Re: GCC AI Policy Announcement

On Thu, Jul 30, 2026 at 9:40=E2=80=AFAM Jonathan Wakely <jwakely.gcc@gmail.=
com> wrote:
>
>
>
> On Thu, 30 Jul 2026, 08:27 Richard Biener, <[email protected]> w=
rote:
>>
>> On Wed, Jul 29, 2026 at 8:51=E2=80=AFPM Jonathan Wakely <jwakely.gcc@gma=
il.com> wrote:
>> >
>> >
>> >
>> > On Wed, 29 Jul 2026, 19:03 Richard Biener, <[email protected]=
> wrote:
>> >>
>> >>
>> >>
>> >> > Am 29.07.2026 um 17:30 schrieb Arthur Cohen <arthur.cohen@embecosm.=
com>:
>> >> >
>> >> > =EF=BB=BFHi Richard,
>> >> >
>> >> > What are the important LLM issues that you think the policy fails t=
o address? As the policy itself states, the it is expected to evolve and fe=
edback is welcome. The goal is to continuously improve it.
>> >>
>> >> The policy grounds its rejection of legally significant LLM contribut=
ions on legal issues (by means of naming).  It does not cover the IMO much =
more severe issues of overwhelming reviewers and keeping GCC maintainable.
>> >
>> >
>> > Such contributions are not allowed by the policy, so we don't need to =
burn reviewer time on them, and we don't need to worry that those contribut=
ions will make GCC unmaintainable.
>> >
>> > If the legal situation changes and somebody wants to argue for relaxin=
g the policy, they'll need to make a convincing case that it won't overwhel=
m reviewers or lead to code nobody understands. Until then, such discussion=
s don't seem productive too me (but that doesn't mean they can't happen any=
way).
>> >
>> >
>> >
>> >>   It also makes LLM assisted contributions by maintainers who might w=
ant to use LLMs to get to prototype stages fall into grey areas or even rej=
ected.  I realize the policy is not a discussion review document but what w=
as the point of having the policy group when the outcome is just the obviou=
s =E2=80=9Alegally unclear situation=E2=80=98 statement?
>> >
>> >
>> > Now at least it's written down and we can point to it as the policy. P=
reviously it was unclear whether this project cared about the legal issues.
>>
>> It still leaves "legally significant" up to interpretation.  When
>> somebody posts a LLM [assisted] patch that is legally significant or
>> when I let an LLM
>> do a prototype to assess viability of an idea, is not any non-LLM
>> implementation of the same solution (as vague as that is) a derivative
>> work of
>> a possibly infringing work?  So is posting LLM produced patches to GCC
>> able to DOS GCC development this way?
>
>
> Is that different to posting non-GPL code?

Of course not.

Richard.

>
>
>> Similar for the allowed
>> use of LLMs to analyze bugs or tooling.  How can that be legally sound?
>>
>> Richard.
>>
>> >
>> > As Arthur said, this isn't expected to be the final policy that lasts =
forever.
>> >