Can AI tell when patients don't need medical tests? Israeli study puts chatbots to the test

Some Israelis are willing to undergo any available tests to see if they are healthy or if a serious but preventable disease is lurking in their future. Then some decline all medical tests, from colonoscopies to mammographies, saying, “What will be will be.”

In an era where patients and clinicians increasingly turn to artificial intelligence for medical advice, a critical question emerges: Does AI know when not to test or treat?

A new joint study from Ben-Gurion University of the Negev (BGU) in Beersheba and Clalit Health Services, the largest public health fund, reveals that leading Large Language Models (LLMs) successfully identify which medical interventions are unnecessary and when they encounter gray areas in which medical consensus itself is divided.

The peer-reviewed study was published in Annals of Internal Medicine, the flagship journal of the American College of Physicians, titled “Overuse and False Restraint in Large Language Model Recommendations for Choosing Wisely Scenarios.”

Patients frequently request extra tests “just to be safe,” said Prof. Guy Laban of BGU’s Department of Industrial Engineering and Management and Dr. Dean Ariel, a family physician and researcher at Clalit Health Services, who worked with Dr. Stav Endelman of Rabin Medical Center-Beilinson Hospital in Petah Tikva.

“Yet in clinical practice, more testing is not always better medicine. Unnecessary tests often uncover incidental, harmless findings that trigger cascades of anxiety, specialist referrals, invasive biopsies, and surgeries that were never needed in the first place, wasting precious healthcare resources and risking patient harm,” asserted Laban and Ariel in a joint interview with The Jerusalem Post.

“Conversely, an overly conservative diagnostic system risks missing crucial interventions when real danger lurks. A clinically safe AI tool must skillfully navigate between these two failure modes.”

The researchers evaluated six leading LLMs developed by Google, Anthropic, and OpenAI across hundreds of clinical scenarios, analyzing thousands of model responses. The scenarios were constructed around recommendations from the international Choosing Wisely campaign, a global medical initiative founded in 2012 by the American Board of Internal Medicine Foundation and active in more than 30 countries to eliminate low-value care.

They were written and designed from the clinician perspective based on 15 Choosing Wisely recommendations. For each of the 15 recommendations, 15 scenarios were developed; for each of those, 10 low-value care scenarios were designed where restraint is appropriate.

For each recommendation, five scenarios were designed for which action is appropriate. Each paired case where an intervention was redundant had closely matched controls where the same intervention was medically necessary.

A young man with acute lower back pain after lifting furniture does not require imaging. However, a woman with a history of breast cancer presenting new back pain at night does.

A common viral cold with runny nose and cough does not warrant antibiotics, but a confirmed bacterial strep throat test does. When an intervention was medically indicated, the models mistakenly recommended withholding care in only 0.6% of cases, they asserted.

When an intervention was explicitly unneeded, the models inappropriately recommended testing or treatment in just 5.4% of cases.

In eight of the 15 recommendations, the LLMs did not recommend a single unnecessary procedure. The vast majority of AI over-testing errors clustered around two notoriously ambiguous clinical practices. Models recommended unneeded electrocardiograms (EKGs) in 28.3% of responses, primarily in cases where patients mentioned palpitations, where an EKG can be a reasonable part of the workup.

They also endorsed unnecessary ultrasounds in 42.8% of responses for small, incidentally discovered nodules, mirroring a real disagreement between the American Thyroid Association and the American College of Radiology over whether such nodules should be imaged at all, the researchers continued.

Does the study demonstrate genuine clinical judgment?

“One of the hardest parts of practicing medicine is knowing when to hold back,” explained Ariel. “Patients find comfort in an extra test, but an unnecessary MRI often catches an asymptomatic disc bulge, setting off unnecessary worry and treatment.

“Language models learned medicine from human clinical literature, including our less disciplined habits. We wanted to see if they also absorbed the ability to exercise restraint. For the most part, they know when to stop, stumbling primarily where we doctors hesitate.”

“People often hear an AI chatbot speak in a single authoritative voice, as if there is a unified ‘AI consensus,’” added Laban. “Yet in gray areas where experts debate, the disparity between models was striking: on the exact same clinical cases, one model recommended an unnecessary test less than 10% of the time, while another recommended it 80% of the time.

“Furthermore, telling a patient ‘you don’t need this test’ is a delicate social act. A human physician must stand firm before an anxious patient, whereas language models tend to go along with the person in front of them. The next critical research question is whether an AI assistant will maintain the correct clinical restraint when an anxious or insistent patient pushes back.”

Laban is unusual in that he has visited more than 60 countries worldwide and has a wide variety of interests. After a bachelor’s degree at Reichman University in Herzliya, he earned a doctorate in brain sciences at Scotland’s University of Glasgow that focused on how people communicate with robots and how such robots can help people with mental health.

He was a visiting researcher at Constructor University in Germany and Tel Aviv University and completed postdoctoral studies in the computer science department at England’s University of Cambridge. He is also an amateur nature photographer.

The central issue is not simply whether AI can recognize unnecessary medical interventions, but whether the study demonstrates genuine clinical judgment or merely the ability to match established medical recommendations in carefully constructed cases.

Asked how well these results would generalize beyond the particular recommendations tested, Laban said: “This is the most central practice among medical doctors, especially primary care doctors. When clinicians or patients consult with AI, they will get output with common practice. Hopefully with new models, we will see more and more online benchmarks.”

Asked if some primary care physicians are upset when patients come in grasping AI printouts about this or that disease, Laban said: “Nobody likes to be second-guessed. It can be disruptive, even unprofessional. But some use it in primary care. Doctors can’t avoid consulting with it, including whether medications they are prescribing conflict with others.”

The models were tested only in English, Ariel commented: “We structured information that was common across studies so they could be standardized.

More work will have to be done before the results can be clinically validated and used in the doctor’s office, added Ariel, who also teaches at Tel Aviv University’s Gray Faculty of Medical and Health Sciences.

The question of who would be accountable if an AI system advised against a necessary test and a patient got into trouble is tricky.

“Safeguards would be necessary. AI is a tool, like reading something in a book. In the end, the doctor is responsible. The model just recommends; it doesn’t tell him or her what to do,” Ariel declared. “The use of AI will have to be studied prospectively, and not only based on saving money. I’m not sure it’s cheaper than sending patients to the test it prevents.”