Computing Machinery and Intelligence

Alan Turing’s 1950 paper replaces the broad question of whether machines think with a more specific question about performance in a conversational game. It combines that proposal with arguments about digital computers, responses to objections, and a program for educating learning machines. It is an argument and research proposal, rather than a report that a machine passed the game.^[1]

What the imitation game asks

Turing begins with three people: a man A, a woman B, and an interrogator C in a separate room. C asks questions and tries to identify which respondent is the man and which is the woman. A tries to induce a mistaken identification; B tries to help. Written or typewritten answers, ideally through a teleprinter, exclude clues from voice and appearance.^[1]

He then asks what happens when a machine takes A’s part, comparing the interrogator’s errors with those in the original human game. Later, section 5 formulates a version with a digital computer as A and a man as B. The paper’s changing formulations matter: its starting point is a particular three-party identification game, rather than a general exam of factual accuracy.^[1]

The separation of respondents lets Turing focus on intellectual performance instead of physical resemblance. Questions may range across poetry, arithmetic, and chess, but the interrogator cannot demand a physical demonstration. He acknowledges that a machine might think in ways unlike a human; the proposed criterion need not exhaust every possible form of Machine Intelligence.^[1]

Turing predicts that in about fifty years an average interrogator would have no more than a 70% chance of correct identification after five minutes, for suitably programmed computers with the storage he envisages. This is his forecast under stated conditions, not an experimental result or an independently established universal passing threshold.^[1]

Computers and objections

Sections 3–5 focus on digital computers and their ability, with enough storage, speed, and suitable programming, to imitate other discrete-state machines. Universality supplies a reason to investigate programming a general-purpose computer; it does not supply the program that would succeed in the game.^[1]

Section 6 addresses nine objections. Against the mathematical objection, Turing accepts limits on particular machines but disputes the unsupported inference that human intellect escapes comparable limits. Against an appeal to consciousness, he argues that requiring direct access to another being’s experience would also undermine ordinary judgments about other people. He explicitly leaves mysteries about consciousness unresolved. Discussing Lovelace’s objection that a machine can only do what it is ordered to do, he distinguishes having programmed a process from already knowing all its consequences, and turns toward learning.^[1]

Educating a child-machine

Section 7 proposes building a simpler child-machine and educating it, rather than directly programming an adult mind. The design and its education would be varied together, with the experimenter judging improvements. This remains a proposed research strategy. Turing reports teaching a few things in a limited experiment, but describes its method as too unorthodox to count as really successful.^[1]

Reward and punishment signals would make preceding events more or less likely to recur, without presupposing feelings. Turing argues that such signals alone communicate too little for practical education; other channels could transmit symbolic instructions and reduce the feedback required. He also anticipates that a teacher may predict some behavior without understanding much of what is happening inside the learner.^[1]

Published in Mind 59(236), pages 433–460, the paper belongs to the October 1950 issue. The timeline uses October 1, the publisher-deposited catalog date, without claiming a separately verified day of physical circulation.^[1]

For the origins guide, this supplies two distinct questions: how to recognize intelligent performance and how to teach desired behavior. Neither conversational imitation nor the proposed educational process establishes that a system will pursue appropriate goals outside the game. The paper does not document a direct historical line from its proposal to modern preference-based training.