Level one: the doctor spots the polyp. Level two: the AI points at the polyp first. Level three, boss fight: the AI disappears, and everyone realizes the doctor’s visual-detection skill tree may have been quietly nerfed in the background. I have read Gerke, Hassan, and Mori’s short Nature Reviews piece several times, and I think - I think - the question is not “Is medical AI useful?” It is “What happens if useful AI makes clinicians worse at doing the job without it?” Tiny legal panic confetti cannon.

Wait, What Is Deskilling?

Deskilling is the “use it or lose it” problem wearing a hospital badge. If a clinician relies on an AI assistant every day, their independent diagnostic muscles might get less exercise. Not because clinicians are lazy. Because humans are spectacularly good at outsourcing effort to tools. We invented GPS and immediately forgot the street two blocks over. Honestly, relatable.

In medicine, this is not just a quirky inconvenience. If an AI tool spots lesions in colonoscopy, flags suspicious scans, or suggests diagnoses, the physician is still supposed to supervise it. But supervision gets awkward if the supervisor slowly becomes less able to catch the machine’s mistakes. That is like hiring a smoke alarm and then forgetting what smoke looks like.

Gerke, Hassan, and Mori focus on whether this deskilling problem could make certain medical AI systems legally shaky under the European Union Artificial Intelligence Act, especially because the Act is built around risk categories, human oversight, and AI literacy, not vibes and crossed fingers Gerke et al., 2026.

The Colonoscopy Plot Twist Nobody Ordered

The paper lands right after a very relevant study in The Lancet Gastroenterology & Hepatology. Budzyń and colleagues looked at colonoscopy centers using AI for polyp detection. When endoscopists later performed non-AI-assisted colonoscopies, adenoma detection dropped from 28.4% before AI exposure to 22.4% after exposure Budzyń et al., 2025. Correct me if I am wrong, but that is the kind of number that makes a quality-improvement committee put down its coffee very slowly.

To be fair, this was observational. Confounding may be lurking in the vents like a tiny methodological horror movie. Workload, patient mix, sedation rates, center effects - all the usual suspects showed up wearing trench coats. But it is still a serious warning shot because AI-assisted colonoscopy also has evidence of benefit. A 2024 systematic review and meta-analysis found that computer-aided detection improved adenoma detection in randomized trials Soleymanjahi et al., 2024. So the story is not “AI bad.” It is “AI good, possibly with side effects, please read the label instead of eating the whole bottle.”

Enter the EU AI Act, Clipboard in Hand

The EU AI Act classifies some systems as prohibited, some as high-risk, and many as lower-risk. Healthcare AI often falls into the high-risk orbit because errors can affect health and safety. High-risk systems need transparency, documentation, risk management, monitoring, and human oversight. Article 14 specifically says humans overseeing high-risk AI should understand its limits, monitor it, avoid over-reliance, interpret outputs, and override or stop the system when needed EU AI Act, Article 14.

This is where Gerke and co-authors are poking the legal bruise. If the law expects meaningful human oversight, what happens when routine AI use makes the human less meaningful as an overseer? If I am reading them right - and I did re-read the paragraph because legal writing has side quests - deskilling probably is not automatically a banned “prohibited AI practice” under Article 5. The prohibited list targets things like manipulative systems, exploitation of vulnerabilities, social scoring, and certain biometric practices. Deskilling is more indirect.

But indirect does not mean irrelevant. Deskilling could undermine the Act’s high-risk safeguards. Article 4 also requires providers and deployers to support AI literacy among staff and users EU AI Act, Article 4. Article 26 says deployers must assign oversight to people with competence, training, authority, and support EU AI Act, Article 26. If the AI slowly sandpapers away that competence, the compliance story gets wobbly.

The Real Problem Is the “Human in the Loop” Sticker

Everyone loves saying “human in the loop.” It sounds comforting, like a seatbelt or a responsible adult. But a human in the loop is not magic. A tired, deskilled, over-trusting human is more like a decorative seatbelt printed on a T-shirt.

Automation bias is the known tendency to favor automated advice even when other evidence disagrees. The AI Act itself names over-reliance as something overseers must remain aware of. A 2025 mixed-method review by Natali and colleagues argues that AI-induced deskilling can affect clinical judgment, differential diagnosis, physical examination, communication, and even institutional resilience Natali et al., 2025. That is not a small bug. That is the software equivalent of “some assembly required” and the box contains a racetrack.

A 2026 scoping review similarly found limited but consistent evidence that AI can impair physician performance or reduce opportunities for skill maintenance Heudel et al., 2026. Meanwhile, another 2026 review argues the goal should be upskilling: designing AI so clinicians learn with it, not shrink beside it Oettl et al., 2026.

Practically, hospitals may need dashboards for skill drift, periodic no-AI practice, interface designs that ask clinicians to commit before revealing AI output, and training that includes AI failures. If you are mapping that workflow, a tool like mapb2.io actually fits the mess: who sees what, when, and who can override whom should not live only in a committee PDF named final_final_v7.

My Anxious Takeaway

Gerke, Hassan, and Mori are not saying medical AI should be banished to the regulatory dungeon. They are asking whether the law has enough grip on a sneaky failure mode: AI that works today while quietly making tomorrow’s human backup weaker.

That is the interesting part. The future of medical AI may depend less on whether the machine can detect the polyp and more on whether the doctor still can. Preferably without needing to respawn at level one.

References

1. Gerke S, Hassan C, Mori Y. “Human deskilling in medical artificial intelligence: prohibited or permissible under the EU Artificial Intelligence Act?” Nature Reviews Gastroenterology & Hepatology. 2026. DOI: 10.1038/s41575-026-01210-y. PMID: 42086927.

2. Budzyń K, Romańczyk M, Kitala D, et al. “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study.” The Lancet Gastroenterology & Hepatology. 2025. DOI: 10.1016/S2468-1253(25)00133-5.

3. Soleymanjahi S, Huebner J, Elmansy L, et al. “Artificial Intelligence-Assisted Colonoscopy for Polyp Detection: A Systematic Review and Meta-analysis.” Annals of Internal Medicine. 2024;177:1652-1663. DOI: 10.7326/ANNALS-24-00981.

4. Natali C, Marconi L, Dias Duran LD, Cabitza F. “AI-induced Deskilling in Medicine: A Mixed-Method Review and Research Agenda for Healthcare and Beyond.” Artificial Intelligence Review. 2025;58:356. DOI: 10.1007/s10462-025-11352-1.

5. Heudel PE, Crochet H, Filori Q, Bachelot T, Blay JY. “Artificial intelligence in medicine: a scoping review of the risk of deskilling and loss of expertise among physicians.” ESMO Real World Data and Digital Oncology. 2026;12:100693. DOI: 10.1016/j.esmorw.2026.100693.

6. European Union. “Artificial Intelligence Act: Article 14, Human oversight.” Regulation (EU) 2024/1689. AI Act Service Desk.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.