Pass for a person
Hold a five-minute chat that most judges can't tell from a human's, in a controlled, repeatable study.
Judged human 73% of the time“I propose to consider the question, ‘Can machines think?’” — A. M. Turing, 1950
Turing set a test and a deadline. A judge talks to two strangers for five minutes, one person and one machine, and has to say which is which. Here is the bar he set, and what happened when it was finally run properly.
He predicted that by about 2000 an average judge would identify the machine correctly no more than 70% of the time after five minutes.
In a controlled three-party test, judges chose GPT-4.5 as the human 73% of the time, so they were right only 27% of the time. They picked the machine over the real person.
Judged human in 23% of games in the same study. The gap between ELIZA and GPT-4.5 is six decades of work.
One witness is a person, one is a machine told to pass as human. Pick the machine.
Example answers written for this page to show the format. They are not taken from the study.
“Computing Machinery and Intelligence” in the journal Mind replaces “can machines think?” with a game anyone can run.
Joseph Weizenbaum's script mirrors your words back as questions. Some users confided in it anyway.
IBM's machine wins a six-game chess match against the reigning world champion.
DeepMind wins 4–1 at Go, a game long thought to need human intuition.
Demis Hassabis and John Jumper share the Chemistry prize for predicting protein structures. The AlphaFold database now holds about 200 million of them.
Jones and Bergen run a pre-registered three-party test. GPT-4.5 is judged human 73% of the time.
Google DeepMind and OpenAI models each solve 5 of 6 IMO problems in plain language, reaching the gold-medal score of 35 out of 42.
Passing for human was the old finish line. These are the things machines do today that Turing's test never asked about.
Hold a five-minute chat that most judges can't tell from a human's, in a controlled, repeatable study.
Judged human 73% of the timeWrite full proofs for the hardest school-level maths competition in the world, within the human time limit.
35 / 42 points, IMO 2025Predict the 3D structure of a protein from its sequence, work that once took a lab months or years per protein.
~200 million structures publishedBeat the best human players at chess and Go, and find moves professionals now study.
AlphaGo 4 – Lee Sedol 1