Turing's 1950 test was a game, not a verdict on thinking

The conversation

In the article published in the October 1950 issue of Mind, Alan Turing set aside the question "Can machines think?" and proposed a substitute he called the imitation game. Three participants take part: a man, a woman, and an interrogator in a separate room who exchanges typewritten questions and answers with both and must decide which is which. Turing then asked what would happen if a machine took the man's place: would the interrogator misidentify the machine as often as a human player fools the interrogator in the original game? The full text of the paper lays out the setup, including the requirement that answers be typewritten or relayed by an intermediary so tone of voice cannot give a speaker away.

What the documents show

Turing's own text states a dated conjecture, not a settled claim: in about fifty years, he wrote, computers with a storage capacity of about 10^9 could play the game "so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning." He calls this a belief, not a proof, and separately calls the original question "too meaningless to deserve discussion." The Mind journal's own catalog record confirms the paper's place as pages 433 to 460 of volume LIX, issue 236. Nothing in the document claims the game predicts or licenses any later system's behavior; it proposes a replacement question for a philosophical one.

The system boundary

The paper describes no software architecture and sets no operator handoff, because it is not a specification for a machine; it is a criterion for a judge. Turing separates two things a reader might conflate: winning the game under its stated conditions (a short session, typewritten exchange, an interrogator who does not know the setup in advance) and possessing whatever internal state "think" or "understand" are meant to describe. He treats the latter question as one he can set aside, not settle. The paper's own replies to objections, including the storage-capacity argument and Lady Lovelace's objection, defend the game's usefulness as a research question, not a claim that any 1950 machine played it.

Where it fails

The recurring failure in later use of this paper is treating a narrow, time-boxed exercise in avoiding identification by one interrogator as a general test of understanding. Turing's conjecture concerned statistical performance against average interrogators, not a threshold that, once crossed, settles the question he set aside. A builder invoking "the Turing test" should say which version is meant and what evidence would count against it.

• Does the claimed test specify its judge, time limit, and topic restrictions, the way the original imitation game does?
• Is a system's success at avoiding identification being treated as evidence of understanding, or only of imitation under stated conditions?
• What would the source itself need to show before "passed the Turing test" is a fair description?

Read this way, the 1950 paper remains a founding document for a debate about method, not a specification any later system can be said to have fulfilled or failed.