of A?" Will the judge decide wrongly as often as when the game is played between a man and a woman?
Strip away that original gender wrapper and you get the Turing test everyone knows: a judge, over a teleprinter, converses at once with a machine and a real human, trying to tell which is the machine, while the machine tries to pass as human. If the judge can do no better than chance at catching the machine, we should grant that the machine "thinks."
The depth of this move is that it swaps the unverifiable "essence of thinking" for observable "conversational performance" — a behaviorist criterion. It has three virtues: first, it compares talk, not looks or voice, cleanly separating "thinking" from "resembling a human"; second, its scope has no ceiling — the judge may ask about poetry, arithmetic, chess, the weather, nearly anything; third, it flips the burden of proof: if you insist the machine can't think, you must say exactly which line gave it away. Turing isn't defining "what thinking is" — he's setting an executable standard for "when we ought to grant that it thinks."
Having swapped the question, Turing rehearses and refutes nine "a machine cannot think" arguments, and this forms the argumentative body of the paper. The weightiest few:
The remaining ones — the theological objection (thinking belongs to the soul), the "heads in the sand" objection (machines thinking would be too dreadful, let's hope they can't), continuity in the nervous system, the informality of behavior, and even his half-serious discussion of extra-sensory perception (ESP/telepathy) — Turing answers each in turn. The whole section has one edge: nearly every "machines can't think" claim rests either on overconfidence about ourselves or on a narrow imagination of machines.
The most far-sighted part is the closing section on learning machines. Turing reckons the engineering cost of directly programming an adult mind is daunting. So he proposes instead: "Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child's? If this were then subjected to an appropriate course of education one would obtain the adult brain."
He splits the plan into three parts: the initial state (the factory "child machine"), the education (a teacher training it with rewards and punishments), and other experience. He even offers an analogy close to an evolutionary algorithm: treat the child machine's structure as hereditary material, changes to it as mutations, and the experimenter's choices as natural selection — keeping the good variants. He also suggests mixing a random element into learning.
In 1950 this reads like science fiction, yet it precisely foretells the main road AI took: intelligence need not be written out in full — it can be learned. Today's gradient-trained deep learning, reinforcement learning's rewards and punishments, and evolutionary algorithms all trace a shadow back to this section.
This is a work of philosophical argument — no experiments, no datasets. Its "result" is a framework of ideas and one bold prediction. Turing forecast: in about fifty years (around 2000), computers with a storage of about 10⁹ bits (roughly a hundred-odd MB today) could be programmed to play the imitation game well enough that an average interrogator, after five minutes of questioning, would have no more than a 70% chance of a correct identification (i.e., fooled roughly 30% of the time). He also predicted that by then, usage of words and educated opinion would shift so much that "one will be able to speak of machines thinking without expecting to be contradicted." He even sized the effort: "At my present rate I produce about a thousand digits of programme a day, so about sixty workers, working steadily through the fifty years, might do the job." These numbers are less exact prophecy than a way of bringing a philosophical stance down to a discussable magnitude.
Its influence far outstrips a single paper: the "Turing test" became a cultural totem of machine intelligence — nearly every AI textbook and every public debate over "does it really understand" begins here. Two deeper contributions: first, it set an operational goal and criterion for a discipline not yet born — the field was only named "artificial intelligence" at Dartmouth six years later; second, the "learning machine / child machine" section pointed, half a century early, to the correct route of letting machines learn for themselves. It also dragged the discussion of "intelligence" out of theology and metaphysics into a place engineering and science could take up. ELIZA in 1966, the Loebner Prize contests from 1991, and today's large-language-model evaluations are all long shadows it cast.
① In one line: replace the unverifiable "can machines think?" with the operational behavioral test "imitation game / Turing test" — if a judge chatting behind a screen can't tell the machine apart, count it as thinking.
② Motivation: the "essence" of thinking can't be verified — you can't even prove other people think — so a higher bar for machines is unfair; Turing cuts "an argument about essence" into "a question you can decide."
③ Mechanism: machine and human are interrogated by teleprinter; only talk counts, not looks or voice; the machine impersonates a human, and if the judge can't catch it, it passes. The criterion flips the burden of proof onto the denier.
④ Argumentative body: rebuts nine objections — mathematical (Gödel), consciousness (anti-solipsism), Lovelace (machines can surprise us), the various "can never X" — all aimed at human overconfidence.
⑤ Most far-sighted idea: rather than hand-code an adult mind, build a "child machine" and educate it with rewards, punishments, and experience, add randomness, and select as in evolution — the herald of machine learning, RL, and evolutionary computation.
⑥ Prediction: around 2000, a 10⁹-bit machine could fool an average judge into ~30% wrong identifications after five minutes; by then "machines think" would go uncontradicted.
⑦ Impact: set the goal and criterion for AI before it was even named (six years before Dartmouth); became a cultural totem of machine intelligence.
⑧ Limits: the Chinese Room shows "faking ≠ understanding"; it measures "human-like," even "how gullible we are" (the ELIZA effect), not intelligence itself; narrow and anthropocentric; perhaps a rhetorical move to end an argument, not a definition.