When do researchers trust AI? Trust as a mediator and experience as a boundary condition in researchers’ intentions to use generative AI

Generative artificial intelligence (GenAI) is increasingly embedded in academic work, yet its opaque processes and the possibility of misleading outputs make reliance on these tools inherently risky. Building on trust theory and stimulus–organism–response (SOR) logic, we examine how perceived accuracy shapes researchers' intention to use GenAI, and whether this relationship operates through trust and varies by users' experience. Across two online studies with researchers (total N = 402), we test a moderated mediation model in which accuracy cues (stimulus) influence trust in GenAI (organism), which in turn predicts usage intention (response), with familiarity/experience moderating the trust–intention link. Study 1 used a three-wave longitudinal design with one-week time lags ( N = 232), temporally separating perceived accuracy (T1), trust (T2), and usage intention (T3). Results were consistent with (a) a positive association between perceived accuracy and trust, (b) an indirect association between perceived accuracy and usage intention via trust, and (c) a negative interaction showing that the trust–intention relationship weakened as familiarity/experience increased. Study 2 conceptually replicated these findings using a vignette experiment manipulating accuracy of ChatGPT recommendations ( N = 170): the accuracy manipulation increased trust, and trust predicted usage intention; the indirect effect via trust was moderated by experience. Together, these findings suggest that trust is a key psychological mechanism through which accuracy cues are linked to researchers’ intentions to use GenAI in academic work, while highlighting experience as a boundary condition: trust functions more like a heuristic for novices and becomes less relevant as experience grows.