If you think it's an unacceptable risk to use a tool you can't trust when your own head is on the line, you're right, and you shouldn't use it. You don't have to guarantee anything. You just have to accept punishment.

Their position is probably that LLM technology itself does not require training on code with incompatible licenses, and they probably also tend to avoid engaging in the philosophical debate over whether LLM-generated output is a derivative copy or an original creation (like how humans produce similar code without copying after being exposed to code). I think that even if they view it as derivative, they're being pragmatic - they don't want to block LLM use across the board, since in principle you can train on properly licensed, GPL-compatible data.