@jamie
This is not my point. Even if it _is_ “fair use”: if the llm produces a 1:1 copy (minus some renamed variables) of some relevant piece of code it is not producing something “new”. As a human I can learn from any code (copyrighted or not), but I cannot just take the code, rename some variables and publish it as my own creation. I would loose in court.¹

So technically if you use an LLM to produce code for you you need to check if any relevant piece of it is a copy of anything that exists.

Clean room implementation requires the programmer to not have seen the original code but only the requirements.

__
¹otherwise you could just take any piece of copyrighted code, rename variables and say it is yours because an LLM has produced it.

@jamie

Additionally, AI generated code can be a copyright infringement if the AI basically generated a copy of some copyrighted code. And if we consider that AI is trained on lots of GPLed code there is a high probability it will generate code that would need to be licensed accordingly.

There is no clean room implementation of anything with AI. The code is immediately tainted.