AI-Generated Medical Deepfakes Fool Radiologists and Detection Models

The rapid advancement of generative artificial intelligence has reached a critical threshold in medical diagnostics, where synthetic imagery has become virtually indistinguishable from authentic patient data. A study published in the journal Radiology by researchers at the Icahn School of Medicine at Mount Sinai demonstrates that AI-generated X-rays can successfully deceive both veteran radiologists and the very large language models (LLMs) used to create them. This development marks a significant turning point for the healthcare industry, shifting the conversation from the benefits of AI-assisted diagnostics to the profound risks of medical deepfakes and the potential for systemic exploitation of digital health records.

The study involved 17 radiologists from 12 hospitals across six countries, tasked with reviewing 264 X-ray images. Half of these images were synthetic, generated by tools such as ChatGPT and RoentGen. The results were startling: when the radiologists were unaware that the dataset contained fakes, they spontaneously identified only 41% of the AI-generated images. Even after being explicitly informed that synthetic images were present, their detection accuracy only rose to 75%. This suggests that even with a high degree of suspicion, a quarter of synthetic medical images would still be accepted as genuine by human experts, posing a direct threat to diagnostic integrity.

“

Detection accuracy among these models ranged from a low of 57% to a high of 85%.

Perhaps more concerning is the failure of advanced AI models to detect their own output. The researchers tested four major LLMs—OpenAI’s GPT-4o and GPT-5, Google’s Gemini 2.5 Pro, and Meta’s Llama 4 Maverick. Detection accuracy among these models ranged from a low of 57% to a high of 85%. Notably, GPT-4o, the model responsible for creating some of the deepfakes, failed to identify all of its own synthetic creations. This recursive vulnerability suggests that the industry cannot currently rely on AI as a foolproof gatekeeper against AI-generated misinformation. If the creators of these models cannot build reliable detection mechanisms, the barrier to entry for bad actors remains dangerously low.

The implications of this research extend far beyond the radiology suite, entering the realms of legal liability and insurance fraud. Dr. Mickael Tordjman, the lead researcher, noted that the ability to fabricate a fracture or a pathology that is indistinguishable from a real one creates a high-stakes vulnerability for fraudulent litigation. In an era where medical evidence is the cornerstone of multi-million dollar personal injury and malpractice claims, the introduction of undetectable synthetic evidence could undermine the entire judicial process. Insurance providers may soon find themselves in a technological arms race, needing to verify the provenance of every digital asset submitted in a claim.