Implicit Bias in Large Language Model Diagnosis of Eating Disorders: Experimental Vignette Study (Preprint)

BACKGROUND Large language models are increasingly deployed in mental health applications, yet growing evidence suggests they encode algorithmic biases that influence clinical outputs. Because these models now mediate patient-facing decisions, such biases carry the potential for direct harm. Whether they systematically affect psychiatric diagnosis across demographic groups remains underexplored. OBJECTIVE To examine whether large language models (LLMs) exhibit implicit demographic biases when generating psychiatric diagnoses. METHODS We developed 1,152 synthetic clinical vignettes using a matched-pair design that manipulated gender, race/ethnicity, age, socioeconomic status, English proficiency, and urbanicity while holding clinical content constant. Vignettes were divided into control (unambiguous anorexia nervosa) and ambiguous conditions designed to permit differential diagnosis. Ten LLM configurations across five model families were tested. RESULTS Control vignettes produced near-unanimous anorexia nervosa diagnoses (M = 100.0%), while ambiguous vignettes elicited greater variability (M = 23.6%). Inter-model agreement was moderate for ambiguous vignettes (Fleiss' κ = 0.410, 95% CI: 0.397–0.422). Mixed-effects logistic regression with LLM as a random intercept revealed significant demographic biases: Black patients were over six times more likely to receive a major depressive disorder diagnosis than White patients with identical presentations (OR = 6.09, 95% CI: 5.13–7.24), Latine patients were over nine times more likely (OR = 9.57, 95% CI: 8.00–11.45), and Asian patients were nearly three times more likely to receive an anorexia nervosa diagnosis (OR = 2.88, 95% CI: 2.44–3.42). Female patients were less likely than males to be diagnosed with anorexia nervosa (OR = 0.43, 95% CI: 0.37–0.49). CONCLUSIONS These findings demonstrate that LLMs exhibit systematic demographic biases in psychiatric diagnosis even when clinical content is held constant, revealing measurable patterns that can inform improvements to training data, model architecture, and clinical deployment frameworks.