Did Your Doctor Use AI in Med School?
Effects of Disclosed AI Use During Training on Trust in Experts

Background

A growing share of professional competence is now acquired with the help of AI. Medical trainees routinely use AI tools to study, prepare for exams, and build clinical skill.

Prior work asks how patients react to AI used at the point of care (Shaffer et al., 2013; Zondag et al., 2024). We ask a different question: When people learn an expert’s competence was developed with AI assistance (with present credentials held fixed) does that change how much they trust the expert, and why?

We extend Mayer, Davis, & Schoorman’s (1995) ability–benevolence–integrity model of trust with a fourth dimension, autonomy of judgment, to explore how disclosure of AI training affects perceptions of a doctor’s trustworthiness, competence, character, and independence of thought.

Method

Participants (recruited on Prolific) read a vignette about a board-certified physician who recently finished training. Training history was manipulated between subjects; all other credentials were held constant. Participants then rated trust/comfort and, on 7-point agree–disagree items (0 = strongly disagree, 6 = strongly agree), the doctor’s ability, benevolence, integrity, and autonomy, and how likely he is to use AI in his current practice.

Composites required Cronbach’s α ≥ .80. Effects were tested with one-way ANOVA + Bonferroni-corrected pairwise contrasts, then parallel mediation (Hayes’ PROCESS, model 4, 10,000 bootstrap samples) with all candidate mediators entered simultaneously.

Study 1 Preregistered

A board-certified cardiologist (“Dr. Smith”) was described as having used AI tools during training, having never used them, or with no information given.

Omnibus: F(2, 213) = 41.32, p < .001, Cohen's f = 0.62, error bars: 95% CIs  ·  ** p < .01, *** p < .001

Figure 1. All three pairwise contrasts are significant (Bonferroni-corrected), and the effect cuts both ways: the trust rating for “did not use AI” is higher than both “used AI” and the no-information baseline. 

Why? Mediation

With all four trust dimensions entered as parallel mediators, the direct effect of condition became non-significant (p ≥ .08), consistent with full mediation.

Figure 2. Only ability and autonomy carry indirect effects (bootstrap CI excludes zero, filled points) for both contrasts. Benevolence’s CI excludes zero for one contrast only, and in the opposite (suppressive) direction. Tested on its own, what participants expected about the doctor’s current use of AI in practice carried no indirect effect either (both bootstrap CIs include zero).

Learning that a doctor trained with AI doesn’t just make him seem less skilled, it makes him seem less likely to think for himself. It does not make him seem less caring or less principled.

Study 2 Preregistered

Are these AI-specific effects, or do they reflect a generic penalty/reward for any disclosed outside help? Study 2 (5 conditions) adds a parallel peer-tutor arm alongside the AI manipulation, sharing one control group.

Omnibus: F(4, 328) = 16.39, p < .001, Cohen's f = 0.45, error bars: 95% CIs  ·  ** p < .01, *** p < .001

Figure 3. The AI-training effect replicates almost exactly (left): using AI reduces comfort, and not using it raises comfort above the no-info baseline (both p < .01, Bonferroni-corrected). The peer-tutor arm (right) shows no significant differences (all ps ≥ .09).

Figure 4. Perceived ability (the strongest mediator in Study 1) robustly explains the AI effect (both CIs exclude zero). The peer-tutor arm has no significant total effect for ability to mediate (both CIs include zero, open points).

The two-sided AI effect does not generalize to human help: the parallel peer-tutor manipulation produced no reliable effect on comfort in either direction. The penalty comes from AI assistance in particular, not assistance in general (Reif et al., 2025).

Discussion

Across two preregistered studies, disclosure of information about the involvement of AI in a doctor’s medical training moves trust in both directions from a shared no-information baseline: using AI sharply reduces it, and explicitly not using it raises it just as reliably (p = .006). This suggests not only a penalty for AI use, but a reward for not using it. Neither direction reflects doubts about the doctor’s goodwill or integrity; both trace to perceived competence and capacity for independent clinical judgment. Study 2 shows this two-sided pattern is specific to AI: a parallel peer-tutor manipulation had no effect on comfort.

People appear to treat AI-assisted training as compromising the self-authored character of expertise itself — a distinct concern from whether the expert is a good or principled person. This has implications for how AI use during training is (or isn’t) disclosed in medicine and other professions, and for how licensing and educational bodies think about AI-assisted skill acquisition.

References