I was mainly inspired by this section:

> Specifically, the judge cited the study's observation that Copilot reportedly "rarely emits memorized code in benign situations, and most memorization occurs only when the model has been prompted with long code excerpts that are very similar to the training data."

That almost sounds like it'd be fine to train an "art transformation model" which takes an image and transforms it, which for all the frames of a specific Disney movie just so happen to output the very next frame...