It is silly, considering there is obviously much higher chance that code-generating LLM generates copy of existing copyrighted code than image-generating diffusion model generates copy of existing copyrighted image.