Work

Two errors in attributing intelligence and a Bayesian Turing test

Year: 2026
Type: preprint
Abstract: Recognising intelligence is a difficult, unsolved problem. The process that produces behaviour is never observed, only inferred, and we are asked to settle a question that only inference can address. This Perspective argues that the two classic mistakes in that judgement (seeing a mind where there is none, and refusing to see one where the evidence supports it) are the same bias in opposite directions. It has a psychological basis in essentialist thinking and a measurable signature: identical behaviour is judged differently depending on who is thought to have produced it. Humans attribute minds along two dimensions, intelligence and experience, which so far have always travelled together; large language models (LLMs) are the first case on which the two come apart, and the first on which both errors appear at scale simultaneously. Re-interpreting Turing's test as an inference in which mechanism counts as well as behaviour, I propose a test any attribution or denial should pass, apply it to sceptics and enthusiasts alike, and outline what behavioural science can contribute.
Language: English
Subfield: Philosophy
Open Access status: green
OpenAlex ID: W7221246146
Cites: 87
Cited by:
Related to: 10