Assuming ML models are causal, then bits of GPL code that fall out of the model have to have the color GPL, because the only way they could've gotten there was

Assuming ML models are causal, then bits of GPL code that fall out of the model have to have the color GPL, because the only way they could've gotten there was to train the ML using GPL-colored bits. It seems to me like the answer here is pretty obvious, it doesn't really matter how you copy a work.