I could not agree more with EFF.

There’s a difference between training a model and using a model. Training involves copyrighted works but fair use is not just about use of copyrighted works, it’s about whether the use is transformative and substitutes the original market. I struggle to see how is not transformative under these criteria.

The use of the model (being able to output copies of GPL software) is a different question. This depends on the circumstances: if GPL code is exactly reproduced then it very well could be subject to the license of the original work.

I don’t understand the legal objections to the fair use of protected IP. Licenses are legal documents, not moral imperatives. GPL only exists because of copyright law, and you can’t write a license that supercedes copyright law if you don’t like the law.

The Claude Code example is completely different, hosting a repo with the leaked code is clearly not fair use.