A structured framework for human-AI resemblance: Evidence, inference and perception.

Abstract

Artificial intelligence is increasingly encountered not as background software but as an interlocutor, adviser and institutional gatekeeper. Yet claims that an AI system is 'human-like' often collapse distinct comparison targets, measurements and inferences into a single label. This Review proposes a structured framework in which every resemblance claim specifies the human reference class, the AI system and version, the task and context, the property compared, the measurement relation, the perturbations considered, the uncertainty of the estimate and the inference that the evidence permits. Four recurring evidence families concern behavioural outputs, representational organization, process dynamics and neural or physiological correspondence. Perceived resemblance is treated separately as an observer attribution that can mediate trust, reliance, disclosure, delegation and moral concern. We distinguish exact from approximate resemblance, symmetric from directional measures, static from dynamical correspondence and observational from interventional evidence. Worked examples show how to code missing or conflicting indicators without forcing them into a single score. Convergence supports broader conclusions only when measures with different assumptions exclude distinct alternatives; otherwise, agreement may reflect shared data, stimuli, objectives or mappings. The framework is intended both for exploratory science and for application-matched evaluation, while preserving strict limits on claims about cognition and consciousness.