Neural and behavioral dynamics of (mis)alignment with AI versus humans in risky decision revision

Artificial intelligence (AI) is increasingly used as social information in risky decision-making, yet little is known about how AI, versus humans, shapes decision revision at the neural level. We combined behavioral measures and event-related potentials (ERPs) to examine how feedback attributed to an AI or human counterpart influenced risky decision revision under alignment and misalignment. Participants completed a two-stage risky decision task: they first chose between a risky and a safe option, then viewed an aligned or misaligned choice attributed to either an AI or a human counterpart, and finally made a revised choice. Switching rates and attitudes toward AI were measured, and ERPs were time-locked to feedback onset. Behaviorally, participants switched more when misaligned with AI than with a human, and higher AI acceptance predicted greater switching under AI misalignment. Neurally, the N1 was more negative for human- than AI-attributed feedback, suggesting early differentiation of counterpart identity. The FRN was more negative for misalignment than alignment, with no counterpart effect. The P3 was larger for alignment than misalignment and larger for AI than human under alignment, and increased with more positive AI attitudes. These findings suggest stage-differentiated processing of AI- and human-attributed social feedback during risky decision revision.