AI Chatbots Usually Know When a Medical Test Is Unneeded

When a doctor orders a scan or a blood panel, the instinct is to assume more testing means better care. AI Chatbots are now facing the same question — and a new study suggests they answer it surprisingly well. Researchers found that six leading large language models correctly recognized when a medical test was unnecessary in the vast majority of cases, only going off track in the gray zones where human doctors themselves disagree.

The joint research from Ben-Gurion University of the Negev and Clalit Health Services, published this week in Annals of Internal Medicine, tested AI systems built by Google, Anthropic and OpenAI against hundreds of clinical scenarios, collecting thousands of model responses in total. The work was led by Dean Ariel, a family physician and medical AI researcher, alongside Guy Laban of the university's industrial engineering department and Stav Endelman of Rabin Medical Center-Beilinson Hospital.

Putting the models on the exam table

To build a fair test, the team borrowed real recommendations from the Choosing Wisely campaign, an international initiative founded in 2012 by the American Board of Internal Medicine Foundation and now active in more than 30 countries, whose entire purpose is stamping out low-value medical care. Each scenario paired a case where an intervention was pointless with a closely matched case where the same intervention was medically necessary. A young man with sudden lower back pain after lifting furniture does not need imaging — but a woman with a history of breast cancer showing up with new nocturnal back pain does. A common viral cold does not warrant antibiotics; a confirmed bacterial strep throat does.

Why does restraint matter this much? Unnecessary tests routinely uncover harmless incidental findings that set off cascades of anxiety, specialist referrals, invasive biopsies and surgeries that were never needed — wasting health care resources and risking real patient harm. An overly cautious system that misses genuine danger is no better. Any AI tool giving medical guidance has to navigate between both failure modes, and until now nobody had measured whether the leading chatbots could.

The results were strong. Missed care was nearly nonexistent: when an intervention was medically indicated, the models wrongly advised skipping it in just 0.6% of responses. Overtesting was rare too — unnecessary tests or treatments were suggested in only 5.4% of cases where the guidance said they were unneeded. In eight of the fifteen recommendations examined, the models did not order a single unnecessary procedure, according to a report in Medical Xpress. For tools that learned medicine by ingesting human clinical literature — including what the researchers describe as medicine's less disciplined habits — the restraint was striking.

The gray zones where machines mirror human disagreement

Almost all of the errors landed in two notoriously ambiguous corners of practice. The models recommended unneeded EKGs for low-risk patients in 28.3% of responses, mostly when patients mentioned palpitations — a situation where an EKG can reasonably be part of the workup. They suggested unnecessary ultrasounds for small, incidentally discovered thyroid nodules in 42.8% of responses, echoing a genuine professional dispute: the American Thyroid Association and the American College of Radiology do not agree on whether such nodules should be imaged at all.

"One of the hardest parts of practicing medicine is knowing when to hold back," Ariel said. "Patients find comfort in an extra test, but an unnecessary MRI often catches an asymptomatic disk bulge, setting off unnecessary worry and treatment." The models absorbed medicine's literature wholesale, he noted — and, for the most part, learned restraint: "For the most part, they know when to stop, stumbling primarily where we doctors ourselves hesitate."

But the finding the researchers flag most urgently is that there is no single AI consensus in those gray areas. On the exact same clinical cases, one model recommended unnecessary tests in less than 10% of its responses while another did so 80% of the time. "People often hear an AI chatbot speak in a single authoritative voice, as if there is a unified 'AI consensus,'" Laban said. "Yet in gray areas where experts debate, the disparity between models was striking." Which chatbot you ask can change the answer you get — a warning for anyone treating one confident reply as settled science.

What it means for anyone asking a chatbot about symptoms

Young people already treat chatbots as a first stop for health questions, so the takeaway is practical rather than theoretical. When the answer is clear-cut, the leading AI Chatbots are about as disciplined as good medical practice — the risk sits in the ambiguous middle, where the same question can produce wildly different answers depending on which system you ask. Treat a chatbot's medical suggestion as a starting point for a conversation with a real clinician, not a verdict.

The team also warns about a social dynamic no benchmark fully captures. Telling an anxious patient that no test is needed is, in Laban's words, a delicate social act — and language models tend to go along with the person in front of them. Whether an AI assistant keeps its clinical restraint when a worried patient pushes back remains, in the researchers' framing, the next critical research question. Until that is answered, the most important medical skill may still be a human one: knowing when to say no, and meaning it.

The full study, "Overuse and False Restraint in Large Language Model Recommendations for Choosing Wisely Scenarios," appears in the journal with DOI 10.7326/annals-26-01667. For more on how digital tools are reshaping care, see recent reporting on virtual coaching for weight-loss drug users — and a reminder that medical data keeps surprising researchers: a spouse predicted severe COVID better than genes.