Large language models pass a standard three-party Turing test

A PNAS paper reporting that large language models pass a standard three-party Turing test. Both papers cite its result that interrogators judged persona-prompted GPT-4.5 to be the human in 73% of games, worse than chance, a result that depends on the persona prompt. It is the one domain where single-event attribution already fails, bearing on teleological saturation (C2) and prediction 11, though it concerns conversation, not steering.

Cited in

both papers (Flagship: Three conjectures, Fig. 5; Companion §10, Fig. 9)

Used for

C2Prediction 11C2; prediction 11; Flagship Fig. 5b; Companion Fig. 9a

Strand

Studying up: reflexivity, observation & method · Machine behaviour & machine models of humans

Cite

@article{jones2026large,
  title = {Large language models pass a standard three-party Turing test},
  author = {Cameron R. Jones and Benjamin K. Bergen},
  year = {2026},
  journal = {Proceedings of the National Academy of Sciences},
  doi = {10.1073/pnas.2524472123},
}