AI-generated images can now be traced back to the model that made them, even without watermarks.
Researchers have developed a post-hoc framework for identifying which image autoregressive model produced a given image. These models, which generate visuals by predicting one token at a time — borrowing the same approach behind large language models — have grown capable enough to produce photorealistic results that are visually indistinguishable from real photographs. The key finding is that the generation process itself leaves characteristic patterns in the output, patterns that persist even when no explicit watermark is embedded. The framework detects those patterns without requiring any changes to the model or its outputs.
That last part matters. Most existing provenance tools depend on watermarking baked into the generation pipeline, which means content already published without those marks is effectively untraceable. This approach works retroactively, covering models that never integrated watermarking in the first place — a much larger share of what is actually circulating online. The practical targets are misinformation, fraud, and attribution of harmful content.
The approach is promising, but the adversarial question remains open: if these fingerprints are detectable, they are also, in principle, removable. Watermarks have faced exactly that challenge for years, and a passive detection method is only as durable as the patterns it relies on staying intact through cropping, compression, and deliberate scrubbing.