> Am 12.10.2026 um 17:04 schrieb Richard Sandiford <[email protected]>:
>
> [email protected] (Richard Kenner) writes:
>>> It seems like it would be difficult to relax the policy much without
>>> also relaxing the copyright/origin rules for contributions that don't
>>> use LLMs.
>>
>> Not really. The policy could take the view that the person
>> instructing the LLM is the copyright holder because they had enough
>> human interaction with the LLM to be able to claim copyright on that basis.
>> That would provide a person to do (a) or (b).
>
> Are you sure?
>
> I remember a case where, despite regular IP training, a colleague copied
> a piece of GPL code wholesale into commercial software (as part of a
> larger change). Fortunately that case was within the company's ability
> to rectify. For copyrightable changes, how could someone using an LLM
> be sure that that the LLM hadn't done the same?
>
> To put it another way: suppose I'd talked with that colleague over Slack,
> describing what was needed. Suppose we'd gone back and forth on it a bit
> and then the colleague had sent the final patch. How would I be able
> to ceritfy (a) or (b), either for myself or the colleague? In this
> hypothetical situation, I would have instigated and guided the
> conversation, but the work was done by the colleague and (in this case,
> and unknown to me) an uncredited third party whose work was under a
> different licence. Couldn't the same thing happen with an LLM?
I guess yes. Note this applies to ‚large‘ additions of functionally new code.
My experience is with bug analysis and fixing where the analysis part can be
done one-shotting and the bugfixing part either requires interaction and/or LLM
switching and I usually just fix the bug (or the suggested) fix myself. The
recent ifcombine fixes had LLM analyses for example - that really saves
precious time. Asking the LLM to summarize that analysis as commit message
also helps and makes them more verbose than you’d see from me usually.
But most fixes are obviously derived from GCC itself, but often more than those
11 lines of code from the policy. Thus me asking for common sense.
Richard
> Obviously GCC is a GPL project and so that particular example wouldn't
> necessarily be a problem. But IIRC there are GPL-v2-only forks of GCC.
> How could someone be sure that all the output of an LLM can be licensed
> as GPL v3 or above?
>
> I suppose the policy also has to be future proof. The trend is clearly
> to try to make LLMs do more with less coaxing, so the level of human
> input is likely to decrease over time.
>
> Still playing devil's advocate. :)
>
> Richard
