Gender Differences in the Use of ChatGPT as Generative Artificial Intelligence for Clinical Research and Decision-Making in Occupational Medicine

Our plain-English summary of Gender Differences in the Use of ChatGPT as Generative Artificial Intelligence for Clinical Research and Decision-Making in Occupational Medicine, by P. Mashburn, F.A. Weuthen, N. Otte, et al (Healthcare, 2025).

The study

• Authors: P. Mashburn, F.A. Weuthen, N. Otte, H. Krabbe, G.M. Fernandez, T. Kraus, J. Krabbe
• Journal: Healthcare
• Published: 2025;13(12)
• Cited by: 11 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Background/Objectives: Artificial intelligence (AI) has evolved from early diagnostic expert systems to advanced generative models, such as GPT-4, which are increasingly being used in healthcare. Concerns persist regarding inaccuracies and input dependency. This study aimed to deliver initial insights into whether gender influences the interaction of medical professionals with generative AI. Methods: This analysis investigated gender differences in medical students' and physicians' interactions with ChatGPT-4 while researching occupational medicine cases in a randomized controlled study. Participants assessed cases involving asbestos-related disease, metal sulfate allergy, and berylliosis using ChatGPT. Inputs and outputs were evaluated for accuracy, confabulations, communication styles, and user satisfaction. Demographic data and self-assessments of occupational medicine knowledge before and after the tasks were also collected. Results: Among 27 participants (63% women, 37% men), women showed greater knowledge improvement after using ChatGPT, particularly in asbestos-related cancer identification. No significant gender differences emerged in diagnostic accuracy, reporting procedures, or satisfaction with ChatGPT. Women exhibited significantly higher self-rated competence after using the ChatGPT application, while men only showed minimal change. Input from the female participants led to more confabulations, although response accuracy remained comparable. Conclusions: This study offers the first real-world insights into the use of generative AI in occupational medicine, highlighting the importance of understanding user-dependent variability in AI-supported clinical practice and decision-making. These findings underscore the need for gender-sensitive AI literacy training in medical education, accommodating diverse interaction styles and strategies to mitigate AI-generated misinformation. Future research with larger and more diverse cohorts could provide deeper insights into the influence of gender, age, and experience on AI utilization in healthcare. Integrating gender-based interaction differences into AI training and applications may improve clinical performance and promote more equitable healthcare practices.

In plain English

This study was not about people exposed to asbestos. It was a gender-stratified subgroup analysis from a randomized study in Germany. In that study, medical students and physicians researched three occupational medicine cases. One case described an outdoor worker with asbestos-related pleural changes on CT; the other two involved metal sulfate allergy and berylliosis. The analysis included the 27 participants assigned to use ChatGPT-4 (17 women, 10 men; 21 students, 6 physicians), and 15 completed all three cases. In the asbestos case, women named significantly more asbestos-related cancer types after using ChatGPT than before, while the change among men was not statistically significant. After using ChatGPT, women rated their own occupational medicine expertise better than men did (school grade 2.8 versus 4.3, where 1 is best). Satisfaction with ChatGPT did not differ significantly by gender. ChatGPT's answers to women contained more confabulations on average (6.0 versus 4.0), but this difference was not statistically significant. The authors concluded that, if replicated, the findings could suggest a need for gender-sensitive AI training. They cautioned that the very small sample means the results should not be generalized.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Gender-stratified subgroup analysis of a randomized controlled study comparing ChatGPT-4 with participants' usual internet research. The gender groups were formed afterward, without matching.
• Who was studied: 27 participants assigned to the ChatGPT group of an online study conducted in German (RWTH Aachen University, Germany): 17 women and 10 men, of whom 21 were medical students and 6 were physicians. 22 completed case 1 and 15 completed all three cases.
• Exposure: Use of ChatGPT (OpenAI's GPT-4 model, GPT-4-0125-preview) to research three occupational medicine cases. Case 1 was an outdoor worker with incidental asbestos-related pleural changes on thoracic CT; case 2 was metal sulfate allergy in galvanization; case 3 was berylliosis in a former dental technician. No human asbestos exposure was studied.
• Main results:
  • In case 1 (n = 22), women named significantly more types of asbestos-caused cancers after using ChatGPT than before (p less than 0.001). The before-versus-after change in men was not statistically significant (p = 0.12). Before research, the mean was 1.6 for women and 2.0 for men (p = 0.41). After research, it was 2.9 for women and 2.8 for men (p = 0.75).
  • Before using ChatGPT, 9 women and 1 man correctly answered whether cholangiocellular carcinoma is recognized as an occupational disease in Germany (p = 0.02). After using ChatGPT, 9 women and 7 men answered correctly (p = 0.24).
  • There was no gender difference in identifying the correct German occupational disease number for asbestos-induced pleural changes: 5 women and 2 men answered correctly (p = 0.60).
  • Among the 16 participants who answered the concluding questions, self-rated occupational medicine expertise after the cases averaged 2.8 (SD 0.8) for women and 4.3 (SD 0.8) for men on German school grades, where 1 is best (p = 0.006). Women's ratings improved from before to after (p = 0.02); men's did not (p = 0.99). In a regression, gender predicted the post-use self-assessment (beta = -1.24, p = 0.03), while age and baseline self-assessment did not.
  • Satisfaction with ChatGPT did not differ significantly by gender: the mean grade was 2.5 for women and 3.3 for men (p = 0.33).
  • ChatGPT's outputs averaged 6.0 (SD 4.7) confabulations for women and 4.0 (SD 2.7) for men; this difference was not statistically significant (p = 0.45). Consecutive mistakes were similar: 2.4 for women and 2.6 for men (p = 0.78).
• The authors' conclusion: The authors concluded that gender differences, possibly linked to communication styles, could influence the use of AI tools like ChatGPT. If replicated, the findings could suggest a need for gender-sensitive, or at least communication-style-targeted, AI training.
• Limitations the authors state:
  • The sample was very small: only 15 participants completed all three cases. The authors advised against generalizing the results.
  • Attrition bias cannot be completely ruled out, even though the dropout analysis found no significant differences.
  • Most participants were students from one university, which introduces potential sample and selection bias.
  • Self-reported gender and medical background were not independently verified.
  • Participants' AI literacy, AI experience and prompt-writing expertise were not characterized.
  • The small sample made complex statistical analysis infeasible, so the combined influence of gender, age and background could not be modeled.

Related studies

• Research cost analyses to aid in decision making in the conduct of a large prevention trial, CARET. Carotene and Retinol Efficacy Trial — M.D. Thornquist, N. Urban, et al. (1993)
• Ki67 (MIB-1) as a Prognostic Marker for Clinical Decision Making Before Extended Pleurectomy Decortication in Malignant Pleural Mesothelioma — R.A. Belderbos, A.P.W.M. Maat, et al. (2021)
• Understanding clinical decision-making in mesothelioma care: a mixed methods study — C. Henshall, P. Dawson, et al. (2022)
• Not just a research method: If used with caution, can job-exposure matrices be a useful tool in the practice of occupational medicine and public health? — M. Fadel, B.A. Evanoff, et al. (2020)
• Communicating imaging findings in peritoneal mesothelioma: the impact of 'PAUSE' on surgical decision-making — A. Chandramohan, N. Shah, et al. (2021)
• Epidemiology: a foundation of environmental decision making — K.C.K. Deener, J.D. Sacks, et al. (2018)

Browse more research in Healthcare, or return to medical research.