The Case for a Neuropsychology of Artificial Agents

Le résumé fourni par la source

Artificial agents now interact with hundreds of millions of people daily in settings of high consequence, including medicine, law, finance, and education. Yet, a validated psychometric framework has never been applied to any of these systems. The elements required to build such a measurement program are now in place: a substrate-agnostic validity methodology, a well-validated dimensional reference framework for affect and personality, mechanistic interpretability findings confirming that internal affective representations organize along valence-arousal axes and causally drive misaligned behavior, and prior theoretical frameworks that make this convergence theoretically intelligible rather than merely analogical. On this foundation, we propose agentic neuropsychology as one such approach: the extension of two methods of clinical neuropsychological assessment to artificial agents, namely latent-construct inference from behavioral output and, where interpretability access permits, direct measurement of internal state. Model activations and emotion vectors serve as the internal-state proxies that neural measures provide in human assessment. To operationalize this approach, we introduce the Nomos Battery as an initial five-domain instrument, the initial measurement framework for the psychometric infrastructure AI safety governance currently lacks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.