I'm not sure this really qualifies as anything new. It's been well known for a long time that many generative AI models are capable of reconstructing copyrighted works (copilot, SD, etc.) from their original training data with nearly 100% fidelity, but until we have a definitive legal ruling it doesn't really matter.
Basically the AI doesn't have a copyrighted image in its data set, but it will go grab YouTube screenshots or images from Google Images or something.