- Detection is imperfect: Even when aware of AI involvement, radiologists achieved ~75% accuracy, meaning roughly 1 in 4 synthetic images were misclassified.
- Clinical realism is high: Diagnostic accuracy for abnormalities was similar in synthetic vs real images (~92% vs 91%), suggesting deepfakes can be clinically plausible.
- Experience does not protect: Years of experience and prior exposure to AI-generated images were not associated with better detection performance.
- Subtle visual cues matter: Common indicators included excessive symmetry, uniform noise, overly smooth bone contours, and atypical soft-tissue textures.
- AI is not a solution (yet): No tested LLM reliably detected all synthetic images, though GPT-4o and GPT-5 outperformed other models.
Source: Radiology
Daily News
Stay up to date with the latest clinical headlines and other information tailored to your specialty.
