pipette
ESEspañol

Attenuated, Not Absent. Trusting Autonomous Systems Tracks Reliability More Closely, With Less Neural Engagement, Than Trusting Humans

D. A. Rogers, A. Weuthen, A. Finn, M. Schlesewsky, I. Bornkessel-Schlesewsky, M. Ullsperger

Preprint

In the authors' words

Users of autonomous systems and AI assistants decide whether to accept a system's output, yet whether such active trust decisions engage the same evaluative architecture as trust between people remains unknown. Forty-four participants made trial-by-trial decisions to trust or distrust a human or autonomous system partner during fMRI, with identical prescripted accuracy for both partners across five difficulty levels. Despite equivalent performance, trust in humans exceeded their reliability. At the highest difficulty level, the estimated probability of trusting the human partner was 72% but only 51% for the system, despite both performing at 52% accuracy. Distrust decisions recruited a prespecified performance monitoring network of valuation, salience, social-perceptual and affective regions. On hard trials, this network remained functionally coupled for both partners. Across the task, trust and distrust were discriminable in every component for both, with greater distrust-related activation and discriminability for human partners. Discriminability differences were expressed on hard trials, where partner accuracy was uncertain and trust diverged between partners. These findings indicate that trust in humans and autonomous systems engages a shared evaluative architecture, but that engagement is attenuated during autonomous system interaction. Whereas our earlier passive monitoring study found the P3a to system errors absent with no behavioural difference, an active decision appears to impose a floor of engagement, so that technology-related cognitive economization, a reduction in cognitive and social-affective resources for non-human partners, takes the form of attenuation rather than absence. Alongside this attenuation, trust in the system tracked its reliability without exceeding it.

Main resultThe abstract does not state a limitation.

Appeared: Wednesday, November 18. bioRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.14.766654