Overview
- A paper published in PNAS in May 2026 reports randomized three‑party Turing tests that ran 1,023 five‑minute games across undergraduate and Prolific participant pools.
- When given an explicit humanlike persona prompt, GPT‑4.5 was judged human in 73% of five‑minute trials, a rate that statistically exceeded human performance in the study.
- The experiment showed persona prompting was essential because GPT‑4.5 without persona instructions was judged human only about 36% of the time.
- A 15‑minute replication that included GPT‑5 found the newer model remained only marginally above chance at 59.3%, showing longer talks and model version change moderate the effect.
- Authors say success came from copying tone, humor, directness and fallibility rather than true reasoning, and they warn the results raise clear risks for deception, authentication, and online trust while stopping short of claims about consciousness.