AI can spot unnecessary tests – whether it prevents them is open
A new study tests medical language models on 225 case examples. For Switzerland, the decisive hurdle lies between a correct answer and better treatment.
Can artificial intelligence also advise against an examination? A study published on 20 October 2026 in the «Annals of Internal Medicine» offers encouraging but limited evidence for this. Six commercial language models rarely recommended medically unnecessary measures in a test. Whether this actually leads to fewer tests or better treatment in a doctor's practice was not examined.
This gap is central for Switzerland. The smarter medicine campaign has campaigned against overuse and misuse for years. Its aim is not as little medicine as possible, but appropriate care. An AI system would therefore have to prove both: it may slow unnecessary workups, but must not prevent necessary ones. A low number of examinations alone would not be a mark of quality.
What the study actually tested
Dean Ariel, Stav Endelman and Guy Laban developed 225 scenarios from 15 Choosing Wisely recommendations. For each recommendation, ten cases required restraint and five a medical action. The models were queried via their programming interfaces between February and March 2026. Three runs per model produced a total of 4050 answers. No humans took part as patients.
In the cases designed for it, the rate of unnecessary recommendations to act was 5.4 percent. The reverse error, failing to recommend an appropriate measure, occurred less often at 0.6 percent. The percentages refer to answers within this test. They are not a success rate for arbitrary health questions and not a measured harm rate in care.
The average hides difficult decisions
Most overuse recommendations concentrated on ultrasound for incidentally discovered small thyroid nodules and routine electrocardiograms in low-risk adults. That does not mean ultrasound or ECG would be fundamentally superfluous. What matters are the conditions described in each case. The Swiss Society of Radiology accordingly stresses that recommendation lists must not prevent individually appropriate diagnostics.
The measurement method also sets limits. The researchers used single prompts without follow-up dialogue. One language model judged the answers, with medical clarification of disputed ratings. That is a controlled comparison, not a substitute for clinical testing. Whether a system reacts the same way with incomplete information or repeated pushing for a test therefore remains open.
Knowledge alone does not remove misaligned incentives
For the Swiss debate, a second look is therefore worthwhile: why does a hardly useful examination happen at all? The health economists Katharina Blankart and Tobias Benjamin Müller explain in a 2024 contribution from Bern University of Applied Sciences how financial incentives can influence medical decisions. Payment per service initially rewards the service delivered. Case-based lump sums in turn create other incentives, such as limiting costs per case.
This background piece is not an evaluation of the new AI study and does not automatically describe every current tariff. Its fundamental point remains important for assessing a digital tool: a correct hint does not by itself change remuneration, workflows or expectations. A hospital would have to clarify who checks a suggestion, justifies deviations and takes responsibility for the decision.
Services saved are also not automatically total costs saved. An assessment would have to include the introduction and operation of the system as well as additional workups following wrong recommendations. That would be a necessary next yardstick, not an already proven effect. The present work contains no Swiss cost calculation for this.
A comparison with existing care would also be important: a good answer in a benchmark does not show whether doctors decide better with AI than without it. That would require a suitable control group, comparable disease risks and follow-up observation. Only then could avoided overdiagnoses be distinguished from a mere shift in testing. These requirements concern practical introduction, not the validity of the test as a first research contribution.
The conversation remains the practical yardstick
Patients can already do something concrete today. smarter medicine recommends asking about treatment alternatives, their advantages and disadvantages, their likelihood, and the consequences of waiting. Add to that the question of what one can contribute oneself. This conversation aid aims at a shared decision, not at blanket rejection of an examination recommended by a doctor.
AI could prepare such a conversation or alert professionals to a recommendation. Whether it actually reduces overuse in Swiss practices would have to be shown by studies in everyday care. The convincing evidence would not be a particularly resolute digital no, but less unnecessary medicine with at least equally good care for the people who need treatment.
Sources
- Ariel, Endelman and Laban: original abstract of the benchmark study, PubMed
- smarter medicine: five questions before treatment
- Swiss Society of Radiology: recommendations and individual diagnostics
- Bern University of Applied Sciences: health costs and financial incentives, 25 November 2024
- Cover image: Daniel Sone / National Cancer Institute via Wikimedia Commons, public domain
- Article image: Harrison Keely / Wikimedia Commons
- License of the article image: Creative Commons BY 4.0
