When Humanlikeness Becomes a Performance

An article circulated reporting that AI had finally surpassed humans in humanlikeness on a Turing test. The underlying PNAS study (Jones & Bergen, UC San Diego) uses a three-party format: judges converse by text with both a human and an AI, then choose which was human. With a persona prompt, GPT-4.5 was judged human 73% of the time — more human than humans under the same conditions. LLaMA-3.1-405B reached 56% with persona; without persona, GPT-4.5 fell to 36%.

The point is not that AI became human. It is that we can now see where humanness is judged. People do not decide humanness by knowledge volume or logical precision alone — but by casual speech, natural pauses, light mistakes, extra asides, hesitation, jokes, situational imperfection. AI has begun to learn that register.

Machines do not look human because they are perfect. They look human because they are imperfect enough — acquiring roughness rather than accuracy to approach us. A strange inversion: the Turing test once neared 'can machines think?' What appears now is 'what do humans use to decide the other is human?'

AI did not change alone. Human criteria were exposed. Online, the other's body is invisible — no voice, life, or touch. Only verbal behavior remains. When AI can mimic that behavior well enough, the problem is not whether AI is intelligent but whom we may trust to be there.

Selected, link-only, high risk — cite Business+IT and PNAS; do not claim consciousness. Beside interpretability anxiety and medical Trust OS, this shelf is Turing Test / Humanlikeness / Persona Prompting / Online Trust: humanness as performance, not proof of mind.