The integration of artificial intelligence (AI), including generative large language models (GLLMs), is rapidly expanding across medical education domains, encompassing admissions, teaching, learning, assessment, and program evaluation [
1]. Several studies have focused on developing AI competencies for medical student graduates [
2] and healthcare professionals more broadly [
3], underscoring the imperative for curricular integration to ensure graduates achieve these essential competencies. Against a backdrop of ever-changing and competing priorities in medical education [
4], current scholarship seems to demonstrate a concerning tendency to treat disparate AI technologies as functionally equivalent, with platforms such as ChatGPT, Claude, or Gemini frequently discussed as interchangeable “AI tools” despite differences in their capabilities, interfaces, and educational impacts [
5,
6]. This conceptual flattening obscures an emerging phenomenon that warrants serious attention: as medical schools increasingly integrate AI into their curricula to meet competency requirements, AI access and proficiency may become a new axis of educational stratification in medical training, potentially creating disparities in learning outcomes based on students’ differential access to, familiarity with, or aptitude for specific AI platforms.
The literature’s tendency to aggregate AI technologies under broad categories overlooks functionally important differences between platforms. In this commentary, we focus on GLLMs given their prevalence in medical education scholarship. However, we feel it is critical to acknowledge that discussion of AI tools must transcend GLLMs, being careful to draw distinctions in how various tools are built based on their intended purposes. For example, ChatGPT and Claude excel at generative synthesis and explanatory reasoning, making them particularly valuable for complex problem-solving tasks. Microsoft Co-Pilot integrates seamlessly with productivity workflows, offering contextual prompts that scaffold procedural learning. Platforms like Gemini and GPT-4o support multimodal interactions, requiring users to possess technical fluency in image and audio integration.
As medical trainees increasingly rely on GLLMs for studying, diagnostic reasoning, and/or other academic endeavors, differential access to platforms can create systematic disparities that extend beyond traditional measures of academic ability or effort. The epistemic consequences of this differential access merit serious consideration. Medical trainees using less capable AI tools may experience systematic exclusion from high-level academic discourse, not due to intellectual limitations but because of technological constraints that compromise their ability to generate, articulate, and defend complex ideas. Fricker [
7] describes this sort of systematic exclusion of certain individuals from these knowledge validation processes as epistemic injustice. In other words, trainees with institutional access to sophisticated AI platforms may demonstrate enhanced performance not because of superior knowledge or skills, but because their tools provide more nuanced feedback, faster processing, and greater functionality. Meanwhile, peers using free-tier versions with usage limits and reduced capabilities face structural barriers to equivalent academic participation. Furthermore, AI integration generates new forms of cultural capital that may correlate with existing privilege structures [
8]. Prompt literacy, platform fluency, and subscription access represent emergent “capital” that are unequally distributed across trainee populations. Medical trainees from higher socioeconomic backgrounds may possess superior capacity for leveraging AI technologies due to prior exposure to technology-rich educational environments and advanced linguistic training. In fact, Claude’s architecture rewards sophisticated prompt construction—a skill that correlates with prior educational privilege and linguistic competency.
The cumulative effect of these differences can shape how trainees encounter information, develop reasoning skills, and participate in educational discourse. Institution-specific AI implementation decisions can also exacerbate these differences, often prioritizing cost efficiency and administrative convenience over educational equity and pedagogical flexibility. Institutional partnerships with specific providers like Microsoft or OpenAI may be driven by existing procurement relationships rather than evidence-based assessments of educational outcomes. As Eubanks notes [
9], this type of infrastructural inequality can inadvertently create tiered access systems where some trainees receive premium AI support while others are relegated to basic versions with significant functional limitations.
Rather than functioning as an equalizing force, AI technologies risk becoming sophisticated sorting mechanisms that maintain existing hierarchies while appearing to offer universal benefits. The challenge for medical education lies in recognizing that AI integration represents not merely a technological upgrade, but a fundamental restructuring of educational opportunity structures. As we begin interrogating the mechanisms through which GLLMs might be producing, magnifying, or disguising stratification, we share an insightful case study reflecting on the integration of GLLMs built into the mandatory third-year core medicine clerkship at the University of Maryland School of Medicine.
Bookended by a pre-post design relying on the Generative AI Literacy Assessment Tool (GLAT) [
10], students (n=25) completed a baseline GLAT assessment, participated in a didactic session on AI fundamentals and LLMs, then applied a GLLM of their choosing to revise a differential diagnosis and management plan from a previous clinical case. Students specifically prompted their chosen GLLM to generate a differential diagnosis, findings supporting and opposing said diagnosis, and potential diagnostic tests to confirm said diagnosis. Students were also asked to critically analyze their GLLM outputs for possible hallucinations. Afterwards, students completed a post-intervention GLAT assessment. GLAT scores were not factored into clerkship evaluations to maximize transparent use and reflection of AI.
Fifty percent of students employed ChatGPT exclusively, and 34.6% used OpenEvidence exclusively. Remaining students relied on either a combination of ChatGPT and OpenEvidence, Glass Health, or Open Intelligence. Interestingly, no students used the institutionally licensed LLM (Microsoft Co-Pilot). Pre- to post-intervention GLAT scores demonstrated improvement across user groups, with ChatGPT-only users (n=13) showing an increase from 68.75 to 80.83, and OpenEvidence-only users (n=9) demonstrating gains from 61.88 to 83.13. While gains in AI literacy are seen for both samples, it is worth noting that ChatGPT-only users entered the clerkship with a higher baseline literacy score, perhaps indicating that students with higher AI literacy may be gravitating to ChatGPT. This deserves further investigation.
Preliminary analysis of student responses to the hallucination item seems to also indicate platform-specific patterns, particularly in relation to error recognition and mitigation strategies. ChatGPT-based hallucination responses predominantly emphasized the need to identify and correct algorithmic assumptions, while OpenEvidence-based responses focused on enhancing accuracy and clinical specificity. Taken together, these distinctions suggest that different GLLMs may require distinct critical appraisal skill sets, with ChatGPT necessitating assumption-based evaluation and OpenEvidence requiring precision-focused assessment. This platform-dependent variation in required skill sets may have important implications for the differential impact of GLLM integration on medical education outcomes. Specifically, these tools may function as cognitive amplifiers that could either mitigate or exacerbate pre-existing academic advantages and disadvantages among learners. A more detailed thematic analysis of student prompts as well as subsequent GLLM responses is still ongoing.
Addressing these challenges requires institutional action grounded in evidence, with particular attention to global disparities in access and infrastructure. Our case study suggests that AI literacy training should move beyond generic “how to” curricula to explicitly teaching platform-specific competencies such as recognizing assumptions being made in GLLMs, evaluating clinical specificity, and understanding how interface design may shape diagnostic reasoning. Medical schools should secure institutional licenses for multiple high-capability platforms, ensuring all trainees can develop fluency across different AI ecosystems, and avoiding single provider partnerships that prioritize administrative efficiency over instructional versatility. These challenges intensify in resource-constrained international contexts. In many low- and middle-income countries, students lack reliable access to more premium GLLMs, face bandwidth limitations that restrict real-time AI interactions, and encounter language barriers as most advanced GLLMs remain optimized for English-language prompts. Interventions that reduce disparities at a well-resourced US medical school may prove inaccessible or misaligned in other contexts. As AI becomes more embedded across the global medical education landscape, striving toward equity across diverse contexts must remain a central commitment, regardless of geographical or institutional variability.
AI is not an inherent equalizer in medical education and its integration demands a critical examination of how these technologies reproduce and amplify existing inequities. By recognizing AI platforms as heterogeneous resources rather than homogeneous solutions, we can better understand how differential access can systematically advantage and/or disadvantage our trainees. The path forward requires deliberate attention to platform equity and proactive measures to ensure that AI tools democratize rather than concentrate educational advantage. Without such attention, medical education risks reproducing existing inequalities under the guise of technological innovation, ultimately undermining the profession’s commitment to equitable patient care.
Acknowledgements
None.
Funding
The authors declare that no funds, grants, or other financial support were received for the conduct of this study.
Conflicts of interest
No potential conflict of interest relevant to this article was reported.
Author contributions
MS, NG, and VK conceived the manuscript perspective. MS drafted the initial manuscript and led the revision process. NG and VK critically reviewed the manuscript. All authors reviewed and approved the final manuscript.
References
- 1. Boscardin CK, Abdulnour RE, Gin BC. Macy Foundation Innovation Report Part I: current landscape of artificial intelligence in medical education. Acad Med. 2025;100(9S Suppl 1):S15-S21. https://doi.org/10.1097/ACM.0000000000006107
- 2. Lee YM, Kim S, Lee YH, et al. Defining medical AI competencies for medical school graduates: outcomes of a Delphi survey and medical student/educator questionnaire of South Korean medical schools. Acad Med. 2024;99(5):524-533. https://doi.org/10.1097/ACM.0000000000005618
- 3. Russell RG, Lovett Novak L, Patel M, et al. Competencies for the use of artificial intelligence-based tools by health care professionals. Acad Med. 2023;98(3):348-356. https://doi.org/10.1097/ACM.0000000000004963
- 4. Kim S, Kim SH, Kim H, Lee YM. Integrating artificial intelligence into medical curricula: perspectives of faculty and students in South Korea. Korean J Med Educ. 2025;37(1):65-70. https://doi.org/10.3946/kjme.2025.324
- 5. Mavrych V, Yaqinuddin A, Bolgova O. Claude, ChatGPT, Copilot, and Gemini performance versus students in different topics of neuroscience. Adv Physiol Educ. 2025;49(2):430-437. https://doi.org/10.1152/advan.00093.2024
- 6. Meyer MK, Kandathil CK, Davis SJ, et al. Evaluation of rhinoplasty information from ChatGPT, Gemini, and Claude for readability and accuracy. Aesthetic Plast Surg. 2025;49(7):1868-1873. https://doi.org/10.1007/s00266-024-04343-0
- 7. Fricker M. Epistemic injustice: power and the ethics of knowing. Oxford, UK: Oxford University Press; 2007.
- 8. Bourdieu P. The forms of capital. In: Richardson J, ed. Handbook of Theory and Research for the Sociology of Education. New York, USA: Greenwood; 1986:241-258
- 9. Eubanks V. Automating inequality: how high-tech tools profile, police, and punish the poor. New York, USA: St. Martin’s Press; 2018.
- 10. Jin Y, Martinez-Maldonado R, Gašević D, Yan L. GLAT: the generative AI literacy assessment test. Comput Educ Artif Intell. 2025;9:100436. https://doi.org/10.1016/j.caeai.2025.100436
Citations
Citations to this article as recorded by
- The new wave of generative artificial intelligence in medical education
Eulho Jung, Anita Samuel
Korean Journal of Medical Education.2026; 38(1): 1. CrossRef
