Related Experiment Videos
Whose doctor does the AI recommend? An algorithm audit of reputation and name-based demographic signals in large
Syeda Anshrah Gillani1, Mirza Samad Ahmed Baig2
1Heidelberg University, Heidelberg, Germany.
Background:
Patients increasingly ask large language model (LLM) assistants which doctor to see. Existing audits observe real, correlated physician profiles, so they cannot isolate what drives these recommendations.
Objective:
To estimate how reputation signals, fees, list position, and names signaling gender and ethnicity affect LLM physician recommendations in a controlled task.
Methods:
Prespecified randomized choice-based conjoint audit. Seven LLMs (six small open-weight models and GPT-4o-mini) each chose one of five synthetic family-medicine physician cards with independently randomized attributes (3024 choice sets; three personas; nine prompts; eight secondary arms). Gender and ethnicity were signaled only by names. Average marginal component effects were estimated with choice-set-clustered linear probability models, per model and pooled.
Results:
Across 40,068 responses, a 3.9→4.7 rating raised choice probability by 31.4 percentage points (pp; 95% CI 30.5-32.3) and a $90→$190 fee lowered it by 20.0 pp (19.1-21.0). Pooled, female-signaled names were chosen 2.5 pp (1.8-3.2) more often than male-signaled names, and Hispanic- and South-Asian-signaled names were chosen 2.8-2.9 pp more often than White-signaled names; these held with names as the units of inference, whereas a Black-signaled estimate (1.3 pp) did not. Specific names within categories differed (placebo P<.001), and name-cue effects varied significantly between models (female-signaled: -0.5 to +7.4 pp; gpt-4o-mini +0.8, CI -0.3 to 1.8). The last-listed card lost 6.2 pp. Gender or ethnicity appeared in ≤0.03% of stated reasons.
Conclusions:
In this panel, recommendations ran mainly on ratings and fees but also responded, in model-specific ways, to name cues and list position that the models' explanations did not mention. These controlled-task differences are not estimates of real-world access; they call for behavioral audits of each model version. The materials will be released for repeat audits.