Physicians and artificial intelligence diverge in evaluating large language models on real clinical cases

Abstract

While multimodal large language models (LLMs) demonstrate significant potential in healthcare applications, their clinical utility is difficult to appraise. Current evaluations of medical-assisting LLMs are often limited by sparse human expertise, narrow specialty scope, and reliance on multiple-choice benchmarks or synthetic vignettes, which can inflate performance and obscure clinical utility. We conducted a multicenter, multidisciplinary study in which more than 400 physicians—spanning seven specialties, varied experience levels, and multiple geographic settings—evaluated LLM-generated free-text responses to real, de-identified clinical cases. In a matched-control design, we also deployed an equivalent number of AI agents configured to mirror physician characteristics to examine whether automated evaluators can supplement or replace human assessment. Our results demonstrated that physician assessments exhibited substantial heterogeneity by clinical seniority and practice environment, leading to notable shifts in relative model rankings across cohorts. While AI agents delivered highly efficient, directionally aligned assessments, they did not fully capture the nuances of human clinical judgment and could not substitute for physician-centered evaluation. Instead, they promise assistive tools that can triage or pre-screen outputs to reduce human burden.

Introduction

Generative artificial intelligence (AI), particularly multimodal large language models (LLMs) capable of processing and generating text, image, and other data modalities, has accelerated adoption of AI in healthcare^1,2, moving beyond the traditional predictive analytics toward assistant-like systems that interact with clinicians and patients³. These models hold great promise as transformative tools in the medical field^4,5, ranging from automating administrative tasks to augmenting clinical decision making^6–8. However, rigorous evaluation has revealed fundamental limitations that must be addressed before their widespread and safe clinical deployment^9–13. To date, medical evaluation of LLMs has centered largely on medical question answering (MQA), leveraging standardized multiple-choice benchmarks such as MedQA (USMLE) and PubMedQA to quantify knowledge^14,15. Remarkably, 84.2% of the research in this field is dedicated to refining QA capabilities¹³. This paradigm enables rapid, low-cost model comparisons but departs from clinical reality. Multiple-choice tests do not reflect how clinicians synthesize unstructured information into free-text assessments, differentials, and management plans. They can overstate performance, obscure reasoning deficits, and miss safety-critical errors^16–20. As LLMs approach high-stakes use, evaluations should prioritize free-response performance on real, de-identified clinical cases.

Concurrently, there is growing recognition that evaluations should be human-centric. Physicians bring experiential knowledge, contextual judgment, and risk calculus that current automated metrics and proxy evaluators often fail to capture^5,6,21–24. Yet large-scale human evaluation faces practical barriers: constrained clinician availability, the diversity of specialties and care settings, and heterogeneity in training and clinical priorities^25–29. This leaves the crucial task of selecting and standardizing expert evaluators an unresolved challenge.

To address the scalability constraints, some have proposed using autonomous AI agents as proxy evaluators³⁰. While agents can be efficient and have shown utility in other domains^31,32, it remains unclear whether they can reliably emulate the nuanced assessments of practicing clinicians—especially for free-text clinical reasoning, safety appraisal, and contextual appropriateness. Overreliance on automated judges’ risks mischaracterizing model readiness and overlooking safety-relevant failure modes.

In this study, we aimed to determine whether AI agents can approximate physician-centered, multidimensional evaluation of large language model (LLM)-generated clinical case interpretations under realistic clinical conditions (see Fig. 1). To address this question, we conducted a multicenter evaluation involving more than 400 physicians across seven specialties and diverse practice environments. In parallel, we deployed a matched cohort of AI agents configured to simulate physician roles. Secondary exploratory analyses examined whether evaluation patterns varied across practice environments and assessed comparative time efficiency between human and agent evaluators. Instead of multiple-choice questions, we use real, de-identified clinical cases presented as case summaries without predefined options, and we assess free-text LLM outputs across key dimensions of clinical quality. Additionally, we conduct a multicenter, multidisciplinary evaluation with >400 physicians spanning seven specialties, varied seniority levels, and diverse geographic practice settings. In parallel, we deploy a matched set of AI agents configured to mirror physician characteristics to test whether automated approaches can supplement—or substitute for—human evaluations at scale. To examine whether physician-centered evaluation remains stable across diverse clinical environments, we included participants from institutions with differing levels of healthcare resources and practice infrastructures, including high-altitude centers that operate under distinct clinical service conditions.

Fig. 1. Schematic of a multicenter, survey-based observational study.

This study encompassed seven medical specialties, evaluated five advanced large language models, involved >400 physicians, and assessed six domains of preference evaluation.

Results

To characterize physician-centered evaluation under realistic clinical conditions, we first analyzed physician rankings of five LLM-generated interpretations across six predefined quality dimensions. Physicians exhibited systematic, structured evaluation patterns, with preferences that varied by clinical seniority (Fig. 2). This experience-related variability indicates that physician evaluation is not monolithic and underscores why approximating physician-centered assessment is a non-trivial target for automated evaluators. Across the six dimensions, junior physicians tended to favor Qwen-Max (the highest mean score, 51.8, 95% CI 44.5–59.0) consistently, whereas senior physicians demonstrated more differentiated preferences, frequently assigning higher relative rankings to GPT-4o and DeepSeek-R1 in dimensions related to coherence, accuracy, and clinical usefulness (Fig. 2). Notably, the identity of the top-ranked model was not stable across seniority strata, suggesting that evaluative emphasis shifts with clinical experience rather than reflecting random dispersion. In contrast, Claude 3.5 and OpenAI o1 were consistently ranked lower across experience levels, with OpenAI o1 (with mean scores typically below 45) demonstrating persistently reduced relative scores across multiple dimensions. Collectively, these findings indicate that clinical seniority shapes multidimensional evaluation priorities and that physician-centered judgment exhibits structured heterogeneity rather than consensus uniformity. Unless otherwise specified, analyses reflect n = 421 physicians evaluating seven cases, with five model outputs per case across six evaluation dimensions.

Fig. 2. Overall preference of Human doctors’ evaluators.

Details of physician preferences for MLLM (Qwen-Max, DeepSeek-R1, GPT-4o, Claude 3.5, and OpenAI o1) across six performance dimensions (comprehensiveness, coherence, accuracy, helpfulness, harmlessness, and structure) among doctors with varying experience levels (junior, middle career, senior).

In an exploratory robustness analysis, physician evaluation patterns differed across heterogeneous practice environments (high-altitude versus plain-region institutions; Fig. 3), suggesting that service context may influence evaluative priorities. Across multiple dimensions, regional differences reached statistical significance, with directional inversions observed for several models. For example, plain-region physicians more frequently favored GPT-4o and Qwen-Max in dimensions related to comprehensiveness and helpfulness, whereas high-altitude physicians assigned relatively higher rankings to OpenAI o1 across several facets. These differences were not confined to a single evaluation dimension and indicate that clinical context may shape what clinicians prioritize when judging LLM-generated outputs.

Fig. 3. Preference divergence of doctors from different regions.

Details of physician preferences for MLLM (Qwen-Max, DeepSeek-R1, GPT-4o, Claude 3.5, and OpenAI o1) across six performance dimensions (comprehensiveness, coherence, accuracy, helpfulness, harmlessness, and structure) among doctors with different regions (high-altitude region and plain region).

We evaluated whether AI agents can approximate physician-centered evaluation by quantifying alignment between agent-assigned and physician-assigned scores across dimensions (Fig. 4). Across the six evaluation dimensions, AI agents showed partial directional alignment with physician preferences, but limited absolute agreement at the higher-ranked models in this evaluation, indicating that agents cannot be treated as perfectly interchangeable with physician evaluators. Figure 4 demonstrates distinct preference patterns between AI Agent doctors and human doctors when evaluating five LLMs across six assessment dimensions. For comprehensiveness, human doctors showed a higher preference for GPT-4o (n = 125, mean score 74, 95% CI 69.4–78.6), while AI Agent doctors strongly favored OpenAI o1 with significantly higher scores (n = 106, mean score 76.9, 95% CI 72.8–81.1, $p<0.001$). DeepSeek-R1 received comparable preference rates between both groups but with higher scores from AI Agents (n = 119, mean score 65.3, 95% CI 59.8-70.9, $p<0.05$). In coherence evaluation, similar patterns emerged with human doctors preferring GPT-4o (n = 122, mean score 74.8, 95% CI 69.9–79.7) and DeepSeek-R1 (n = 123, mean score 73.8, 95% CI 59.8–70.9), while AI Agents showed a significantly higher preference for OpenAI o1 (n = 106, mean score 76.6, 95% CI 72.4–80.8, $p<0.001$). The accuracy dimension revealed that human doctors predominantly selected GPT-4o (n = 122, mean score 75.8, 95% CI 71.1–80.6, $p<0.05$), whereas AI Agents demonstrated a strong preference for OpenAI o1 with notably higher scores (n = 108, mean score 81.5, 95% CI 77.1–85.9, $p<0.001$). For helpfulness, human doctors maintained their preference for DeepSeek-R1 (n = 128, mean score 73.2, 95% CI 69.4–77.1) and GPT-4o (n = 125, mean score 73, 95% CI 68.8–77.8), while AI Agents consistently favored OpenAI o1 (n = 104, mean score 77.9, 95% CI 73.5-82.3, $p<0.001$). Similar divergent patterns were observed in harmlessness and structure dimensions, where AI Agents consistently rated OpenAI o1 significantly higher than human doctors ($p<0.001$ and $p<0.01$, respectively), while human doctors showed stronger preference for DeepSeek-R1 and GPT-4o in structural quality ($p<0.01$ and $p<0.001$, respectively). Notably, Claude 3.5 consistently received the lowest preference rates and scores from both AI Agent and human doctors across all six dimensions, suggesting unanimous agreement between both evaluator types regarding its relatively inferior performance in this specific medical context.

Fig. 4. Overall preference of AI agent evaluators and human evaluators.

Details of physician preferences for MLLM (Qwen-Max, DeepSeek-R1, GPT-4o, Claude 3.5, and OpenAI o1) across six performance dimensions (comprehensiveness, coherence, accuracy, helpfulness, harmlessness, and structure) among doctors with corresponding agents.

To directly evaluate whether AI agents approximate physician-centered assessment beyond these specific top-model preferences, we conducted a global rank concordance analysis across all models and evaluation dimensions using Kendall’s coefficient of concordance (Fig. 5). Despite the specific divergences noted at the top of the rankings, a substantial level of overall concordance was observed between AI agents and physicians ($W=0.798,p=2.19\times {10}^{-2}$). Scatter plots with regression trends across all six evaluation dimensions demonstrate a broadly similar rank-ordering of model performance between the two evaluator types. These findings indicate that while AI agents exhibit distinct top-choice preferences and do not perfectly mirror absolute human scoring nuances, their evaluations remain highly directionally aligned with human clinical judgment across the broader spectrum of model performance.

Fig. 5. Concordance analysis between AI-agent and human evaluator scores across six assessment dimensions.

Scatter plots with fitted regression lines compare mean scores assigned by AI agents and human physicians across six dimensions (comprehensiveness, coherence, accuracy, helpfulness, harmlessness, and structure), with Kendall’s W indicating overall agreement and highlighting moderate alignment alongside dimension-specific variability.

We further examined within-group consistency of rankings. AI agents produced highly stable ranking patterns across simulated experience profiles, whereas physician rankings varied substantially across seniority strata (Fig. 6). Among AI Agent doctors, remarkable consistency was observed across senior, middle career, and junior levels, with 23 out of 30 ranking pairs (76.7%) maintaining identical rankings regardless of experience level. OpenAI o1 consistently ranked first across all dimensions and experience levels, while DeepSeek-R1 uniformly ranked second. Only minor variations appeared in seven pairs (23.3%), such as the tied rankings between DeepSeek-R1 and GPT-4o in coherence evaluation by middle career physicians, and between GPT-4o and Claude 3.5 in structure assessment by junior doctors. In stark contrast, human doctors demonstrated substantial heterogeneity in their rankings across experience levels, with only 1 out of 30 pairs (3.3%) showing complete consistency—Claude 3.5 maintaining fifth rank in helpfulness across all groups. The remaining 29 pairs (96.7%) exhibited varying degrees of ranking disagreement. Notably, Qwen-Max rankings fluctuated between first and fifth place depending on experience level and evaluation dimension, while GPT-4o showed rankings ranging from first to fifth across different doctor groups. Human doctors of senior group tended to favor Qwen-Max for comprehensiveness and coherence, while junior doctors showed more diverse preferences, often ranking Claude 3.5 and OpenAI o1 higher than their senior colleagues. These findings suggest that AI Agent doctors exhibit algorithmic consistency in their evaluation patterns regardless of simulated experience levels. These findings indicate that while AI agents capture broad evaluative trends, they do not achieve sufficient agreement to function as substitutes for physician-centered assessment.

Fig. 6. Preference divergence of AI agent and human doctor with different years of service experience.

a Aggregated rankings of five different AI models (Qwen-Max, DeepSeek-R1, GPT-4o, Claude 3.5, and OpenAI o1) across six performance criteria, as evaluated by AI Agent Doctors simulating varying years of experience (Junior, Middle career, Senior). Rankings were assigned on a scale of 1 to 5, where 1 indicates the highest-ranked performance and 5 indicates the lowest-ranked performance within each group. Colors correspond to rank, with Yellow indicating higher performance (rank 1) and Blue indicating lower performance (rank 5). b Aggregated rankings of five different AI models (Qwen-Max, DeepSeek-R1, GPT-4o, Claude 3.5, and OpenAI o1) across six performance criteria, as evaluated by human doctors with varying years of experience (Junior, Middle career, Senior).

As an exploratory efficiency analysis, evaluation time varied across specialties and physician seniority levels (Fig. S3). Across all specialties, senior physicians required the longest evaluation times, followed by middle-career and junior physicians, with this hierarchical pattern reaching statistical significance. Evaluation of cases within Oncology demanded the longest average time, approximately 135 minutes per case. Conversely, cases in Ophthalmology exhibited the shortest evaluation duration, averaging around 95 minutes. In contrast, AI agents completed identical evaluation tasks with minimal computational time relative to human assessors. These findings highlight substantial efficiency differences between human and automated evaluation. Evaluators were also invited to suggest additional assessment dimensions; the distributions are shown in Supplementary Fig. S4. This exploratory analysis provides contextual perspective and does not affect the primary alignment findings.

Discussion

Our study provides evidence on the human-centric evaluation of LLM-generated clinical case interpretations, revealing two principal findings. First, physician assessments are not monolithic but are shaped by clinical seniority and institutional context, producing structured heterogeneity in model rankings. Second, while AI-agent evaluators show broad directional alignment with physician rankings, they do not replicate the nuanced variation of human clinical judgment and cannot currently substitute for physician-centered evaluation.

The variability in model rankings across physician seniority groups underscores the complexity of evaluating LLM-generated clinical case interpretations. Our findings demonstrate that clinical experience shapes how physicians rank LLM outputs in this case-note evaluation setting. Junior and senior physicians produced distinct model-ranking patterns across the six evaluation dimensions (Figs. 2 and 6). However, this study measured model preferences rather than the evaluative reasoning underlying them; the cognitive factors driving experience-associated variation remain to be investigated. Importantly, the study did not adjudicate which seniority group provided the most clinically valid assessments. Closer alignment between AI-agent evaluators and any physician subgroup should therefore not be interpreted as evidence of alignment with superior evaluation quality.

The significant differences observed between high-altitude and plain region physicians reveal another layer of complexity in evaluation of LLM-generated clinical case interpretations. Physicians from high-altitude and plain-region institutions showed different model-ranking patterns across multiple dimensions (Fig. 3). Because the study was not designed to isolate the causes of these differences, we cannot determine whether they reflect case mix, institutional practice norms, training background, language familiarity, or other unmeasured factors. These findings indicate that practice environment should be considered as a stratification variable in future LLM evaluation studies.

AI agents consistently assigned the highest rankings to OpenAI o1 across all six dimensions, whereas physicians favored DeepSeek-R1 and GPT-4o—a systematic preference divergence at the top of the ranking (Fig. 5). However, this divergence was not uniform: for GPT-4o, human and AI-agent scores appeared closely aligned across several dimensions, suggesting that concordance varies by model and dimension. The global rank concordance analysis (W = 0.798, p = 2.19 × 10⁻²) confirmed that, across the full spectrum of models, both evaluator types maintained a broadly similar rank-ordering. These findings support a nuanced interpretation: automated evaluators capture broad relative performance but cannot be assumed to agree with physician judgment for any given model without empirical verification. Both evaluator types also showed consensus on lower-ranked models, such as Claude 3.5, and relative agreement on mid-ranked models such as Qwen-Max, reinforcing the finding that divergence was concentrated at the top of the ranking.

In contrast to physicians, AI-agent evaluators produced highly stable rankings across simulated experience profiles: 76.7% of ranking pairs were identical across seniority levels, compared with 3.3% among physicians (Fig. 6). AI agents consistently favored OpenAI o1 across all dimensions and experience configurations, whereas physician rankings of the same models varied substantially by seniority. This consistency, however, should not be equated with validity. Rather, it indicates that current AI-agent evaluators, configured as single-model instances with identical prompts and contextual parameters, do not reproduce the structured heterogeneity of physician judgment. The procedural homogeneity required for experimental control may have further contributed to convergent agent outputs, although the relative contributions of model architecture, training data, and prompt design to this consistency were not directly assessed.

Our findings align with and extend a growing body of literature focused on moving beyond simplistic benchmarks. While MedFound⁵ established a critical precedent for in-depth qualitative assessment with a panel of six physicians, our study’s scale (N > 400) allows us to characterize a crucial phenomenon that is not apparent in smaller cohorts: the profound and systematic heterogeneity of expert clinical judgment. This underscores that frameworks like HealthBench³³ introduced by OpenAI and evaluation paradigms like MedArena³⁴ aim for scalability, the composition and diversity of the human evaluation panel are as important as its size. Our results also provide context for the deployment of models like AIME^35,36, by demonstrating that evaluation panel composition shapes the conclusions drawn about LLM performance. The goal is not merely to create accurate models, but to ensure they are validated by a representative cross-section of their intended users.

Within this specific case-note evaluation task, the observed global rank concordance suggests that AI-agent evaluators may provide a preliminary signal about broad relative model performance. Whether this directional alignment translates into practical utility for preliminary screening of LLM-generated case interpretations remains untested and would require prospective, task-specific validation with safeguards against missing clinically important errors. These findings do not establish suitability for patient triage, clinical decision-making, treatment recommendation, or autonomous quality assurance. More broadly, our results underscore that meaningful validation of LLM-generated clinical case interpretations requires evaluation frameworks that are not only human-centric but also intentionally diverse, capturing the perspectives of clinicians across specialties, experience levels, and practice settings.

This study has several limitations. First, although the physician cohort was large (N = 421), the evaluation was based on seven real-world clinical cases, each corresponding to a specific specialty. Because each case was evaluated exclusively by physicians within the matching specialty, the effects of case content and evaluator specialty are confounded and cannot be disentangled without cross-specialty evaluation of the same case. The observed ranking patterns may therefore be case-sensitive and specialty-sensitive, and should be interpreted as evidence from a multicenter, multidisciplinary case-note evaluation setting rather than as a comprehensive assessment across all clinical tasks. Second, all cases and evaluations were conducted in Chinese; generalizability to other languages and clinical documentation systems requires independent validation. Third, the study compared evaluation patterns across physician seniority groups but did not establish which group provided the most clinically valid assessments. Senior clinicians may be better positioned to evaluate certain dimensions, whereas junior clinicians may apply different criteria. Closer alignment between AI-agent evaluators and any physician subgroup should not be interpreted as evidence of alignment with superior evaluation quality. Future studies should incorporate expert adjudication, predefined reference standards, or outcome-linked assessments to distinguish evaluator preference from evaluation quality. Fourth, the ordinal ranking-based scoring captures relative preference but does not measure absolute quality differences between closely ranked models. Fifth, all AI agents were powered by a single underlying LLM architecture; agent-evaluation findings may not generalize to other agent frameworks or base models. Finally, all findings are restricted to the task of evaluating LLM-generated clinical case interpretations and should not be extrapolated to predictive modeling, diagnostic decision support, treatment recommendation, or other medical LLM applications.

Automated evaluators may be useful for exploratory benchmarking or preliminary comparison of LLM-generated clinical case interpretations, but final assessment of clinical relevance, safety, and utility should remain anchored in appropriately diverse human clinical evaluation. The composition, seniority distribution, and practice-environment diversity of evaluation panels are as important as their size.

Methods

This study employed a multi-agent, multicenter evaluation framework to compare human physician and AI agent assessments of LLM-generated medical case interpretations. The design featured two parallel evaluation arms: a human cohort of 450 physicians, with 421 valid responses collected, and an AI cohort of 421 evaluator agents. Participation was voluntary and uncompensated. No financial or material incentives were provided. Both cohorts were tasked with assessing the quality of medical case reports generated by a standardized set of five high-performing LLMs. The physicians’ collective judgments served as the benchmark for evaluating the performance and alignment of the AI agents. To generate the medical interpretations for evaluation, five LLMs were selected, including GPT-4o (gpt-4o-2024-11-20), Claude 3.5 (claude-3-5-sonnet-20240620), Qwen-Max (qwen-max-0919), DeepSeek-R1 (deepseek-r1), and OpenAI o1 (o1-mini-2024-09-12), all of which consistently exhibited strong performance. To ensure comparability across models, all LLMs were evaluated under identical input conditions. A standardized prompt template was applied uniformly across all five models. Each model received the same case materials in an identical sequence. No model-specific prompt tuning or adaptive modification was introduced.

We retrospectively compiled a multicenter cohort comprising seven real-world clinical cases collected between November and December 2024 in China. All case materials were presented in Chinese, consistent with the primary clinical documentation language of the participating institutions, thereby preserving ecological validity in evaluator interpretation. Rigorous de-identification protocols were implemented, and patient data contamination was prevented by excluding any cases previously disseminated online or at academic conferences. For each case, the definitive diagnosis and reference treatment strategy were established through consensus by a panel of senior specialists, including multidisciplinary team (MDT) discussions where appropriate. The final case material provided to human doctors and agents included summarized clinical records and definitive diagnosis (ground truth) but omitted the established treatment plan to allow for unbiased assessment of the LLM-generated therapeutic suggestions. For each case, a definitive diagnosis and reference treatment strategy were established through consensus among senior specialists to ensure internal clinical validity of the case material. The definitive diagnosis was provided to both physician and AI-agent evaluators to standardize the clinical context and anchor assessment of reasoning quality relative to a confirmed ground truth. The reference treatment strategy, however, was intentionally withheld from evaluators. This design choice was made to prevent evaluators from anchoring their judgments to a predefined therapeutic plan or assessing outputs based solely on concordance with that plan. Instead, evaluators were instructed to assess the quality, coherence, accuracy, and clinical appropriateness of the LLM-generated interpretations independently. Evaluators were not tasked with generating independent diagnoses but with assessing the quality of model-generated interpretations relative to a confirmed diagnostic context.

We recruited physicians from eight medical institutions in this evaluation, and they were divided into different groups based on medical experience: Junior (1–3 years of clinical experience, n = 252), Middle career (4–14 years of clinical experience, n = 127), and Senior (over 15 years of clinical experience, n = 42) for comparison. Physicians were recruited from eight medical institutions across China through institutional collaborations and established professional clinical networks. Invitations were distributed via departmental communications and specialty-based networks to licensed physicians actively practicing in relevant clinical disciplines. Eligibility required active clinical engagement within the corresponding specialty. Participation was open during the designated recruitment period and was entirely voluntary. The recruitment strategy was designed to capture heterogeneous real-world clinical perspectives rather than a selectively incentivized sample. Before the experiment began, each doctor completed an anonymous questionnaire, including their background details such as medical title, years of practice in the corresponding specialty, and the highest academic degree obtained. Participants were assigned to evaluate cases relevant to their medical specialty. For each clinical case, five LLM-generated interpretations were produced under identical prompt conditions. Participating physicians evaluated all five interpretations corresponding to the case within their clinical specialty. To minimize positional bias, the order of the five model outputs was randomized independently for each evaluator. Evaluations were completed within a single structured session, during which physicians ranked all five model outputs across six predefined quality dimensions. Because the five interpretations referred to the same underlying case and were presented concurrently, judgments were made comparatively within-case rather than sequentially across separate time points. Case allocation followed specialty alignment rather than cross-specialty randomization, ensuring that physicians evaluated cases relevant to their clinical expertise while preserving ecological validity.

Physician assessments were conducted in designated evaluation rooms at participating institutions using a standardized digital interface developed for this study. Evaluations were completed in person rather than through remote or online submission. Internet access was not available in the evaluation setting. Participants were instructed not to communicate with other evaluators or consult external resources during assessment. Each evaluation session was limited to a maximum duration of 120 min, and all model outputs corresponding to a case were assessed consecutively within the same session. Submissions were finalized at the end of the session and could not be revised thereafter. Although continuous physical supervision was not implemented, adherence to instructions relied on professional compliance.

As a parallel evaluation arm, we constructed an agentic evaluation system (see Fig. S1) based on the CAMEL architecture³⁷. Specifically, each agent participating in the evaluation was assigned a unified task, analogous to the task of physicians evaluating AI-assisted case interpretations. AI Agent evaluations were conducted using the identical case materials and informational inputs provided to physicians. Each agent performed a single-pass evaluation per case without iterative self-review or repeated scoring, mirroring the single-submission structure of physician assessments. Each AI agent performed a single complete evaluation for each assigned case–LLM interpretation set. No repeated runs or iterative self-revision procedures were implemented. This single-pass design mirrored the human evaluation protocol, in which physicians submitted a finalized assessment without subsequent revision. And each agent was provided with configurable identity profiles (classes) and personalized settings (properties). The evaluation was conducted across six dimensions, which were developed with medical specialists following discussion with reference to previous literature^5,13. For each case and for each of the six predefined evaluation dimensions (comprehensiveness, coherence, accuracy, helpfulness, harmlessness, and structural quality), evaluators were instructed to rank the five LLM-generated interpretations from highest to lowest quality. These ordinal rankings were then converted into standardized scores of 100 (rank 1), 75 (rank 2), 50 (rank 3), 25 (rank 4), and 0 (rank 5) to enable quantitative comparison across evaluators. In addition to mean converted ranking scores, we report the frequency with which each LLM was assigned first rank within each subgroup (“mentions”), defined as the count of first-rank assignments across evaluators.

A structured evaluation rubric was provided to physicians, describing the intended meaning of each dimension to promote consistent interpretation across participants. AI-agent evaluators were configured to follow the identical ranking framework and score conversion rules. A centralized AI agent outside the evaluation system then consolidated the results, which were recorded in a structured spreadsheet for subsequent analysis. Complete information of a case (referred to as ‘Case Information’ in Fig. S1) consists of clinical records, diagnosis, and corresponding LLM-generated diagnostic reports. Both human and AI evaluators assessed the LLM-generated reports across six predefined dimensions and results from all evaluators were consolidated into a unified dataset for analysis. To better resolve the fine details of AI models preference comparison between AI Agents and human doctors, a higher-magnification image of the area marked in Fig. S1 is presented in Fig. S2.

To ensure analytical rigor and transparency, prespecified criteria were applied to define a valid response. A submission was considered valid if it met all of the following conditions:

1. Completeness: all mandatory evaluation fields were fully completed without missing primary data;

2. Internal Consistency: absence of logically contradictory scoring patterns;

3. Technical Integrity: successful finalization through the study interface without duplication, corruption, or partial saving.

Responses that were incomplete, structurally inconsistent, or duplicate entries were excluded prior to analysis. Of 450 physician submissions, 421 met these criteria and were retained for final statistical analysis.

To evaluate the overall agreement in the rank-ordering of the LLMs between human physicians and AI agent evaluators, we calculated Kendall’s coefficient of concordance (W). This non-parametric statistical test was selected to quantify the degree of global rank concordance across all five evaluated LLMs and the six assessment dimensions, independent of absolute score magnitudes. Kendall’s W values range from 0 to 1, where 0 indicates no agreement and 1 indicates perfect, unanimous agreement in rankings between the evaluator groups.

All physician evaluations were conducted under standardized conditions. Participants were instructed not to communicate with other evaluators, consult external resources, or use electronic devices during assessment. Each evaluation session was limited to a maximum duration of 120 minutes. Evaluations were completed independently and submitted as a single finalized entry without revision. Agent evaluations were conducted under identical informational conditions. Each agent performed a single-pass evaluation per case without iterative self-review or repeated scoring, mirroring the single-submission structure of physician assessments. The clinical reader study component of this research involved the participation of physicians. This study adhered to the principles outlined in the Declaration of Helsinki. Informed consent was obtained from each physician before their participation. De-identified data were handled in accordance with the approvals and requirements of the relevant institutional review boards. This study was approved by the Ethics Committee of Qinghai University Affiliated Hospital (approval no. QHU-P-SL-202219) and approved by the Ethics Committee of Beijing Tongren Hospital (approval no. TREC2025-KY106).

Author contributions

P.S., J.L., Z.Y., Y.C., and L.J. conceived and designed the study. J.L., Y.C., H.B., H.X., Y.F., T.R., J.Z., Y.D., X.H., Y.T., and Z.W. performed data collection. P.S., J.L., Y.C., Z.Y., J.Z., Y.X., and J.Z., performed data processing and experimental analysis. J.L., Z.Y., H.X., Y.F., Y.X., L.J., C.L., and M.W. interpreted experimental results and provided feedback on the study. P.S., J.L., and Z.Y. wrote the original manuscript. J.L., Y.X., J.C., Y.X., J.Z., B.T., Y.K., and P.A.K. contributed to the content revision and proofreading. H.X., L.J., C.L., J.G., and N.W. supervised the research. All authors reviewed and approved the manuscript.

Data availability

The de-identified datasets generated and/or analyzed during the current study are not publicly available because of institutional data protection policies and restrictions under ethics approvals. Access may be requested from the corresponding author Prof. Ningli Wang at wningli@vip.163.com and will be considered on reasonable request, subject to institutional approval and applicable data-use agreements.

Code availability

The code used for this project is available at https://github.com/payen-shi/Evaluation-Bias, which is based on PyTorch.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02942-6.