Blog

Think First, Ask Second: What the Evidence Shows About AI and Clinical Judgment

7 July 2026 By SwissMed AI

⏱️ Reading time: ~4 minutes

Here's the uncomfortable part about AI in diagnosis: it makes you better when it's right, and worse when it's wrong, and you often can't tell which one you're getting. A randomised JAMA study [1] put that to the test.

Three distinct risks are often confused:

  • Automation bias: accepting AI output without critically evaluating it

  • Deskilling: losing a skill you already have because AI is doing it for you

  • Never-skilling: never developing a skill because AI was always there first

All three are documented. All three need different responses.

AI-generated illustration.

🚨 When faulty AI is worse than no AI

In the JAMA study [1], 457 clinicians diagnosed the cause of acute respiratory failure from written cases, scored against expert reference diagnoses:

  • No AI: 73.0% correct

  • Standard AI: 75.9%

  • Biased AI: 61.7%, worse than no AI at all

  • Adding explanations to faulty AI output didn’t help significantly (64.0% vs. 61.7%, p=.37)

The same pattern recurs across specialties:

  • ECG (JAMIA 2003 [2], n=30): physicians agreed with a wrong computer diagnosis 67.7% of the time vs. 34.6% with no computer

  • ECG, experience gradient (J Electrocardiol 2018 [3], 30 physicians, 9,000 interpretations): cardiology fellows resisted incorrect AI better than non-cardiology fellows, but neither group was immune

  • Dermatology (Nature Medicine 2020 [4], ~300 clinicians): correct AI helped everyone, especially juniors; faulty AI degraded performance across experience levels, including experts

  • Prescribing (BMC Med Inform 2017 [5]): incorrect decision support raised commission errors by up to 65.8% vs. no support

  • Differential diagnosis (IJERPH 2021 [6], n=22, exploratory): AI-generated lists shifted physicians’ own differentials in at least 15% of cases

A 2012 systematic review [7] found workload, task complexity, and cognitive load worsen automation bias. It predates modern AI but its mechanisms remain plausible.

⏳ Deskilling and never-skilling

  • Colonoscopy (Lancet GE 2025 [8], 4 centres, Poland, observational): adenoma detection without AI assistance dropped from 28.4% to 22.4% after months of AI-assisted work. Observational, so causality isn’t proven, but the pattern fits reduced independent skill use

  • Scoping review (ESMO RWD & Digital Oncology 2026 [9]): radiology, pathology, endoscopy, and clinical decision support are flagged as areas of particular concern

  • Mixed-methods review (AI Review 2025 [10]): AI reduces existing skills and the opportunities to build new ones

  • A 2026 Nature Medicine perspective [17] argues trainees relying on AI before building competence may never develop foundational reasoning, and introduces “mis-skilling,” internalising wrong AI output as fact. A related Lancet piece [15] raises similar concerns. Both are conceptual arguments, not empirical findings.

🧠 A limit of the evidence

Most findings above involve perceptual tasks: polyps, ECGs, skin lesions, drug interactions, not the reasoning used daily in general practice or internal medicine. The Harada study [6] and Budzyń study [8] come closest to touching clinical reasoning, but neither covers it directly. A 2026 BMJ commentary [16] argues AI risks making clinical reasoning optional, but this too is an argument, not a controlled study. Evidence on AI’s effect on broader clinical reasoning remains limited.

⚠️ Who is most at risk

  • Physicians with lower diagnostic expertise are most likely to follow incorrect AI output unquestioned [11]

  • Trainees face added never-skilling risk if they use AI before developing skills independently [17]

  • Time pressure increases the severity of automation bias, even if not its frequency [12]

  • AI literacy training helps but isn’t enough: trained physicians (NEJM AI 2026 [13]) still showed a 14.0-point drop in diagnostic reasoning with incorrect AI advice (73.3% vs. 84.9% error-free); the study didn’t compare against untrained peers

✅ What helps

Correct AI genuinely helps: accuracy rose from 73.0% to 75.9% in the JAMA study [1], especially for less experienced clinicians [4]. Reasoning through a case before checking AI output is consistently framed as protective across the literature [7], though no single study isolates it as definitive. Tools that withhold recommendations when confidence is low cut automation bias by 41.7% in one model [14], not yet standard, but a meaningful criterion when choosing a tool.

💡 Five practical takeaways (SwissMedAI)

  1. Think first, then ask the AI: form your own differential before checking AI output.

  2. A wrong AI recommendation can actively harm your diagnosis: use it to pressure-test your thinking, not replace it.

  3. Document when you override the AI: it preserves independent reasoning and creates a record.

  4. Protect early training at programme level: sequencing (AI after competency) is a curriculum decision, not an individual one [17].

  5. New skills will be needed, which ones isn’t yet defined, but critically evaluating AI output will be one of them.

🔍 What we still don’t know

  • Do these risks apply to broader clinical reasoning, or just pattern-recognition tasks?

  • Are AI-heavy training programmes already producing reasoning gaps?

  • Can AI-induced skill gaps be detected before independent practice?

  • Once mis-skilling takes hold, can it be corrected?

📚 Sources

  1. Jabbour S et al. Measuring the Impact of AI in the Diagnosis of Hospitalized Patients. JAMA. 2023;330(23):2275–2284.
    https://doi.org/10.1001/jama.2023.22295

  2. Tsai TL, Fridsma DB, Gatti G. Computer Decision Support as a Source of Interpretation Error: The Case of Electrocardiograms. JAMIA. 2003;10(5):478–483. https://doi.org/10.1197/jamia.M1279

  3. Bond RR et al. Automation bias in medicine: The influence of automated diagnoses on interpreter accuracy and uncertainty when reading electrocardiograms. J Electrocardiol. 2018;51(6S):S6–S11.
    https://doi.org/10.1016/j.jelectrocard.2018.08.007

  4. Tschandl P et al. Human-computer collaboration for skin cancer recognition. Nature Medicine. 2020;26(8):1229–1234.
    https://doi.org/10.1038/s41591-020-0942-0

  5. Lyell D et al. Automation bias in electronic prescribing. BMC Med Inform Decis Mak. 2017;17:28.
    https://doi.org/10.1186/s12911-017-0425-5

  6. Harada Y et al. Effects of a Differential Diagnosis List of Artificial Intelligence on Differential Diagnoses by Physicians. Int J Environ Res Public Health. 2021;18(11):5562.
    https://doi.org/10.3390/ijerph18115562

  7. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. JAMIA. 2012;19(1):121–127.
    https://doi.org/10.1136/amiajnl-2011-000089

  8. Budzyń K et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10(10):896–903.
    https://doi.org/10.1016/S2468-1253(25)00133-5

  9. 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.
    https://doi.org/10.1016/j.esmorw.2026.100693

  10. Natali C, Marconi L, Dias Duran LD, Cabitza F, et al. AI-induced Deskilling in Medicine: A Mixed-Method Review and Research Agenda for Healthcare and Beyond. Artificial Intelligence Review. 2025;58(11).
    https://doi.org/10.1007/s10462-025-11352-1

  11. Kücking F, Hübner U, Przysucha M, et al. Automation Bias in AI-Decision Support: Results from an Empirical Study. Stud Health Technol Inform. 2024;317:298–304.
    https://doi.org/10.3233/SHTI240871

  12. Rosbach E, Ganz J, Ammeling J, Riener R, Aubreville M. Automation Bias in AI-Assisted Medical Decision-Making under Time Pressure in Computational Pathology. arXiv. 2024.
    https://arxiv.org/abs/2411.00998

  13. Qazi IA, Ali A, Khawaja AU et al. Automation Bias in Large Language Model–Assisted Diagnostic Reasoning among Physicians Trained in AI Literacy, A Randomized Clinical Trial. NEJM AI. 2026;3(5).
    https://doi.org/10.1056/AIoa2501001

  14. Wang DY et al. Artificial intelligence suppression as a strategy to mitigate artificial intelligence automation bias. JAMIA. 2023;30(10):1684–1691.
    https://doi.org/10.1093/jamia/ocad118

  15. Berzin TM, Topol EJ. Preserving clinical skills in the age of AI assistance. Lancet. 2025;406(10513):1719.
    https://doi.org/10.1016/S0140-6736(25)02075-6

  16. Nguyen Phu Nghia. AI is making clinical reasoning optional, and that should worry us. BMJ. 2026;393:s871.
    https://doi.org/10.1136/bmj.s871

  17. Ke Y, Jin L, Ong JCL et al. AI-induced never-skilling in medical education. Nature Medicine. 2026;32:1997–2006.
    https://doi.org/10.1038/s41591-026-04438-y

Liked this? Get new articles in your inbox.