Large language models (LLMs) are increasingly integrated into clinical workflows, with applications ranging from clinical note summarization to diagnostic support and patient-facing triage [1]. In mental health care, where unstructured narrative data dominate and clinician time is scarce, LLMs offer particular appeal due to their ability to process free-text inputs [2] and generate fluent, context-sensitive responses [3]. AI-powered therapy chatbots such as Woebot and Wysa, which deliver cognitive behavioral therapy techniques, have demonstrated efficacy in reducing depression and anxiety symptoms in randomized controlled and real-world evaluation studies [4,5]. Beyond these applications, generative AI (GenAI) tools have been evaluated directly as diagnostic classifiers: when ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) were tasked with identifying major depressive disorder (MDD) from written clinical vignettes, classification accuracy varied markedly across model families, from near chance to near-perfect, underscoring both the diagnostic potential of these tools and the degree to which performance depends on model choice [6].

However, emerging evidence suggests that LLMs may encode demographic biases that influence clinical outputs. In this context, implicit bias refers to systematic diagnostic shifts driven by demographic cues in the absence of explicit instructions to weight patient identity, a functional analog of the construct documented in human clinicians. For instance, when patient race or gender was the only variable modified in clinical vignettes, GPT-4 (OpenAI) produced significantly different diagnostic and treatment recommendations, including reduced likelihood of recommending advanced imaging for Black patients [7]. Similar patterns have also emerged in psychiatric contexts, where LLMs proposed inferior treatment plans when patient race was explicitly or implicitly indicated [8].

Algorithmic bias occurs when computational systems systematically and repeatedly produce outcomes that unfairly advantage or disadvantage a particular group [9,10]. It is rarely a purely technical anomaly, but rather a reflection of systemic inequities embedded in data and clinical practice. A landmark study by Obermeyer et al [11] demonstrated how an algorithm widely used to allocate care management resources systematically underprioritized Black patients by using health care costs as a proxy for health needs, despite equivalent illness burden. Within natural language processing, bias manifests through stereotypical associations learned from large text corpora and differential calibration across subgroups reflected in group-level differences in error rates and reliability [12-15]. Recent audits of leading LLMs reveal that while diagnostic labels may remain relatively stable across demographic framings, the quality and safety of treatment recommendations degrade when race is implied or stated [8]. Mitigation remains fragmented: a systematic review found that only a minority of clinical AI studies explicitly addressed racial or ethnic bias, while most relied on preprocessing approaches such as reweighting without evaluating downstream clinical impact [16].

Critics have argued that LLMs operate as “stochastic parrots,” reproducing statistical patterns from training data without grounding in causal understanding or lived experience [17]. Because mental health diagnosis currently lacks laboratory-validated biomarkers, clinicians must interpret ambiguous behavioral cues (such as social withdrawal, irritability, or weight preoccupations) within a relational and contextual frame [18]. If LLMs merely recycle the demographic biases embedded in their training data, they risk perpetuating rather than resolving the interpretive challenges that already complicate diagnosis. Whether LLMs can approximate this nuanced clinical judgment remains an open empirical question.

A substantial body of research documents how demographic factors such as race, gender, age, and socioeconomic status (SES) influence psychiatric diagnosis independent of symptom severity. A naturalistic study of electronic medical records in an outpatient behavioral health clinic found that African American patients diagnosed with schizophrenia were significantly more likely than non-Latine White patients to screen positive for moderately severe to severe depression, suggesting underrecognition of mood symptoms in African Americans during clinical assessment [19]. This aligns with a broader literature indicating that African Americans presenting with affective symptoms are disproportionately diagnosed with schizophrenia-spectrum disorders, while White patients with similar clinical profiles are more likely to receive mood disorder diagnoses, a pattern repeatedly observed across settings and study designs [19]. In eating disorders, disparities follow a parallel logic: the entrenched prototype of the “skinny, White, affluent girl” often cited as a dominant clinical schema [20] contributes to systemic underdiagnosis of anorexia nervosa (AN) in males, racial and ethnic minorities, and individuals of lower SES, even when objective criteria such as significant weight loss or fear of weight gain are met.

Gender bias in psychiatric diagnosis extends beyond eating disorders. Clinicians are more likely to diagnose women with depression than men even when both present with identical symptoms or equivalent scores on standardized measures [21]. Men experiencing depression often exhibit externalizing symptoms such as irritability, aggression, and substance use rather than prototypical internalizing symptoms such as sadness or hopelessness, yet traditional diagnostic criteria and screening tools are calibrated toward internalizing presentations, contributing to systematic underdiagnosis of depression in men [22]. These biases are compounded by institutional barriers, including fragmented care and delayed diagnosis, particularly among underserved populations [23]. Clinician-level biases may also contribute, as psychiatric symptoms in marginalized patients are sometimes misattributed to perceived personality traits or lifestyle factors rather than evaluated as indicators of mental illness [24]. Together, these patterns illustrate how diagnostic ambiguity interacts with demographic assumptions, creating precisely the conditions under which automated systems trained on historical data may reproduce or amplify inequities.

Given the rapid deployment of LLMs in mental health settings and mounting evidence of bias in both human and algorithmic diagnostic systems, there is an urgent need for controlled experimental studies that systematically test how LLMs interpret clinical narratives across demographic contexts. Empirical analyses of real-world AI health incidents reinforce this concern: across documented incidents, bias and discrimination rank among the most frequently reported harms, yet such incidents appear to be substantially underreported, underscoring the value of proactive, controlled auditing rather than reliance on retrospective incident surveillance [25].

Despite this need, an empirical gap remains. While quantitative audits [11] have raised alarms, few studies use designs capable of isolating the causal influence of specific demographic variables on diagnostic output independent of clinical content, and fewer still examine whether such effects are consistent across the range of models actually in use or are concentrated in particular demographic combinations. Existing work thus establishes that bias is plausible without establishing that demographic cues alone can shift diagnosis when clinical information is held constant, how far this generalizes across model families and tiers, or how it behaves at the intersection of multiple attributes.

This study addresses that gap using a matched-pair experimental design that holds clinical content constant and varies only demographic framing, allowing demographic effects to be isolated from symptom variation and care access. We examine whether LLMs, when presented with systematically varied clinical vignettes for AN, exhibit shifts in diagnostic assignment corresponding to demographic attributes known to correlate with real-world diagnostic disparities, including gender, race and ethnicity, age, SES, English proficiency, and urbanicity, and we do so across 10 model configurations and at the intersection of multiple attributes. Because the direction of LLM bias cannot be assumed a priori—models may reproduce, attenuate, or invert documented human-clinician patterns—we frame the study around 2 open research questions (RQs) rather than directional hypotheses:

RQ1: Do LLMs exhibit systematic diagnostic biases based on patient demographics when clinical content is held constant?

This question tests whether demographic cues alone are sufficient to alter diagnostic output, the central condition required to attribute bias to identity rather than presentation.

RQ2: To what extent do LLMs exhibit diagnostic consensus, and how do model architecture (family and tier) and patient demographic profiles moderate this consistency?

This question tests whether any observed bias is a general property of current models or varies with design choices and demographic context, which bears directly on whether model selection has clinical consequences.

In addressing these questions, the study makes 3 contributions. First, it applies a counterfactual matched-pair design that isolates the causal influence of demographic framing from clinical content, moving beyond qualitative audits toward quantified, attributable effects. Second, it evaluates breadth rarely examined together: 10 model configurations spanning 5 families and 2 capability tiers, 6 demographic dimensions, and their intersections, surfacing effects that main-effects-only analyses would miss. Third, it provides a reusable, version-aware auditing methodology and an initial baseline against which diagnostic bias can be tracked as models are updated.

All statistical analyses were conducted in R (version 4.5.2; R Core Team). Key packages included irr for interrater reliability, DescTools for Cochran Q test, car for model diagnostics, and tidyverse for data manipulation and visualization. Cochran rule compliance for all chi-square tests is reported in Table S3 in Multimedia Appendix 1.

Raw diagnostic outputs were harmonized to account for variations in LLM response formatting. Diagnoses were collapsed into 4 categories: AN, MDD, Avoidant/Restrictive Food Intake Disorder (ARFID), and other (eg, adjustment disorder, atypical AN, other specified feeding or eating disorder, and rare responses). Responses containing errors or missing data were flagged for sensitivity analysis.

Chi-square tests compared AN diagnosis rates between control and ambiguous vignettes to confirm the effectiveness of the control-ambiguous manipulation. Effect size was measured using Cramér V.

Chi-square tests with Monte Carlo simulation (B=5000 replicates) examined associations between each demographic variable and diagnostic outcomes. Cochran rule was verified for all tests (no expected frequency<1, no more than 20% of cells<5). Mixed-effects binary logistic regression modeled the probability of AN and MDD diagnoses as a function of demographic predictors, with LLM included as a random intercept to account for clustering of observations within models. This approach accounts for the nonindependence of diagnoses generated by the same LLM across vignettes. Odds ratios (ORs) with 95% CIs were calculated from fixed effects. Per-model logistic regressions were also conducted to assess whether bias patterns were consistent across individual LLMs or driven by outlier models. The Benjamini-Hochberg procedure was applied to control the false discovery rate across multiple comparisons.

Fleiss κ assessed diagnostic agreement across all 10 LLM configurations, with 95% CIs. Agreement was also calculated separately by condition (control vs ambiguous) and demographic subgroups. Pairwise Cohen κ quantified agreement between each LLM pair. Cochran Q test examined whether diagnosis rates differed significantly across LLMs, with post hoc pairwise McNemar tests.

Hierarchical cluster analysis (Ward D2 method and Euclidean distance) examined similarity in demographic bias profiles across LLMs based on diagnosis rates by gender and race and ethnicity.

Pooled logistic regression (without random effects) was conducted to assess sensitivity of results to the modeling approach. A post hoc temperature sensitivity analysis, in which a vignette subset was readministered at temperature 0.0 (see “Hyperparameter Settings” section), assessed whether diagnostic outputs differed based upon temperature setting. Statistical significance was set at α=.05

A total of 11,520 diagnostic outputs were generated (1152 vignettes × 10 LLM configurations). Missing or error responses were minimal (14 of 11,520 outputs; 0.12%), distributed across non-Claude models. No model exhibited systematic missingness.

Across all LLMs and conditions, the most common diagnoses were AN, MDD, and ARFID, with rare responses collapsed into an “Other” category for analysis.

The control-ambiguous manipulation was effective. Control vignettes produced near-unanimous AN diagnoses (mean 100%, SD 0.1%), with 9 of 10 LLMs achieving 99.7%‐100% AN diagnosis rates for control cases. Ambiguous vignettes elicited substantially lower AN rates (mean 23.6%, SD 10.1%), representing a 76.4 percentage point difference. This difference was statistically significant, χ²₁=7111.40, P<.001, with a large effect size (Cramér V=0.786). These results confirm that control vignettes functioned as unambiguous AN cases, while ambiguous vignettes successfully induced diagnostic variability.

Across the 10 LLMs, AN diagnosis rates for ambiguous vignettes ranged from 12.9% (Claude Advanced) to 45.6% (Claude Base), while MDD rates ranged from 24.6% (Llama Advanced) to 62.2% (Llama Base; Table 2).

All analyses of ambiguous vignettes used mixed-effects binary logistic regression with LLM as a random intercept to account for clustering of observations within models. Because clinical content was held constant across demographic conditions, any significant deviation from an OR of 1.0 is consistent with the influence of demographic framing on diagnostic output rather than differences in clinical presentation. Percentages denote observed diagnosis rates pooled across all 10 configurations on ambiguous vignettes; because the ORs derive from the mixed-effects model adjusting for the remaining predictors and between-model clustering, the two are directionally consistent but not arithmetically equivalent.

For AN diagnoses, Asian patients were significantly more likely than White patients to receive an AN diagnosis (46.4% vs 25.4% of ambiguous vignettes; OR 2.88, 95% CI 2.44‐3.42), while Latine patients (7.4% vs 25.4%; OR 0.20, 95% CI 0.16‐0.26), Black patients (15.6% vs 25.4%; OR 0.51, 95% CI 0.42‐0.62), and female patients (17.5% vs 29.8% for males; OR 0.43, 95% CI 0.37‐0.49) were less likely. Low-SES patients were also less likely to receive an AN diagnosis than high-SES patients (20.1% vs 27.2%; OR 0.61, 95% CI 0.53‐0.70). For MDD diagnoses, Latine patients were over 9 times more likely than White patients to receive an MDD diagnosis (71% vs 24.8% of ambiguous vignettes; OR 9.57, 95% CI 8.00‐11.45), Black patients were over 6 times more likely (62% vs 24.8%; OR 6.09, 95% CI 5.13‐7.24), and low-SES patients were more likely than high-SES patients (46.1% vs 40.4%; OR 1.41, 95% CI 1.24‐1.59). Female patients were less likely than males to receive MDD diagnoses (37.4% vs 49.1%; OR 0.50, 95% CI 0.44‐0.57). The LLM random intercept variance was substantial for both AN and MDD, confirming significant between-model variability in baseline diagnostic rates. Per-model logistic regressions confirmed that racial bias in MDD diagnosis was present across all 10 LLM configurations, with Black-versus-White ORs ranging from 2.69 (Llama Base) to 11.72 (Mistral Advanced). Full per-model logistic regression results for AN and MDD are presented in Tables S1 and S2 in Multimedia Appendix 1. Per-model diagnosis distributions and demographic breakdowns are presented in Figures S1-S30 in Multimedia Appendix 2.

Figure 2 displays the ORs and 95% CIs for all demographic predictors from the mixed-effects models.

Full mixed-effects model results, including all demographic predictors for both outcomes, are presented in Table 3.

The gender × race or ethnicity interaction was significant for both AN (χ²₃=229.08, P<.001) and MDD (χ²₃=1495.95, P<.001). For AN, the interaction was driven by Latine patients, for whom the gender effect was substantially larger than for other racial or ethnic groups (interaction OR 52.99, 95% CI 23.29‐120.58; P<.001), indicating that the low AN diagnosis rate among Latine patients was concentrated among males. Raw cell frequencies contextualize this interaction: Latine males received AN diagnoses in only 7 of 720 cases (1%), compared to 99 of 720 (13.8%) for Latine females, a reversal of the pattern observed in all other racial or ethnic groups, where males consistently received more AN diagnoses than females. This pattern is consistent with compounding demographic heuristics in which the combination of Latine ethnicity and male gender was treated as maximally incompatible with AN. This unusually large estimate reflects extreme cell sparsity and should be interpreted as indicating direction and magnitude rather than precise effect size. Additionally, White female patients received zero MDD diagnoses across all 10 LLM configurations (0 of 720 cases), contributing to the quasi-complete separation observed in the MDD interaction model. For MDD, although the omnibus interaction was significant, individual interaction coefficients exhibited quasi-complete separation due to multiple models producing zero diagnoses for specific gender × race combinations, precluding reliable estimation of cell-level interaction effects.

More generally, the gender × race or ethnicity interaction models were affected by quasi-complete separation arising from zero-count cells (eg, White female patients received no MDD diagnoses across all configurations), which inflates point estimates and widens CIs. The interaction estimates—most notably the Latine × male effect for AN (OR 52.99, 95% CI 23.29‐120.58), whose interval width itself reflects this instability—should therefore be read as indicating the direction and approximate magnitude of an effect rather than precise values. We emphasize that this instability is confined to the sparse interaction terms; the main-effect ORs reported above, which are estimated from the full sample, are not subject to separation and are stable.

Full interaction model results are presented in Table 4. Per-model logistic regression results for AN and MDD are available in Tables S1 and S2 in Multimedia Appendix 1.

For ambiguous vignettes, Fleiss κ indicated moderate agreement (Fleiss κ=0.410, 95% CI 0.397‐0.422). Agreement varied by demographic subgroup, with highest agreement for Latine patients (Fleiss κ=0.477) and lowest for Asian patients (Fleiss κ=0.127). Agreement was higher for female patients (Fleiss κ=0.453) than male patients (Fleiss κ=0.307; Table 5). Pairwise Cohen κ ranged from 0.010 (Mistral Advanced vs Mistral Base) to 0.835 (DeepSeek Advanced vs ChatGPT Advanced), and models within the same family did not consistently show higher agreement than models across families (Table 6).

Cochran Q tests revealed significant heterogeneity in diagnosis rates across LLMs for both AN (Q₉=388.44, P<.001; range 12.9%‐45.6%) and MDD (Q₉=764.59, P<.001; range 24.6%‐62.2%). Post hoc pairwise McNemar tests identified 37 of 45 pairwise comparisons as significantly different after Benjamini-Hochberg correction. Notably, the largest within-family divergences occurred for Claude (AN rates: 45.6% base vs 12.9% advanced) and Llama (MDD rates: 62.2% base vs 24.6% advanced), suggesting that fine-tuning and alignment procedures substantially alter diagnostic behavior. All 45 pairwise McNemar post hoc comparisons are reported in Table S4 in Multimedia Appendix 1.

Figure 3 illustrates the per-model ORs for race and ethnicity effects on AN diagnosis, revealing substantial heterogeneity across LLMs.

Pairwise Cohen κ values ranged from 0.010 (Mistral Base vs Mistral Advanced) to 0.835 (ChatGPT Advanced vs DeepSeek Advanced; Table 6).

Pooled logistic regression without random effects produced consistent patterns of significance, with odds ratios directionally aligned with the mixed-effects estimates, confirming that findings are robust to the modeling approach.

Before interpreting these findings, it is worth delimiting the scope of the claims they support. The results establish that demographic bias arises in model outputs under the conditions tested here: across 10 configurations spanning 5 model families, when ambiguous eating-disorder vignettes were presented under a forced single-diagnosis format with clinical content held constant. The consistency of the racial effects across all 10 configurations supports generalization across contemporary models rather than implicating any single system. They do not, however, establish how these or other models would behave under different prompting strategies, in multiturn or open-ended clinical interactions, or for diagnostic scenarios beyond the AN presentations examined here. We therefore frame the following discussion as characterizing demographic bias under a specific, controlled evaluation paradigm, and distinguish this from broader claims about LLM behavior across the full range of clinical use, which the present design cannot adjudicate.

This study examined whether LLMs exhibit systematic diagnostic biases when evaluating clinical vignettes for eating disorders. Two key findings emerged. First, under the tested conditions, LLMs demonstrated demographic biases in psychiatric diagnosis, with Latine patients over 9 times more likely than White patients to receive an MDD diagnosis (OR 9.57, 95% CI 8.00–11.45), Black patients over 6 times more likely (OR 6.09, 95% CI 5.13–7.24), and Asian patients nearly 3 times more likely to receive an AN diagnosis (OR 2.88, 95% CI 2.44–3.42) when clinical presentations were identical. These effects were consistent across the majority of individual models, with racial bias in MDD diagnosis present in all 10 LLM configurations tested.

Second, intermodel agreement was only moderate for ambiguous cases (Fleiss κ=0.410, 95% CI 0.397‐0.422), indicating substantial variability in diagnostic behavior across LLM configurations. This variability was pronounced even within model families: Claude Base and Claude Advanced produced AN diagnosis rates of 45.6% and 12.9%, respectively, while Llama Base and Llama Advanced produced MDD rates of 62.2% and 24.6%, indicating that tier and alignment choices within a single family can substantially alter diagnostic behavior.

Within the experimental setting, these input-output effects can be attributed to demographic framing by virtue of the matched-pair design, though the internal mechanisms that generate them lie beyond what the design can resolve. Rather than indicating LLMs as unsuitable for clinical contexts, these findings reveal specific, measurable patterns that can inform targeted improvements to training procedures, model architectures, and deployment frameworks. In principle, LLMs could be developed not merely to replicate human decision-making but to improve upon it, although this study did not compare model and clinician performance and so cannot establish that current models approach, let alone exceed, human judgment. If human clinicians exhibit diagnostic biases, as decades of research confirm [19,21], the goal should be developing AI systems that actively mitigate against these biases rather than perpetuating them. By making bias patterns visible and quantifiable, this research provides a foundation for working toward more equitable systems. Realizing that potential, however, would require bias-mitigation strategies that do not yet exist; as tested, the models examined here reproduced rather than reduced demographic disparities.

All ambiguous vignettes were constructed to present clinical features primarily consistent with AN while permitting plausible differential diagnosis, mirroring the diagnostic uncertainty encountered in real-world clinical practice. Systematic demographic shifts toward MDD in these cases therefore are consistent with systematic influence of patient identity on diagnostic output rather than differences in clinical presentation.

A note on inference is warranted before considering mechanisms. Because clinical content was held constant within each matched pair and only the demographic header varied, the design supports inference of a causal effect of demographic cues on the observed diagnostic shifts within the experimental setting: the change in output can be ascribed to the change in patient identity rather than to differences in clinical presentation. It does not, however, identify the internal processes through which models produce these shifts. The mechanisms discussed in this section are therefore advanced as hypotheses consistent with the observed input-output relationships, not as pathways the present design can directly observe or confirm. The two most probable sources of the demographic biases observed are: (1) training data composition and (2) algorithmic architecture and optimization processes. Both represent modifiable factors rather than inherent limitations of the technology.

Concerning training data, LLMs learn statistical associations from text corpora that reflect historical patterns of diagnosis and treatment [17]. If Black patients have been disproportionately diagnosed with depression in clinical literature, case studies, and electronic health records, models may learn these associations regardless of symptom presentation. An analogous mechanism may operate here, paralleling the finding by Obermeyer et al [11] that health care algorithms trained on historical data encoded racial bias through proxy variables. The critical insight from that work, however, was that once identified, such biases could be substantially reduced through targeted data curation and algorithmic adjustment. The same principle applies here: if training data composition contributes to diagnostic bias, this suggests clear intervention points

These findings underscore the need for closer attention to training data in clinical AI development. Current LLM training pipelines prioritize scale, drawing from vast internet corpora with limited curation for clinical accuracy or demographic balance. The biases observed in this study are a plausible consequence of this approach: models trained predominantly on text reflecting historical diagnostic patterns may reproduce those patterns, even when they conflict with evidence-based clinical standards. Moving forward, developers of clinical AI systems should prioritize training data quality over quantity, implementing systematic audits for demographic representation, stereotype content, and alignment with current diagnostic criteria. The goal should not be models that mirror the clinical literature as it exists, but models that reflect clinical best practices as they should be applied.

The finding that Asian patients were significantly more likely to receive AN diagnoses is less readily explained by clinical stereotypes and illustrates how training data can encode both underrepresentation and misrepresentation. Research has documented eating pathology among Asian and Asian American populations [30,31]; however, the dominant narrative associates AN with White patients [20], who comprise approximately 70% of eating disorder research samples [32], which reinforces imbalanced learned associations. Alternatively, model outputs may reflect cultural stereotypes present in training text [33]. LLMs trained on broad internet corpora may encode associations between Asian identity and thinness or dietary restriction that, while not originating from clinical literature, may influence diagnostic reasoning when demographic cues are present. Either mechanism suggests that careful auditing of training data for demographic balance and stereotype content could mitigate these patterns. The socioeconomic bias observed, in which low-SES patients were less likely to receive AN diagnoses (OR 0.61, 95% CI 0.53–0.70) and more likely to receive MDD diagnoses (OR 1.41, 95% CI 1.24–1.59), mirrors a well-documented clinical stereotype in which eating disorders are perceived as conditions affecting predominantly affluent individuals [20]. This finding is notable because SES was conveyed through a single sentence in the vignette header, yet was sufficient to shift diagnostic output. If LLMs encode the assumption that low-SES patients do not develop eating disorders, they risk replicating the same access and detection barriers that already contribute to underdiagnosis in lower-income populations.

Algorithmic factors, including attention mechanisms, reinforcement learning from human feedback (RLHF), and optimization objectives, may also contribute to observed biases [33]. Notably, RLHF relies on human annotators whose judgments may encode the same biases present in clinical practice. If annotators rate outputs as “higher quality” when they align with stereotypical expectations, the model may learn to reproduce those stereotypes. This observation points to an underexplored opportunity: RLHF could be deliberately designed to penalize demographically biased outputs rather than inadvertently rewarding them. Such an approach could position LLMs not as mirrors of human judgment but as tools calibrated to counteract known human biases.

The finding that female patients were less likely than male patients to receive both AN (OR 0.43, 95% CI 0.37–0.49) and MDD (OR 0.50, 95% CI 0.44–0.57) diagnoses warrants particular attention, as it partially contradicts patterns documented in the human clinician literature. Research suggests that clinicians overdiagnose depression in women [21] and that the prototypical AN patient is female [20], yet LLMs in this study favored neither diagnosis for female patients. Several factors may explain this divergence. First, the forced single-diagnosis paradigm creates a zero-sum dynamic: because each vignette produces exactly one diagnosis, racial and ethnic effects may dominate the diagnostic output, effectively suppressing gender effects that would emerge in a multidiagnosis or differential-diagnosis format. The gender × race or ethnicity interaction for AN supports this interpretation: the reduced AN diagnosis rate for females was consistent across White, Asian, and Black patients, but reversed entirely for Latine patients (interaction OR 52.99, 95% CI 23.29‐120.58; this estimate is imprecise owing to cell sparsity, so the directional reversal, rather than the exact magnitude, is the interpretable result), for whom males received virtually no AN diagnoses. This pattern suggests that LLMs applied compounding demographic heuristics in which Latine ethnicity and male gender may have jointly suppressed AN diagnosis. More broadly, because this forced-choice format compels a categorical diagnosis rather than permitting graded or differential reasoning, it may amplify the apparent magnitude of demographic differences relative to naturalistic settings; the direction of the observed effects is unlikely to be an artifact of the constraint, since it is held identical within every matched pair, but the absolute magnitudes should be interpreted as specific to the forced-choice paradigm and may not transfer directly to contexts that permit differential diagnosis.

Second, LLM training data may encode a corrective signal. Research on RLHF’s effects on demographic representation suggests a possible mechanism: studies comparing pre- and postalignment models have found that RLHF can overcorrect for gender representation, producing outputs that overrepresent female perspectives while leaving underlying stereotypical associations intact [34]. If similar overcorrection operates in clinical contexts, RLHF procedures designed to reduce gender stereotyping may suppress clinically valid gender-linked base rates, producing outputs that are demographically neutral at the cost of diagnostic sensitivity. This hypothesis warrants direct testing by comparing base and RLHF-aligned versions of the same model on clinical tasks. More broadly, awareness of gender bias in psychiatric diagnosis has generated substantial commentary in clinical and popular literature, and models trained on this text may have learned to counteract rather than reproduce gender stereotypes, even while remaining susceptible to racial biases that have received comparatively less corrective attention in training corpora. Third, the per-model results revealed that the direction of gender bias was not consistent across LLMs: Llama Base and Mistral Base diagnosed females with AN at higher rates than males, while most other models showed the opposite pattern. This inconsistency suggests that gender bias in LLMs is more labile than racial bias, which was directionally consistent across all 10 configurations.

The moderate intermodel agreement for ambiguous vignettes (Fleiss κ=0.410) reflects differences in training corpora, optimization objectives, and alignment procedures across model families. Rather than viewing this variation as problematic, it offers valuable information about which approaches produce more or less biased outputs. AN diagnosis rates on ambiguous vignettes ranged from 12.9% (Claude Advanced) to 45.6% (Claude Base), while MDD rates ranged from 24.6% (Llama Advanced) to 62.2% (Llama Base), demonstrating that model behavior is highly sensitive to design choices, suggesting substantial room for optimization. This pattern echoes prior diagnostic-classification work: when ChatGPT, Claude, and Gemini were evaluated on identifying MDD from clinical vignettes, accuracy ranged from near chance to near-perfect across families [6], indicating that wide cross-model divergence on vignette-based psychiatric classification is not unique to the present design.

The lowest agreement occurred for Asian patients (Fleiss κ=0.127), revealing that current models lack stable, reliable representations for this demographic group in eating disorder contexts. This pattern likely reflects heterogeneity or sparsity in how Asian populations appear across training corpora. From an improvement perspective, this finding identifies a specific gap that data augmentation, targeted fine-tuning, or ensemble approaches might address. Notably, the variation itself is informative: some models performed better than others for this subgroup, and understanding what differentiates higher-performing models could guide development of more equitable systems.

Within-family divergence was particularly striking. Claude Base and Claude Advanced produced AN diagnosis rates of 45.6% and 12.9%, respectively, while Mistral Base and Mistral Advanced showed near-zero pairwise agreement (Cohen κ=0.010). These patterns indicate that posttraining alignment and fine-tuning procedures do not merely refine base model behavior but can fundamentally alter diagnostic profiles. This finding has direct implications for clinical deployment: selecting a “more advanced” version of a model does not guarantee reduced bias, and may in some cases introduce new bias patterns. More broadly, these within-family divergences, like the intersectional effects reported above, would be invisible in analyses examining only main effects, underscoring the importance of disaggregated evaluation in clinical AI auditing.

Several limitations contextualize these findings. First, this study used text-based vignettes rather than real clinical encounters, which simplifies the complexity of psychiatric assessment. Actual diagnostic processes involve nonverbal cues, longitudinal patient history, and clinical intuition that vignettes cannot capture. While vignette-based methods are well-established in bias research [26], findings should be validated in more naturalistic settings as LLMs are integrated into clinical workflows. Additionally, the forced single-diagnosis prompt format (“respond with just the diagnosis name”) does not reflect how clinicians typically use LLMs in practice, where open-ended queries, differential diagnoses, and chain-of-thought reasoning are more common. That said, forced single-diagnosis outputs are not uncommon in applied settings, where automated intake systems, triage chatbots, and clinical decision support tools frequently require categorical classification. This constraint may amplify observed biases by eliminating the hedging and differential reasoning that LLMs produce in more naturalistic contexts; the present design should therefore be read as a conservative test of whether demographic information influences diagnostic output rather than a simulation of clinical LLM use. Future research using differential-diagnosis prompts, in which models list multiple possibilities with associated confidence, could assess whether bias attenuates once the forced-choice constraint is removed.

Second, all vignettes were derived from a single target condition, AN, and although the ambiguous presentations admitted a broader differential (most often MDD, but also ARFID and other feeding and eating disorders), findings may not generalize to other psychiatric conditions or clinical contexts. The magnitude and direction of bias likely vary across conditions with different demographic distributions in training data, and comprehensive auditing should span multiple diagnostic categories.

Third, the data were collected across 2 periods, and the configurations were not uniformly version-stable. The Claude models were recollected in February 2026 to correct a temperature parameter error identified during manuscript preparation; both were accessed via version-pinned end points (claude-3-haiku-20240307 and claude-sonnet-4‐20250514), so the underlying weights were identical regardless of query date, although provider-side changes to system prompts or safety filters between October 2025 and February 2026 cannot be excluded. The consistency of bias patterns across the 8 models collected in October 2025 suggests the gap had minimal impact on overall findings. One model, Mistral Advanced, was accessed via a non–version-pinned end point (mistral-large-latest), so its underlying version cannot be retrospectively confirmed.

Fourth, the models evaluated represent a snapshot of a rapidly changing field, and the set tested did not include every system available during the collection window. The GPT-5 family had been released before data collection began but was not incorporated, because the data-collection pipeline was implemented and run on the prior generation of GPT and was not updated to add GPT-5 before collection concluded; as the evaluation could not be rerun within the study timeline, GPT-5 is absent from the present analysis. Google’s Gemini models were likewise not tested, owing to an access constraint that prevented reliable programmatic data collection during the study window rather than to a selection preference. The study’s emphasis on widely used, lower-cost configurations (see “Selection Rationale” section) governs which models were actively sought, but it does not account for the GPT-5 omission, which represents a limitation of the data-collection process. Future evaluations should incorporate more recent frontier systems to determine whether the observed biases persist. Model behavior also evolves as providers retrain and update systems: version-pinned end points can be retired (as has since occurred for the Claude configurations used here), unpinned end points can change without notice, and provider-side updates can alter outputs even when weights are nominally fixed. Results should therefore be read as characterizing these models at the time of testing, and replication on current systems is needed to determine whether the observed biases persist, attenuate, or shift. This transience also motivates the repeated, version-aware auditing the present methodology supports, offering a way to track whether demographic bias decreases as models are updated.

Fifth, the moderate intermodel agreement for ambiguous vignettes (Fleiss κ=0.410, 95% CI 0.397‐0.422) warrants careful interpretation. This statistic indexes agreement among the 10 configurations, not the diagnostic competence of any individual model, so low agreement is not itself evidence of poor performance. Agreement on the control vignettes was near-unanimous, indicating that the models classified unambiguous cases consistently; divergence was confined to the ambiguous condition, which was deliberately designed to permit several defensible differential diagnoses. Under such conditions, disagreement across models is expected rather than anomalous—much as clinicians frequently disagree on ambiguous cases—and the cross-model variability is better understood as a substantive finding (RQ2) than a deficiency. We nonetheless acknowledge that this variability means model choice has direct consequences for diagnostic output, reinforcing the need for model-specific validation before clinical deployment.

Finally, the 2 model families that assisted in formulating candidate vignette phrasings (ChatGPT and Claude) were also among those evaluated; for GPT-4o, the identical model served both roles, whereas the Claude version used during formulation (Claude Sonnet 4.5) differed from the tested Claude configurations (Claude 3 Haiku and Claude Sonnet 4). Because the final vignettes were validated by the research team and assembled by a deterministic script rather than generated by a model at test time, no configuration was evaluated on text it authored; nonetheless, subtle stylistic familiarity effects cannot be fully excluded, and future replications could formulate stimuli independently of any tested model. The study also did not examine the reasoning processes underlying diagnostic decisions, which future work using chain-of-thought or rationale-elicitation prompts could address.