Ask a language model whether a patient needs a scan and the answer depends heavily on which model you ask. Israeli researchers put six of them to the test and found the divergence concentrates in exactly the cases where clinicians argue.
Clalit Health Services and Ben-Gurion University of the Negev built scenarios from the Choosing Wisely campaign, an international effort to discourage low-value tests. Each scenario appeared twice, once where the test was pointless and once where the same test was essential.
Across thousands of responses from models built by Google, Anthropic and OpenAI, the systems proposed an unnecessary test or treatment in 5.4% of cases where intervention was not warranted. They wrongly advised against needed care in 0.6% of cases where it was.
Those totals look reassuring. The pattern beneath them does not. Agreement between models collapsed on scenarios with contested evidence or thin guidelines, the same ground where two experienced physicians might part ways.
Dean Ariel, a family physician at Clalit who led the work, offered a clinical example. Take two backs. A young man who strained his lifting a couch almost never warrants a scan, whereas a woman previously treated for breast cancer whose back pain arrives at night and worsens deserves urgent investigation.
Ariel described the models as inheriting medicine from the text they trained on, including its less disciplined habits. They generally know when to stop, he said, and stumble where doctors hesitate.
The cost of over-ordering is rarely financial alone. An unnecessary MRI can surface a harmless herniation, which then drives referrals, anxiety and occasionally a procedure nobody needed.
The study suggests a specific role for these tools. They may be safest handling the clear cases and flagging contested ones for human judgment rather than resolving them alone.
