The doctor who forgot and the doctor who never learned

I've spent about twenty years learning to read a sick patient before the labs come back. Residents starting today may never learn it the same way.

Some of that is progress. The research from the past two years also says AI is eroding physician skill in two separate ways, and which one hits you depends on when you trained.

Three ways to lose a skill

In 2025 the New England Journal of Medicine put names on the problem:

• Deskilling: losing a skill you already had.
• Never skilling: never building it at all.
• Mis skilling: learning the AI's mistakes as if they were facts.

My generation sits in the first group. Today's trainees sit in the other two.

The veterans

Experience doesn't make anyone immune. In Poland, 19 endoscopists who had each done more than 2,000 colonoscopies started using AI to spot polyps. Within months, their detection rate on procedures done without AI fell from 28.4% to 22.4%.

These were experts. The study was observational, so other factors may have contributed, and the authors say so. A Lancet commentary still called it the first real world clinical evidence of deskilling.

Radiology tells a similar story. When an AI suggested the wrong answer on mammograms, radiologists with more than 15 years of experience dropped from 82% accuracy to 45.5%. Seniority helped. It didn't save them.

The trainees

Never skilling is harder to catch, because there's no baseline to fall from. A 2026 Nature Medicine paper named it as its own risk and admitted the direct evidence is still thin.

The early numbers point one way. In a Nature Medicine study of nearly 800 people, practicing physicians held their accuracy when an AI gave bad advice with a convincing explanation. Medical students deferred to it more often. Years at the bedside work like a firewall, and students haven't built theirs yet.

Training on the tool won't close the gap by itself. In a randomized trial, 44 physicians who had completed AI literacy training worked through diagnostic cases. Those given clean AI suggestions scored 84.9%. Those given suggestions with planted errors scored 73.3%.

Why the struggle matters

Neuroplasticity cuts both ways. Circuits you use get stronger, and circuits you skip get pruned.

London taxi drivers showed this in 2000. Their posterior hippocampi, the region linked to spatial memory, were larger than those of people who didn't drive taxis, and the size grew with years on the job.

Clinical reasoning is built the same way. You commit to a differential at 3 a.m., get it wrong, and get corrected by an attending or by the patient's course. That correction is the workout. When a model hands you the differential first, you skip it.

An MIT Media Lab preprint adds an early hint. Among 54 adults writing essays, the group using a large language model showed the weakest brain connectivity on EEG. It isn't peer reviewed and it didn't study doctors, so I treat it as a signal and nothing more.

What I'd change tomorrow

For residency programs and medical schools:

1. Commit before you consult. Write the assessment and plan first, then open the AI and talk through the differences.
2. Protect AI off reps. Schedule cases with the assistance off, the way pilots still hand fly.
3. Test unassisted. Milestone exams should prove the skill exists without the tool.
4. Plant the errors. Hand learners AI output with seeded mistakes and score what they catch.

For physicians already in practice:

1. Track your unassisted numbers wherever a metric exists, such as adenoma detection rate.
2. Form your impression before you read the AI's. In that Nature Medicine study, seeing the AI recommendation first made people more deferential.
3. Don't lean on a course alone. Trained physicians still followed planted errors, so the workflow has to carry some of the protection.

The trade between generations

Doctors who trained before AI carry a record of what unassisted competence looks like, and that record retires with my generation. Residents bring a fluency with the tools that my generation never had.

Pair them on purpose, each teaching the other. It's the cheapest fix on this list.

If you trained before AI, what's the one skill you'd hate to lose? I'd like to hear your answer in the comments on my Substack.

Sources

• Abdulnour et al., NEJM, 2025: deskilling, never skilling and mis skilling. Summary
• Budzyn et al., The Lancet Gastroenterology & Hepatology, 2025: endoscopist deskilling. Lancet release
• Dratsch et al., Radiology, 2023: automation bias in mammography. Summary
• Ke et al., Nature Medicine, 2026: AI induced never skilling. Coverage
• Xu et al., Nature Medicine: AI explanations across expertise levels. Columbia DBMI summary
• Randomized trial of automation bias in physicians trained in AI literacy. Summary
• Maguire et al., PNAS, 2000: taxi driver hippocampi. DOI
• Kosmyna et al., MIT Media Lab, 2025: Your Brain on ChatGPT (preprint). Project page

For a broader discussion with your team, see my guide to AI governance for healthcare boards. If you're planning a medical education or clinical meeting on this topic, explore my healthcare AI keynotes.