Implicit neural measures of trust in artificial intelligence

Trust in AI systems is critical for effective collaboration, yet traditional measures—such as self-reports and behavioral proxies—are limited in capturing its dynamic and latent nature. This study introduces the contralateral delay activity (CDA), a neural marker of visual working memory load, as a novel, implicit index of trust. While the CDA has been widely used in change detection tasks to track memory load, we repurpose it here to measure how many working memory resources users offload to an AI-framed automated partner. Participants performed a lateralized memory task under low and high working memory load, collaborating with an AI whose reliability was experimentally manipulated across three phases: trust formation, violation, and restoration. The system was a rule-based automated agent framed to participants as an AI collaborator, as the psychological effects of AI framing are central to trust and offloading behavior. In dyadic trials, where the AI was responsible for one hemifield, the CDA amplitude served as an index of how much information participants chose to maintain themselves versus offload to the AI. As AI reliability increased, CDA amplitude rose, indicating greater trust and reliance. When AI reliability dropped during the violation phase, participants encoded more from both hemifields, and CDA amplitude declined. During trust restoration, CDA amplitude returned to pre-violation levels, indicating renewed reliance—though it did not match the high amplitude of solo trials, suggesting lingering mistrust. Behavioral measures (e.g., reliance, compliance, response time) tracked the CDA dynamics but lacked the resolution and specificity of the CDA. Together, these results establish the CDA as a powerful neural index of dynamic cognitive offloading that closely tracks behavioral and reliability-based indicators of trust. It captures trial-by-trial fluctuations in offloading behavior that reflect users’ evolving confidence in AI assistance, offering a continuous, covert, and cognitively grounded measure of trust in interactive settings.