A new paper on arXiv, titled "Beyond Generation: Using Multimodal LLMs as Detectors of LLM-Generated Deepfake Facial Images," proposes a shift in how we catch AI-generated faces. The authors note that image generation models are now producing highly realistic content, and conventional forensic methods are no longer reliable enough to keep pace.

The paper's abstract outlines an investigation into whether multimodal large language models (MLLMs) can serve as detectors instead. Rather than relying on pixel-level artifacts or statistical traces, the idea is to leverage the broad visual and reasoning capabilities of MLLMs to judge whether a face is synthetic.

Because the abstract is truncated, the article cannot report on the study's findings, methodology details, or accuracy. What is clear is the motivation: as generation improves, detection must move beyond traditional forensics, and MLLMs are being explored as a possible answer. Further reading of the full paper would be needed to assess whether the approach works.