A new study finds that CLIP, the vision-language model behind many AI tools, already knows how to spot AI generated images without any extra training.
Researchers built a pipeline that keeps a CLIP model frozen, extracts its visual embeddings, and trains only a small classifier on top. On the CIFAKE benchmark, that setup hit 95% accuracy without any language reasoning. Few-shot adaptation using just 20% of a custom, more varied dataset brought accuracy down to 85%. A closed-source baseline, Gemini-2.0, posted the best zero-shot accuracy overall but failed on specific image styles.
The interesting failure mode is the point: wide-angle photographs and oil paintings tripped up classifiers across the board, meaning detectors trained on typical AI outputs may miss entire categories of fakes. That matters more for platforms moderating images at scale than any single accuracy number. A 95% score on one benchmark says little about content that was never represented in that benchmark to begin with.
Treat any AI image detector, CLIP-based or otherwise, as a probabilistic guess rather than a verdict.