What AI in Healthcare Can and Cannot Do
Software now reads scans, writes up consultations and warns of sepsis, and its record differs sharply from one job to the next. Where randomised trials exist the gains are measurable, and where they do not, the failures have reached patients first.
Answer to the question
Can AI diagnose illness better than a doctor?
At specific tasks, alongside a doctor. In a Swedish trial of more than 105,000 women, AI-supported breast screening found about 29 per cent more cancers without more false alarms. Prediction is weaker: in 2021 Andrew Wong and colleagues found a widely used sepsis model missed 67 per cent of sepsis cases at Michigan Medicine.
Key takeaways
- In a Swedish trial of more than 105,000 women, AI-supported breast screening found about 29 per cent more cancers, with no rise in false alarms and 44 per cent less reading for radiologists.
- A 2024 review of 44 randomised trials found AI-assisted colonoscopy raised the adenoma detection rate from 36.7 to 44.7 per cent, with no change in advanced growths per procedure.
- In 2021 Andrew Wong and colleagues found the widely used Epic Sepsis Model missed 67 per cent of sepsis cases at Michigan Medicine while alerting on 18 per cent of all patients.
- Keeping a doctor in charge is not automatically protective: in Thomas Dratsch's 2023 study, very experienced radiologists rated 82 per cent of mammograms correctly with right AI suggestions and 46 per cent with wrong ones.
This is general information on AI in healthcare research, not medical advice. See a qualified professional; for emergencies, 999 or 911. Medical review:.
Would you want software to help read your mammogram? In a Swedish screening trial of more than 105,000 women, half had their scans sorted and checked with help from an AI system and half had the usual two radiologists. The AI group had about 29 per cent more cancers found, the radiologists' reading load fell by 44 per cent, and false alarms did not rise. Trials like that one make a good case for AI in healthcare as a second reader in screening. Tools that predict who will fall ill have a far weaker record, and they deserve doubt until they have been checked on patients outside the hospitals that built them.
Where does AI help?
Colonoscopy has the deepest evidence. Pooling 44 randomised trials and more than 36,000 procedures, a 2024 review in Annals of Internal Medicine found that computer-aided detection raised the share of patients with at least one adenoma, the kind of polyp that can turn into cancer, from 36.7 to 44.7 per cent. Advanced growths per colonoscopy, the ones closest to becoming cancer, stayed about the same, and almost two extra polyps were removed for every ten procedures that turned out not to be precancerous.
Breast screening is the other strong case. The Swedish trial, known as MASAI, reported a second result in the Lancet in 2026. Over the two years after screening, cancers that surfaced between screens were no more common in the AI group, and its screening caught about 81 per cent of the cancers present, against 74 per cent with two radiologists.
Software that listens to a consultation and drafts the doctor's note has spread quickly. A 2025 randomised trial in a large academic health system in California gave 238 outpatient doctors one of two of these ambient scribes, or neither. With one product, time spent writing notes fell by 9.5 per cent compared with no scribe; with the other, it did not change. Users of both reported that clinically significant errors in the drafts turned up "occasionally".
Most AI in healthcare reaches patients through a regulator. The FDA's list of AI-enabled devices, in its September 2026 version, names 1,614 products, about three in four of them for radiology, and about 96 per cent reached the market through the 510(k) route. That route asks a maker to show a device is "substantially equivalent" to one already legally sold, a comparison with an older product rather than proof that patients end up healthier.
Where does AI fail?
A sepsis warning tool in use at hundreds of US hospitals missed two in three of the cases it was meant to catch. That was the result when Andrew Wong and colleagues checked the proprietary Epic Sepsis Model against 38,455 hospital stays at Michigan Medicine in 2021. Sepsis, the body's runaway response to an infection, occurred in 2,552 of those stays. The model failed to flag 67 per cent of them while sending alerts on 18 per cent of all the patients in hospital, and the authors wrote that its ability to identify sepsis had not been adequately evaluated despite its widespread use.
A sepsis warning tool in use at hundreds of US hospitals missed two in three of the cases it was meant to catch.
A check of this kind, on new patients in a new place rather than on the data a tool was built from, is known as external validation. For anyone who has had a worrying symptom dismissed or delayed, the Michigan study is uncomfortable reading, because the tool was in wide use before such a check appeared.
Dermatology supplies the second warning. At Stanford in 2022, Roxana Daneshjou's team assembled 656 biopsy-confirmed skin images, matched across light and dark skin, and ran three published AI models on them. Each scored below its developers' reported accuracy, and lower on dark skin than on light; one caught 69 per cent of cancers on light skin and 23 per cent on dark. The dermatologists in the study also struggled with dark skin. Once two of the models were retrained on the new images, the gap closed, which places much of the problem in what they had learned from. This is algorithmic bias, a tool serving the patients it was trained on better than everyone else.
| Task | Study | What happened |
|---|---|---|
| Breast screening | Swedish trial, 2025 and 2026 | About 29 per cent more cancers found, no rise in false alarms |
| Colonoscopy | Review of 44 trials, 2024 | More adenomas found, advanced growths per procedure unchanged |
| Note writing | California trial, 2025 | One of two scribes cut note time by 9.5 per cent |
| Sepsis warning | Michigan study, 2021 | 67 per cent of cases missed |
| Skin cancer photos | Stanford study, 2022 | One model's detection fell from 69 to 23 per cent on dark skin |
Should doctors keep control?
Keeping a person in charge is the usual answer, and it has a careful defender in Guardrails, the 2024 book by Urs Gasser and Viktor Mayer-Schönberger on guiding human decisions in the age of AI. They argue that the structures around a decision should mark its limits rather than take it over. Guardrails, they write, "show you where the edge is, making it less likely that you'll step over without meaning to. But they aren't like prison walls" (page 3). Their sixth chapter turns this into a design principle, that a good guardrail should empower the person deciding rather than supplant them, and they present it as a value they hold.
In a clinic, that principle means software that informs a doctor's judgement and leaves the decision with someone accountable. Whether the arrangement also produces better results is a separate question, and the research on it is mixed.
Wrong suggestions from AI in healthcare can pull even experienced readers off course. In a Cologne experiment published in 2023, 27 radiologists read 50 mammograms alongside advice from what they were told was an AI system, and some of that advice was deliberately wrong. Very experienced readers rated 82 per cent of scans correctly when the advice was right and 46 per cent when it was wrong. For the least experienced, the figure fell from 80 to 20 per cent. The study, led by Thomas Dratsch, measured automation bias, the pull to trust a machine's suggestion over your own reading, and experience softened it without removing it.
Very experienced readers rated 82 per cent of scans correctly when the advice was right and 46 per cent when it was wrong.
Four Polish endoscopy centres that brought in AI polyp detection at the end of 2021 offered a different test: what happens when the machine is off. Comparing colonoscopies done without AI in the three months before and after the tools arrived, Krzysztof Budzyń and colleagues found that the share of patients with an adenoma detected fell from 28.4 to 22.4 per cent. Widely reported in 2025, the result is an early clinical sign of deskilling, the loss of a skill through reliance on a tool.
Both results bear on the book's case for keeping doctors in charge of AI in healthcare. Its argument that people should keep the right to decide stands as a value. Its hope that a person in the loop will improve the outcome gets little support here, because a doctor who checks AI can be steered by it and can lose practice without it.
What should patients ask?
The colonoscopy review did not measure deaths from bowel cancer, so finding more adenomas has yet to be shown to save lives. Only one of the two scribes saved time, and the reports of less exhaustion among their users were secondary results that the authors said needed confirming. The Polish deskilling study was observational, so other changes over the same months could explain part of the fall. The FDA describes its own device list as not comprehensive.
A patient can still ask useful questions about AI in healthcare. Is software involved in reading my test? Has it been checked on people like me? Who looks at its result, and can they overrule it? A clinician using such a tool should be able to answer all three.
Sweden's screening trial also counted the work. Its AI-supported arm needed 61,248 screen readings by radiologists against 109,692 for standard double reading, about 44 per cent fewer, while still finding more cancers than the standard arm did over the same period, April 2021 to December 2022.
Common questions
Can AI diagnose illness better than a doctor?
At specific tasks, alongside a doctor. In a Swedish trial of more than 105,000 women, AI-supported breast screening found about 29 per cent more cancers without more false alarms. Prediction tools have done worse: a 2021 Michigan study found a widely used sepsis model missed 67 per cent of sepsis cases.
Does relying on AI in healthcare make doctors worse at their jobs?
It may. In Krzysztof Budzyń's 2025 study at four Polish centres, the rate at which doctors found adenomas in colonoscopies done without AI fell from 28.4 to 22.4 per cent after AI tools arrived. The study was observational, so it cannot show that the AI caused the fall.
Is AI in healthcare less accurate on darker skin?
In dermatology it has been. Roxana Daneshjou's 2022 study of 656 biopsy-confirmed images found three published models performed worse on dark skin, with one catching 69 per cent of cancers on light skin and 23 per cent on dark skin. Retraining two models on more varied images closed the gap.
Sources
Hernström, V., Josefsson, V., Sartor, H., Schmidt, D., Larsson, A.-M., Hofvind, S., Andersson, I., Rosso, A., Hagberg, O., & Lång, K. (2025). 'Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study.' The Lancet Digital Health, 7(3), e175-e183. doi:10.1016/S2589-7500(24)00267-X. Abstract read: 105,934 women randomised; cancer detection 6.4 against 5.0 per 1,000, ratio 1.29; recall and false-positive rates not significantly higher; screen-reading workload reduced by 44.2 per cent.
Gommers, J., Hernström, V., Josefsson, V., Sartor, H., Schmidt, D., Hjelmgren, A., Larsson, A.-M., Hofvind, S., Andersson, I., Rosso, A., Hagberg, O., & Lång, K. (2026). 'Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study.' The Lancet, 407(10527), 505-514. doi:10.1016/S0140-6736(25)02464-X. Abstract read: interval cancer rates 1.55 against 1.76 per 1,000, non-inferior; sensitivity 80.5 against 73.8 per cent; specificity 98.5 per cent in both groups.
Soleymanjahi, S., Huebner, J., Elmansy, L., Rajashekar, N., Lüdtke, N., Paracha, R., Thompson, R., Grimshaw, A. A., Foroutan, F., Sultan, S., & Shung, D. L. (2024). 'Artificial intelligence-assisted colonoscopy for polyp detection: a systematic review and meta-analysis.' Annals of Internal Medicine, 177(12), 1652-1663. doi:10.7326/ANNALS-24-00981. Abstract read: 44 RCTs, 36,201 cases; ADR 44.7 against 36.7 per cent, RR 1.21; advanced colorectal neoplasia per colonoscopy 0.16 against 0.15; almost 2 extra non-neoplastic polyps resected per 10 colonoscopies.
Lukac, P. J., Turner, W., Vangala, S., Chin, A. T., Khalili, J., Shih, Y. T., Sarkisian, C., Cheng, E. M., & Mafi, J. N. (2025). 'Ambient AI scribes in clinical practice: a randomized trial.' NEJM AI, 2(12). doi:10.1056/AIoa2501000. Abstract read: 238 outpatient physicians, 14 specialties; Nabla reduced time-in-note by 9.5 per cent against control; DAX no significant change; clinically significant inaccuracies noted 'occasionally'.
US Food and Drug Administration. 'Artificial Intelligence-Enabled Medical Devices' list, page modified 4 September 2026, and 'Premarket Notification 510(k)'. Full text read: 1,614 listed devices, 1,230 in radiology, 1,553 with 510(k) submission numbers; the FDA states the list is not a comprehensive resource; a 510(k) demonstrates a device is 'substantially equivalent' to a legally marketed device.
Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M., & Singh, K. (2021). 'External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients.' JAMA Internal Medicine, 181(8), 1065-1070. doi:10.1001/jamainternmed.2021.2626. Abstract read: 27,697 patients, 38,455 hospitalisations; sepsis in 2,552; AUC 0.63; 1,709 sepsis cases (67 per cent) not identified; alerts for 6,971 of 38,455 (18 per cent); 183 identified who did not receive timely antibiotics.
Daneshjou, R., Vodrahalli, K., Novoa, R. A., Jenkins, M., Liang, W., Rotemberg, V., Ko, J., Swetter, S. M., Bailey, E. E., Gevaert, O., Mukherjee, P., Phung, M., Yekrang, K., Fong, B., Sahasrabudhe, R., Allerup, J. A. C., Okata-Karigane, U., Zou, J., & Chiou, A. S. (2022). 'Disparities in dermatology AI performance on a diverse, curated clinical image set.' Science Advances, 8(32), eabq6147. doi:10.1126/sciadv.abq6147. Full text read: 656 images; three models below their reported ROC-AUC; DeepDerm sensitivity 0.69 on Fitzpatrick I-II against 0.23 on V-VI; dermatologists also worse on dark skin; fine-tuning DeepDerm and HAM10000 on DDI closed the gap.
Dratsch, T., Chen, X., Rezazade Mehrizi, M., Kloeckner, R., Mähringer-Kunz, A., Püsken, M., Baeßler, B., Sauer, S., Maintz, D., & Pinto dos Santos, D. (2023). 'Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance.' Radiology, 307(4), e222176. doi:10.1148/radiol.222176. Abstract read: 27 radiologists, 50 mammograms, 12 incorrect suggestions among 40; correct ratings 82.3 against 45.5 per cent for very experienced and 79.7 against 19.8 per cent for inexperienced readers.
Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H. O., Blom, J., Buszkiewicz, M., Halvorsen, N., Hassan, C., et al., & Mori, Y. (2025). 'Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study.' The Lancet Gastroenterology & Hepatology, 10(10), 896-903. doi:10.1016/S2468-1253(25)00133-5. Abstract read: four Polish centres in the ACCEPT trial; 1,443 non-AI colonoscopies; ADR 28.4 per cent before against 22.4 per cent after AI exposure.
Gasser, U., & Mayer-Schönberger, V. (2024). Guardrails: Guiding Human Decisions in the Age of AI. Princeton University Press. Read in full by this publication, with a reliability audit. Ch. 6; p. 3.