Beyond Generation: Using Multimodal LLMs as Detectors of LLM-Generated Deepfake Facial Images
Researchers tested Multimodal LLMs (MLLMs) to detect fake images by looking for structural inconsistencies rather than relying on watermarks. GPT-5.2 showed high accuracy in finding physical flaws, proving structural reasoning is more reliable than watermarking. However, invisible watermarks can reduce detection accuracy in some models.