AI Tricked Nearly Half of New Radiologists into a Fake Diagnosis

Here's the scenario: you go to the hospital with a tricky set of symptoms. A doctor uses an AI assistant to help think through possible causes. The AI might be wrong—and if the doctor is early in their training, they might trust it anyway. That's the finding of a new study that quietly added a fake disease called "neurocadmiumatosis" to an AI diagnostic tool.

The results were striking. Out of 41 doctors, 18 (44%) included that made-up disease in their final possible diagnosis list. But here's the twist: nearly all of those were doctors with six months or less of specialty training—69% of them fell for it. Zero percent of the more experienced radiologists did. In other words, experience acts as a shield against AI hallucinations.

Why should you care? Because AI tools are being rolled out in hospitals right now to help doctors diagnose faster. They can be genuinely useful, but they're also "black boxes"—even the creators don't always know why they produce a certain answer. When the AI is wrong, it's often confidently wrong. And newer doctors, who might not yet have the pattern recognition to spot an impossible disease, are the most likely to absorb that error into their final call.

The researchers call this "hallucination by proxy": the AI hallucinates, and the doctor passes it along. The fix isn't to ban AI—it's to teach doctors, especially trainees, to question AI suggestions and check them against real medical knowledge. Until then, this study is a warning: the computer's confidence doesn't make it right.