If someone were to train a coding-oriented LLM on only GPL code, I would assert that any code it outputs is covered by GPL because an LLM is fundamentally storing and reproducing its input, not understanding it and generating new things based on that understanding.