(IANAL)

A couple of problems with this article:

1. not written by a copyright lawyer
2. it is not established that code produced by AI is not derivative from the code it was trained on
3. chardet's code regeneration has not been confirmed as an actual clean room implementation as the AI writing the code has seen the original, and the AI writing the specs is actually reading the code and translating it to a different language (which is a derivative). In a clean room implementation, nobody has seen the original code.
4. US copyright office said that LLMs do not produce original work, therefore it cannot be copyrighted, it does not mean that what an LLM produces isn't copyrighted to begin with. It has been proven quite a few times that LLMs can produce material it was trained on verbatim
5. US copyright office only has a say over the US, not internationally

Does LLM break F/LOSS? too many unanswered legal questions; also too soon to tell any cultural/network effects.

Is LLM produced code safe to be used anywhere? given the number of unanswered legal questions, I highly doubt it.

Will F/LOSS projects die when they significantly adopt LLM outputs? Too soon to tell, but I agree there is a significant risk for this. Only other LLM users would probably contribute. If the project is valuable, it would be forked by people who would not allow the same level of LLM involvement.