can't be. if there is gpled code in the training data, all code must be gpl since the model is one work. and since it can't, that's a breach.

I recall articles saying that significant input/prompt may influence that. Whether this is true or not I can’t say.

That's not true. Until these issues have run through the courts, we won't know what is or isn't legal. The biggest court case I'm familiar with atm is the New York Times' case that has yet to be resolved, and even then the ruling there will probably be too narrow to decide anything for code licensing. Hell, OpenAI/Microsoft might even settle to avoid a bad ruling, leaving all those questions unanswered.

So the Tl;Dr is that anyone using LLMs to code (in any capacity) is taking a risk, since the legality of it is still uncertain. It's not outside the realm of possibility that some court will rule that LLMs trained on stolen data are illegal, and the outputs of those models are infringing.