I can't imagine a scenario where pre-training on someone else's works is fair-use, but distilling from a proprietary LLM isn't.