Overview
- The peer-reviewed Radiology study, published Tuesday, found AI-generated X-rays realistic enough to mislead experts and leading AI models.
- Seventeen radiologists from 12 centers reviewed 264 images split evenly between real and synthetic, with only 41% suspecting fakes before being told and average accuracy rising to 75% afterward.
- Four multimodal AI systems—OpenAI’s GPT-4o and GPT-5, Google’s Gemini 2.5 Pro, and Meta’s Llama 4 Maverick—reached only about 57% to 85% accuracy in spotting fakes.
- Researchers cataloged telltale signs of synthetic images, including overly smooth bones, unnaturally straight spines, overly symmetrical lungs, uniform vessel patterns, and unusually clean, one-sided fractures.
- The team warned of fraud and cyber risks and urged invisible watermarks and technologist-linked cryptographic signatures, released an educational deepfake dataset with quizzes, and cautioned that 3D fakes in CT and MRI are likely next.