Succi, Chang, and Rao argue that medical education needs a deliberate redesign for an AI-rich clinical world, not just a bolt-on “AI lecture” or a new tool in the curriculum.
Core argument
- AI, especially large language models (LLMs), is already strong at many clinical-adjacent tasks (documentation, communication support, test-style questions), but performance on benchmarks does not equal genuine clinical reasoning.
- Medical education’s job is to teach reasoning processes and adaptability, not just factual recall or pattern recognition. LLMs can look convincing while still producing plausible but shallow outputs.
Why current LLM success is not the same as clinical reasoning
- The authors emphasize that LLMs often operate via statistical pattern matching, so they can generate confident answers triggered by “buzzwords” or common feature clusters.
- Real clinical reasoning is dynamic: new symptoms appear, data conflict, hypotheses evolve, uncertainty persists. Exams with single best answers do not capture that.
What needs to change in assessment and benchmarking
- They call for new benchmarks that require models to reason step by step through complex cases, justify decisions, and iteratively refine a diagnosis or plan as information changes.
- Validating AI in the education setting, where reasoning can be scrutinized, is presented as a pathway toward trustworthy clinical decision support later.
How AI could reshape teaching and learning
- If LLMs become better at transparent reasoning, they could function as case-based learning partners: tutors, critics of student logic, graders, and discussion counterparts.
- LLMs could help learners at all stages parse difficult materials, including curricula, textbooks, and biomedical literature, which supports lifelong learning in a fast-moving field.
- AI could expand clinical exposure beyond “the patients you happen to see” by generating many varied presentations, including rare diseases and culturally distinct scenarios.
SP-LLMs (standardized patient LLMs)
- The article highlights the idea of LLM-powered standardized patient interactions that can be used for practice and evaluation of communication skills, including exposure to rare and diverse presentations.
Equity and access
- The authors argue LLMs could democratize medical education by distributing expertise at scale, supporting resource-limited settings and schools with lower patient diversity or volume.
- They note that equitable access will require thoughtful licensing models and partnerships between well-resourced and resource-constrained institutions.
What the “AI-enabled physician” must become
- As AI takes on routine tasks, physicians should shift toward higher-level responsibilities: strong clinical reasoning, data interpretation, and ethical oversight of algorithmic outputs.
- Curricula should include “data systems literacy” so future physicians can critically appraise and safely integrate AI outputs into care.
A non-negotiable: dual competency
- The authors stress that technical sophistication must not erode foundational clinical skills. Systems fail, downtime happens, breaches occur, and public health crises arise.
- Training should explicitly reinforce operating both with and without AI, through exercises that require history, exam, and differential diagnosis without digital aids.
Bottom line
Medical schools should integrate AI in ways that strengthen, rather than replace, rigorous reasoning, empathy, and moral judgment. This requires honest engagement with AI limits, new forms of assessment, and collaboration between clinicians, educators, and machine learning experts.
Read more on Building the AI-Enabled Medical School of the Future by Succi, Chang, and Rao.
JACQUELINE P. ASHBY 