AI image generators can now hide their training data — even after being compressed for deployment

Researchers found a way to detect whether a compressed AI image model was trained on copyrighted data, even when the model no longer remembers individual training images. When AI companies compress large models into faster versions for real-world use, they've been able to erase the fingerprints that prove what data was used — creating a legal loophole. This method closes that loophole by measuring whether the compressed model's overall output pattern still matches the original training data.

AI image models are now deployed through a compression pipeline called distillation, which strips away the evidence of what they learned from. This matters because copyright holders have no way to prove unauthorized training happened — the model laundering is technically invisible. The paper shows that even after compression, a statistical signature of the original training data survives in the model's output distribution. This doesn't solve the legal problem, but it does restore the ability to audit what data a deployed model actually came from, which is the prerequisite for any enforcement.