Trust but verify: a longitudinal study of epistemic trust calibration in doctoral students’ evolving relationships with generative AI

Generative AI in academic knowledge construction raises concerns about epistemic trust-the ability to assess the credibility and reliability of AI-generated knowledge. This study examines how doctoral students' trust in ChatGPT evolves over nine months, identifying distinct trajectories, mechanisms and implications for academic knowledge practices. A qualitative longitudinal study involved 32 doctoral students from eight countries. Data comprised monthly reflective journals (N = 288), five bimonthly in-depth interviews (N = 150 after attrition), and AI interaction logs with stimulated recall. Grounded theory and trajectory analysis guided interpretation. Four trust-calibration trajectories emerged: calibrated trusters (34%), over-trusters (28%), under-trusters (22%) and ambivalent oscillators (16%). Key calibration mechanisms included boundary testing, error analysis and social comparison. Calibrated trusters demonstrated stronger epistemic vigilance, critically evaluating AI outputs while using ChatGPT productively. Preliminary disciplinary differences were observed, with over-trust more common in engineering and under-trust in humanities, while no strong cultural patterns emerged. The study develops an initial model of epistemic trust calibration in human–AI academic partnerships and identifies mechanisms for fostering epistemic vigilance through structured reflection. Findings suggest that doctoral education should promote calibrated trust rather than unquestioning acceptance or wholesale rejection of generative AI.