RFC: The Binutils and LLM generated patches

[Paul Koning]

> On Oct 2, 2025, at 10:48 AM, Collin Funk <collin.funk1@gmail.com> wrote:
> 
>> ...
> 
> There was discussion on banning AI assisted contributions in glibc that
> is currently still not resolved [1].
> 
> I think that ideally the two projects should have the same/similar
> policy if possible.

Agreed, and more strongly, consistent among all the major GNU projects.  It would be odd if GCC, for example, were to pick a different policy.

> I responded with my thoughts in that thread, which are similar to yours.
> Given the lack of case law, I am not confident that output of LLMs is
> copyrightable.

IANAL, but I like to refer to LLMs as "derived work generators" because they inhale stuff (which is nearly 100% subject to copyright, in our subject domain) and transmogrify that through some unknowable algorithm to produce outputs.  Admittedly for a given output it may be infeasible to say just what inputs it was derived from, but the nature of LLMs is that it was indeed derived.  I think this is why company lawyers are wary of using code generated by LLMs, unless it was a private LLM trained only on training material that is owned by that same company.

	paul