Turing’s 1950 paper replaces an argument about whether machines think with a question about what they can do in a conversation. It then asks why a programmable computer might eventually succeed, considers objections, and proposes learning as a possible route. The paper is a theoretical argument: its most famous numerical claim is a prediction, not an experimental result. Read the original paper.
Why replace the original question?
“Can machines think?” seems to invite a yes-or-no answer. But before answering, we would need to agree on what counts as a machine and what counts as thinking. Turing worries that defining these words by ordinary usage could turn the question into a survey of opinions rather than an investigation of capabilities. Paper, §1.
His alternative is an operational question: describe a situation, specify what participants can observe, and ask how a machine would perform. That move makes disagreement more precise. Someone can accept that the game is observable while still disputing whether it captures the kind of thinking they care about.
The paper’s central move is not to define consciousness away; it is to substitute an observable comparison for a contested definition.
From a party game to a machine comparison
Turing first describes a game involving a man, a woman, and an interrogator. The interrogator cannot see the participants and exchanges written messages with them. The man tries to lead the interrogator to a wrong identification; the woman tries to help. Turing then asks what happens if a machine takes the man’s part: would the interrogator be wrong as often as in the original game? Paper, §1.
The familiar simplified version compares a hidden machine with a hidden human and asks the judge to identify the machine. That is the version illustrated above. It is useful, but it is worth noticing that the paper initially frames the comparison through the original gender-identification game, not as a fully specified modern benchmark. Paper, §§1–2.
The text channel removes easy clues from appearance and voice. It also restricts the evidence: the judge sees the answers, not the program or the machinery producing them. Questions can range across subjects and include follow-ups, so success requires more than one convincing sentence. Turing illustrates this with imagined exchanges about poetry, arithmetic, and chess; these are examples of questions, not transcripts from an actual experiment.
For example, a judge could ask both participants to explain a joke and then ask why an alternative ending is less funny. A fluent first answer would not settle the matter: the follow-up tests whether the answers remain coherent. The particular questions, time limit, judge, and participants would all affect the outcome.
Why a general-purpose computer enters the argument
The game does not automatically show that a computer can play it. Turing therefore narrows his discussion to digital computers: machines that follow specified rules through discrete operations. He describes a store for information, an executive unit carrying out operations, and a control governing the sequence of instructions. Paper, §4.
A program makes the same physical computer behave differently on different tasks. Turing explains the universal-machine idea: subject to practical limits such as storage and speed, a digital computer can be programmed to imitate the rule-governed steps of another discrete-state machine. This is why he can discuss the right program for a general-purpose machine instead of requiring a new physical device for each task. Paper, §5.
That is an argument about possibility, not a demonstration that a particular program can sustain human-like conversation. The universal-machine idea says what can be represented under appropriate assumptions; it does not supply the necessary program, training, computing resources, or evidence that the imitation game has been won.
The objections: what Turing answers and what remains open
Turing devotes much of the paper to objections. He does not treat them all as versions of the same problem. Paper, §6.
Consciousness. The objection says that producing the right words is not enough: a machine must feel or understand what it says. Turing points out the difficulty of directly verifying another being’s inner experience and argues that behavioral evidence is normally relevant to our judgments about others. This is a response to the objection, not a measurement of machine experience. A successful conversation would still leave room to debate what, if anything, the system experiences.
Mathematical limits. Results associated with Gödel and computability show that formal systems and machines have limits. Turing argues that demonstrating a limit for a particular machine does not establish that humans are free of comparable limits. His reply challenges a comparison; it does not eliminate mathematical limits or prove humans and machines reason identically.
“Machines can only do what we tell them.” Turing discusses Lady Lovelace’s objection that a machine cannot originate anything. He replies that machines can surprise their operators: consequences of rules and inputs need not be obvious to the person who designed them. Surprise to an observer is evidence against a simple predictability claim, but it is not by itself a complete definition or proof of creativity.
Informal behavior. Another objection holds that no set of rules could describe everything a person should do in every circumstance. Turing distinguishes explicit rules for behavior from laws governing a system’s operation: not knowing a compact rule for every situation does not establish that an underlying mechanism cannot generate varied behavior. This is a conceptual distinction, not a complete account of human behavior.
The remaining objections include theological claims, arguments from perceived consequences, supposed specific disabilities, continuity of the nervous system, and extrasensory perception. Their historical context matters; Turing’s treatment of a claim should not be mistaken for modern evidence that the claim is true.
Learning instead of writing every answer
Turing closes by asking how a machine might be prepared for the game. Rather than hand-coding the behavior of an educated adult in full detail, he suggests building a simpler child machine and subjecting it to an education. The starting structure and the learning process would both matter. Paper, §7.
The analogy is useful because it separates two design questions: what capabilities and mechanisms must the initial system possess, and what experiences or feedback will change it? Turing discusses rewards and punishments, instruction, and variation and selection as possible ingredients. The paper does not provide a tested recipe that shows which combination will work.
Consider the difference between programming every possible answer to a question and designing a system that changes its behavior after feedback. The latter shifts some work from the programmer’s explicit instructions to the learning process. It also creates new problems: feedback can be incomplete, training can fail, and a convincing response need not mean the system acquired the concept as a person does.
The learning proposal is a research direction, not an empirical result or a description of a modern model’s training pipeline.
Results: a proposal and a forecast, not a measured score
The paper reports no experiment running the imitation game. It gives no sample of judges, observed success rate, experimental data table, or statistical uncertainty. Its substantive outputs are the proposed comparison, the arguments for considering digital computers, responses to objections, and a suggested research direction involving learning.
Turing also forecasts that in about fifty years, computers with roughly 10⁹ bits of storage might play the game so well that an average interrogator would have no more than a 70% chance of making the right identification after five minutes of questioning. He explicitly presents this as a belief about future performance. Paper, §6.
| Claim | What the paper provides | What it does not provide |
|---|---|---|
| Imitation game | A proposed conversational comparison | A reported trial of that comparison |
| 70% after five minutes | A forecast about an average interrogator | An observed accuracy or a universal passing threshold |
| Learning machines | An outline of a possible approach | A validated training procedure |
The 70% refers to the interrogator’s correct identification, not the machine’s score. In a hypothetical two-choice setting, 70% correct would mean 30% incorrect; the paper does not report observing either number. The interpretation of any real rate would also depend on the protocol and participants.
What evidence would a later test need?
An actual study would need to say which machines and people participated, how judges were selected, how long conversations lasted, what instructions and questions were allowed, and how identifications were scored. Repeated trials and uncertainty estimates would help distinguish a robust result from a lucky exchange. These are requirements for interpreting a later experiment—not measurements supplied by the 1950 paper.
Even a carefully measured result has boundaries. A high rate of mistaken identification would support a claim about indistinguishability in that test, with those judges and conditions. It would not by itself establish reliability in other tasks, factual accuracy, human-like understanding, or subjective experience. Conversely, a failure might reflect the chosen conditions rather than settle every question about a system’s capabilities.
What to carry forward
Read the paper as a chain of distinct moves: replace a vague question with a behavioral comparison; explain why programmable computers are relevant; test the proposal against objections; and suggest learning as a route to more capable systems. None of these steps is interchangeable with an experimental measurement.
The durable lesson is to connect a claim to the evidence that could actually support it. When evaluating an AI system, ask what behavior was tested, who judged it, what comparison was made, and which conclusions go beyond the observations. That habit remains useful even if you disagree with Turing about whether the imitation game is the right test of intelligence.
References and further reading
- Turing, “Computing Machinery and Intelligence” (1950) — primary source; see §§1–2 for the game, §§4–5 for digital computers, §6 for objections and the forecast, and §7 for learning machines.
- Carnegie Mellon University Libraries: The Imitation Game — historical context for the paper.