Article Structure
1. Highlights
2. Background and Clinical Relevance
3. Study Design and Methods
4. Key Findings
5. Interpretation for Surgical Practice and Academic Medicine
6. Strengths and Limitations
7. Research, Policy, and Educational Implications
8. Conclusion
9. Funding, Registration, and Citation
10. References
Highlights
ChatGPT 3.5 generated different descriptors for a “great surgeon” depending on whether prompts included race, ethnicity, gender, or sexual orientation.
Male surgeons received the same descriptors as the control response, while women and gay surgeons were more often framed as compassionate and empathetic rather than confident.
Black surgeons were characterized with terms such as resilient, trailblazing, exceptional, and inspirational, and Latin/Hispanic surgeons were linked to bilingualism, cultural competence, and advocacy.
These patterns do not diagnose intent or prejudice in an individual system output, but they strongly suggest that large language models can reproduce social stereotypes embedded in their training data, with potential relevance for hiring, promotion, mentorship, and leadership evaluation in surgery.
Background and Clinical Relevance
Diversity in the physician workforce is not merely an institutional value statement; it is associated with clinically meaningful benefits. A more diverse workforce can improve communication, strengthen trust, reduce cultural and linguistic barriers, and increase access for historically underserved populations. In surgery, where teamwork, rapid decision-making, and leadership assessments are central to training and advancement, perceptions of competence and authority matter enormously.
Yet the pathway to advancement in academic surgery remains shaped by structural inequities. Women and underrepresented-in-medicine physicians continue to experience disparities in sponsorship, compensation, leadership representation, speaking opportunities, authorship visibility, and promotion. Bias can be explicit, but more often it is implicit: subtle assumptions about who appears decisive, technically gifted, confident, nurturing, or “leader-like.” These assumptions can influence narrative evaluations, recommendation letters, trainee feedback, operative autonomy, and appointment to leadership roles.
Large language models such as ChatGPT are trained on massive corpora of human-generated text. As a result, they can reflect both useful patterns and harmful biases present in society and in professional discourse. This matters for medicine because AI tools are increasingly being used in education, documentation, decision support, recruitment, and public-facing communication. If these systems systematically associate certain identities with communal traits and others with agentic or leadership traits, they may amplify pre-existing inequities under the appearance of technological neutrality.
The study by Avelar and colleagues examines this issue through a focused and clinically resonant question: how does ChatGPT describe a “great surgeon” when demographic identifiers are added to the prompt?
Study Design and Methods
This was a qualitative prompt-based analysis using ChatGPT 3.5. The investigators first asked the model to provide the “top 5 descriptors of a great surgeon,” which served as the control response. They then repeated the query while adding descriptors related to gender, race, ethnicity, and sexual orientation, generating 14 surgeon phenotypes for comparison with the control output.
According to the abstract, the control descriptors were “skilled, confident, meticulous, diligent, and innovative.” Outputs for the other phenotypes were then qualitatively compared with this baseline to identify patterns suggestive of bias.
The study is best understood as an exploratory audit of generated language rather than a formal validation experiment. There were no clinical endpoints, no patient outcomes, and no inferential statistics reported in the abstract. The central outcome was variation in descriptive language across identity-based prompts.
Key Findings
The most striking observation was that the default image of a “great surgeon” aligned perfectly with the description of a great male surgeon. In other words, the male identity prompt reproduced the control descriptors unchanged: skilled, confident, meticulous, diligent, and innovative. This suggests that, within the model’s learned language patterns, the unmarked or default concept of surgical greatness may be implicitly male.
By contrast, great female surgeons, great gay male surgeons, and great gay female surgeons were characterized as “compassionate and empathetic” rather than confident. That shift is important. Compassion and empathy are unquestionably valuable clinical attributes, especially in perioperative communication and longitudinal patient trust. However, when these terms replace rather than complement descriptors such as confidence, the output risks reinforcing a familiar stereotype: some groups are framed as caring, while others are framed as authoritative.
The descriptors assigned to Black surgeons followed a different pattern. Great Black male, gay male, and gay female surgeons were described as “skilled, exceptional, resilient, trailblazing, and inspirational.” Great Black female surgeons were uniquely labeled tenacious. On one level, these are highly positive terms. On another, they imply that excellence in Black surgeons is narrated through struggle, exceptionalism, and symbolic leadership rather than being treated as normative professional competence. Such language can seem affirming while still “othering” the individual, suggesting that success is remarkable because it occurs despite barriers or because it carries representational meaning.
For Latin/Hispanic surgeons, ChatGPT emphasized being “bilingual and effective communicators, culturally competent and empathetic, resilient and determined, and advocates and leaders.” For Latin/Hispanic female and gay female surgeons, “mentorship and community engagement” replaced “bilingual and effective communicators.” Again, these are largely complimentary descriptors, but they tether professional identity to service, communication, and community roles in ways that may differ from default assumptions about technical authority or leadership confidence.
Importantly, the issue is not whether empathy, resilience, cultural competence, mentorship, or advocacy are desirable. They clearly are. The issue is asymmetry. If one group is described in terms of confidence and innovation, while another is described in terms of empathy and resilience, AI output may mirror the unequal distribution of status-laden traits that has long characterized biased evaluations in academic medicine.
Interpretation for Surgical Practice and Academic Medicine
This study speaks less to surgical skill itself than to the social language that surrounds judgment in surgery. Career trajectories are influenced by how faculty, trainees, and institutions talk about excellence. Recommendation letters for men, for example, have historically been more likely to emphasize achievement, leadership, and independence, whereas letters for women may more often stress diligence, warmth, and teamwork. Similar patterns have been reported in trainee assessment language and professional evaluations. If AI systems absorb these patterns, they may reproduce them in educational materials, performance summaries, search committee drafts, institutional communications, and feedback tools.
There are at least four practical implications.
First, AI-generated text should not be assumed to be neutral. Institutions adopting large language models for faculty development, recruitment support, evaluation narratives, or public-facing biographies should recognize that seemingly polished language can contain patterned social bias.
Second, descriptor shifts matter because they map onto status hierarchies. In leadership science and organizational psychology, agentic traits such as confidence, decisiveness, and assertiveness are often rewarded more strongly in selection and promotion than communal traits such as warmth and empathy. Surgery, a field historically associated with decisiveness and technical mastery, may be particularly vulnerable to this distortion.
Third, positive stereotyping is still stereotyping. Calling underrepresented surgeons inspirational, resilient, or community-oriented may sound affirming, but it can impose extra identity labor. It may also shift attention away from technical achievement, scholarship, and executive leadership.
Fourth, this study raises a translational concern for medical AI governance. Bias in language generation may not directly harm a patient in the way a dosing error could, but it can influence workforce culture, educational opportunity, and institutional decision-making. These downstream effects are clinically relevant because workforce equity affects access, trust, and quality of care.
Strengths and Limitations
The study’s main strength is conceptual clarity. It uses a simple, reproducible prompt framework to make an abstract concern visible. By comparing outputs against a control and across multiple identity phenotypes, it demonstrates that prompt wording related to identity can change the model’s representation of professional excellence.
Its limitations are also important. The analysis appears qualitative and based on a single model version, ChatGPT 3.5. Large language model outputs can vary over time, across sessions, and with minor prompt changes. The abstract does not specify whether responses were repeated, whether temperature settings were standardized, or whether inter-rater methods were used for qualitative interpretation. Nor does the study establish the prevalence or magnitude of the effect across multiple models or use cases.
In addition, the findings should not be overextended into claims about clinical decision-making or patient safety. The study evaluates representational bias in generated language, not medical recommendations. Still, representational bias can shape professional ecosystems in consequential ways, especially when language models are deployed in high-volume administrative or educational workflows.
Another nuance is that identity-linked descriptors may reflect real-world roles some surgeons intentionally cultivate, such as mentorship, advocacy, or language-concordant care. The problem arises when these characteristics are selectively assigned or substituted for high-status descriptors rather than integrated alongside them.
Research, Policy, and Educational Implications
Future work should move beyond single-prompt demonstration studies toward more rigorous AI bias auditing in medicine. Useful next steps would include repeated-prompt sampling, blinded coding, comparison across model versions and vendors, and quantitative linguistic analysis of agentic versus communal descriptors. Studies should also test more realistic use cases, such as drafting recommendation letters, faculty bios, annual reviews, trainee milestone summaries, and leadership nominations.
For institutions, several safeguards are reasonable now. Health systems and academic departments should require human review of AI-generated evaluative text, especially in hiring, promotion, trainee assessment, and award nomination processes. Faculty should be educated that AI can inherit social biases from its training data. Procurement and governance committees should ask vendors how bias monitoring, mitigation, and updating are handled. Internal audits of AI-assisted documents may be as important as audits of traditional evaluation language.
For educators in surgery, the study also offers a teaching opportunity. Discussions of professionalism, mentorship, and leadership development can include not only interpersonal bias but also algorithmic bias. Trainees and faculty should learn to recognize how language frames competence, authority, and belonging.
More broadly, the findings align with a substantial literature showing that bias in medicine can be embedded in systems, not only individuals. AI may become another layer where these inequities are reproduced unless actively monitored. This is especially relevant in surgery, where prestige, selectivity, and hierarchical culture can magnify small differences in perception over the course of a career.
Conclusion
Avelar and colleagues present a concise but provocative demonstration that ChatGPT 3.5 does not describe a “great surgeon” uniformly across demographic identities. The default and male-associated image of greatness emphasized confidence, whereas women and gay surgeons were more often described in terms of compassion and empathy. Black and Latin/Hispanic surgeons were linked to resilience, trailblazing, bilingualism, cultural competence, advocacy, and community engagement. These descriptors are often complimentary, but their uneven distribution suggests that the model reflects social stereotypes embedded in human language.
For clinicians and academic leaders, the core message is practical: AI-generated language should be treated as assistive, not authoritative, especially in domains involving professional evaluation and representation. For researchers, the study underscores the need for systematic bias testing of language models in healthcare environments. For surgery as a profession, it is a reminder that equity work now includes not only people and institutions, but also the tools we increasingly use to describe excellence.
Funding, Registration, and Citation
Funding: Not reported in the abstract provided.
ClinicalTrials.gov registration: Not applicable for this qualitative AI prompt study.
Primary citation: Avelar E, Desai P, Javaid M, Ullmann TM, DiBrito S. What it takes to be great: ChatGPT’s top 5 descriptors of great surgeons by race, ethnicity, gender, and sexual orientation. Surgery. 2026-05-11;196:110244. PMID: 42114469. Available at: https://pubmed.ncbi.nlm.nih.gov/42114469/
References
1. Avelar E, Desai P, Javaid M, Ullmann TM, DiBrito S. What it takes to be great: ChatGPT’s top 5 descriptors of great surgeons by race, ethnicity, gender, and sexual orientation. Surgery. 2026-05-11;196:110244. PMID: 42114469.
2. Salles A, Awad M, Goldin L, Krus K, Lee JV, Schwabe MT, et al. Estimating implicit and explicit gender bias among health care professionals and surgeons. JAMA Netw Open. 2019;2(7):e196545.
3. Turrentine FE, Hanks JB, Schroen AT, Stukenborg GJ. Surgical resident gender does not affect patient outcomes. Ann Surg. 2016;264(3):448-455.
4. Silver JK, Poorman JA, Reilly JM, Spector ND, Goldstein R, Zafonte RD. Assessment of women physicians among authors of perspective-type articles published in high-impact pediatric journals. JAMA Netw Open. 2018;1(3):e180802.
5. Shah NH, Milstein A, Bagley SC. Making machine learning models clinically useful. JAMA. 2019;322(14):1351-1352.
6. Ferryman K, Pitcan M. Fairness in precision medicine. Data & Society. 2018. This report is widely cited in discussions of algorithmic bias in health technologies, though it is not a PubMed-indexed clinical study.
