AI Should Teach Like a Book, Not Like a Consultant

Generative AI has entered residency faster than any curriculum, assessment system, or governance framework designed to accommodate it. The research is no longer theoretical.

One large study found a strong negative correlation between frequent AI tool usage and critical thinking scores, with cognitive offloading as a mediating factor. Younger participants showed the highest dependence and the lowest scores. A separate randomised controlled trial found students who studied with ChatGPT scored noticeably lower on a surprise retention test weeks later than those who studied traditionally.

We have seen this pattern before, outside medicine. Studies link heavy GPS reliance to weaker spatial memory, and London taxi drivers who learned the city through sheer effort showed measurable growth in the brain's memory center. The brain builds the map only when it has to.

The clinical evidence is more sobering still. Early research suggests that de-skilling can reach even experienced physicians, with measurable declines in performance after routine reliance on AI assistance.

A recent review in medical education identified three cognitive vulnerabilities: deskilling among advanced learners, never-skilling among junior trainees who bypass foundational mental-model construction, and automation bias — the uncritical acceptance of machine output even when it is wrong. A BMJ opinion piece framed it directly: AI is making clinical reasoning optional. In a national survey of family medicine residents, most reported using AI in the past month, about a third for clinical decision support, and only about a third had received any AI education in the previous six months.

The distinction I draw is between AI as reference and AI as reasoner. Used as a textbook, AI can support knowledge acquisition the way a well-indexed reference does, provided the cognitive load stays on the learner. Reading AI-generated material to review a topic, much as one would read a textbook or a journal article, is very reasonable. Used to generate a differential diagnosis, it performs the precise cognitive act residency exists to train. The resident who never constructs a differential independently never develops one.

That distinction matters most when you consider who is using the tool. For a fully developed clinician, AI can be genuine augmentation. It can sometimes surface edge cases, differentials or considerations that were not on the list. That physician already has the reasoning to judge whether the suggestion fits, though even experienced hands should keep practicing without it, as early de-skilling research suggests. In the trainee setting, educators worry the same tool may do the opposite. When AI builds the differential and reaches the conclusion without the learner doing the cognitive work, it is a disservice to the trainee and to the profession. The long-term cost may be slower development of the complex reasoning that comes from working through the history and physical examination. The tool that can augment the expert risks becoming a crutch for the novice.

A multicentre randomised trial of attendings and residents in family, internal, and emergency medicine found that physicians given GPT-4 plus conventional resources showed no meaningful improvement in diagnostic reasoning over conventional resources alone, while, in a secondary analysis, GPT-4 on its own outperformed the physician control group. Having the tool did not sharpen the reasoning.

The governing principle is sequence: independent reasoning and unassisted practice must precede augmentation. Residents draft their own differential before seeing the machine's. They then compare, and the divergence between the two becomes the curriculum. A follow-up experiment by the same critical thinking researcher found that unguided AI use fostered offloading and added only modestly to reasoning quality, whereas structured prompting significantly reduced offloading and enhanced both critical reasoning and reflective engagement.

For training programs, this means three commitments. Protected AI-free deliberate practice during formative years. AI literacy taught as a skill — knowing when the output is sound, when it is fluent nonsense, when it reflects training-data bias. And assessment that continues to measure the unassisted mind, because a physician who cannot reason without a screen is not a physician.

The question was never whether residents would use AI. It is whether we will graduate them having never once thought alone.

The same rule holds well beyond medicine. When someone is still learning a skill, AI should check their work, not do the thinking for them. Once the skill is well developed, built through years of real cognitive work, AI can sharpen it and make them more efficient. At that point, it's just another tool in the toolbox.

A shorter version of this piece was shared on LinkedIn on October 29, 2026.

References

1. Gerlich M. Societies. 2025;15(1):6. doi:10.3390/soc15010006
2. Barcaui A. Soc Sci Humanit Open. 2025;12:102287. doi:10.1016/j.ssaho.2025.102287
3. Dahmani L, Bohbot VD. Sci Rep. 2020;10:6310. doi:10.1038/s41598-020-62877-0
4. Maguire EA, Gadian DG, Johnsrude IS, et al. Proc Natl Acad Sci USA. 2000;97(8):4398–4403. doi:10.1073/pnas.070039597
5. Woollett K, Maguire EA. Curr Biol. 2011;21(24):2109–2114. doi:10.1016/j.cub.2011.11.018
6. Budzyń K, et al. Lancet Gastroenterol Hepatol. 2025;10(10):896–903. doi:10.1016/S2468-1253(25)00133-5
7. Liu Y, et al. Front Med. 2026;13:1867722. doi:10.3389/fmed.2026.1867722
8. Nghia NP. BMJ. 2026;393:s871. doi:10.1136/bmj.s871
9. Barr WB, et al. Fam Med. 2026 (online first). doi:10.22454/FamMed.2026.0347
10. Goh E, et al. JAMA Netw Open. 2024;7(10):e2440969. doi:10.1001/jamanetworkopen.2024.40969
11. Gerlich M. Data. 2025;10(11):172. doi:10.3390/data10110172