While I don't think significant code produced by any popular LLM is realistically going to work without problems on the SNES, unfortunately I think it is still possible for someone to take a generative AI-branded program trained on copyrighted datasets they don't own to make some graphics and audio samples, and then scale it down themselves and convert the result into a format a SNES can understand before putting it in the ROM.

I'm sure this is possible, because pre-rendered genAI images are already not uncommon among other retro console dev spaces, especially in those targeting systems like the NES and Mega Drive.