Radiologists who knew they might be looking at AI-made images did better (75% accuracy), but those kept in the dark only got it right 41% of the time.
Specialists in musculoskeletal imaging were a bit better at spotting fakes than others.
The researchers also tested popular AI models (including ChatGPT-4o) and their accuracy ranged roughly from 52% to 89% across the tests.
Lead researcher Dr. Mickael Tordjman says we need better tools and training so doctors can keep up with these ultra-realistic fakes, which often look too smooth and symmetrical compared to real scans.