Tom, I've been pondering that very same question. And--for myself--I'll just declare for every MR/Issue I'm going to open in the future that LLMs may have been used in the creation. This gives me the peace of mind that I'm not lying to anyone, while also allowing me to not have to spend excessive mental capacity and bookkeeping to ensure that I did not intentionally or unintentionally use any LLM (Copilot, ...) during the creation of any such contribution. I simply can not guarantee that over a span of a week or more, I didn't somehow end up using LLM tech these days.
I accept that the net result will be that I'll be viewed as a second class contributor; however I will not compromise my integrity by accidentally lying about LLM usage. I do not want to open myself up to the debate of weather or not LLMs have been used during the creation. There is no way I could even exhaustively prove that I haven't and the policy gives human-written work a presumptive normative advantage. The policy in essence reads to me like this: 1. LLMs are ubiquitous. 2. LLM usage is part of life. 3. We prefer human written code over AI assisted code. 4. AI assistance must be declared. This leads (with the other notes from the document to the following understanding--AI was used to phrase this hopefully in a way that's easier to understand and get my argument across): P4 calls disclosure “merely an acknowledgement, not a weakness.” But > elsewhere the policy says that: > > - human-written code and documentation are strongly preferred; > > > - reviewers may adjust their “feedback and discourse” after > disclosure; > > > - AI-generated submissions are associated with inadequate > understanding, reviewer burden, technical debt, de-skilling, and weakened > human relationships. > > Consequently, disclosure cannot realistically be completely neutral. It > communicates membership in a category that the policy has already > characterised as riskier and less desirable. A contributor who > over-discloses borderline autocomplete use may therefore suffer a > reputational or review disadvantage The policy states "I'm guessing that this MR was authored at least partly by an LLM", which puts suspicion on the faithful declaration of AI use by the author. The policy does not say A contributor’s good-faith declaration concerning LLM use should be > accepted unless there is specific evidence that it is false. Stylistic > impressions or unverified AI-detection results are not sufficient grounds > for adverse treatment. Consequently, because I can't prove to myself that I didn't use any LLM assistance, I will rather disclose that LLMs may have been used, to be on the safe side (“err on the side” of declaration). Subsequently I accept that I'll be viewed as a second class contributor. I guess that is why I find this policy (and discussion around) in part hostile to contributions. In an age of LLM proliferation, I think this will not attract fewer contributors. There is a book that has had a very lasting effect on me: https://en.wikipedia.org/wiki/Don%27t_Make_Me_Think. Best, Moritz On Thu, 16 Jul 2026 at 06:34, amindfv--- via ghc-devs <[email protected]> wrote: > Thank you for your work on this, Simon. > > This draft to me looks like an agreeable compromise. However, I'll call > out one bullet point which may require further discussion: under "P4: > Declare use of LLMs," you ask that "if you used an LLM in a substantive way > to generate your code or documentation, that should be acknowledged > explicitly." Bullet point #2 says "How substantive is "substantive"? 90%? > 50%? 10%? There is no hard and fast rule, but you will probably know." > > It's hard for me to imagine a scenario where a 90% or even 50% > AI-generated contribution doesn't merit a declaration of AI usage. I'd > suggest at _minimum_ dividing this number by 10, i.e.: "How substantive is > "substantive"? 9%? 5%? 1%?," but really even 9% or 5% AI generation does > not really qualify for the de minimis concerns of a person asking whether > using Copilot to generate a for-loop counts as AI usage. > > Setting 10% as the lower bound for "substantial" is also certainly not > what I'd want as a reviewer. My preference for these numbers would be more > along the lines of "How substantive is "substantive"? 2%? 1%? 0.1%?" > > If I were a new contributor using LLMs and saw that a 90% contribution > might be viewed as not worth acknowledging, I might be much less likely to > add an acknowledgement on my own contributions, and it might seem > unreasonable to me if a reviewer later takes issue with my not having > acknowledged a large LLM contribution. > > Thanks, > Tom > > > > On 07/16/2026 12:53 AM CEST Simon Peyton Jones via ghc-devs < > [email protected]> wrote: > > > > > > Thanks to everyone for their input. I have been editing the draft in > direct response, as you'll have seen. > > > > Moritz would like to make it non-LLM-specific, and helpfully offered > some concrete wording. But I think we really do need something that tackles > LLMs head on. (I tried to say "we value the contribution, regardless of how > it was produced", although I know that is not really satisfying to you.) > > > > Moritz thanks for the "no right to review" draft. I adapted it to give > it a more positive spin -- see under "Human conversation". > > > > Some would like a briefer policy, and I see merit in brevity. But there > is a lot of nuance here, as we can see in the debate, and my (fairly > strong) preference is to value clarity, precision, and explicitness over > length. > > > > I'm grateful that a few people with strong opinions have generously said > (I am paraphrasing) "It's not what I really want, but it's a reasonable > basis to move forward". That's what I meant about disagreeing agreeably. We > all truly want to work together, even though our views differ. > > > > We may well want to iterate the draft in the light of experience. > > > > I think we may have reached the point of diminishing returns, where we > can freeze the wording for now. Would you like to take a last look at the > current state, and say if there are any drafting changes you'd like to make? > > > > Thanks! > > > > Simon > > > > > > On Wed, 15 Jul 2026 at 18:18, Wolfgang Jeltsch via ghc-devs < > [email protected]> wrote: > > > Am Mi 15.07.2026 16:20 schrieb Julian Ospald: > > > > Given that I have way more strong opinions on this and can still get > > > > behind this policy, I would suggest that it’s a good deal for > > > > everyone. > > > > > > Given that I have way more strong opinions on this, I cannot get behind > > > this policy. That said, if the goal is to somehow move forward with > more > > > or less everybody still on board, then we likely don’t have any > > > considerably different option than using this policy that Simon > > > proposes. Maybe it would be a good next step to just adopt it for now > > > and check in a year how things have been going. > > > > > > All the best, > > > Wolfgang > > > _______________________________________________ > > > ghc-devs mailing list -- [email protected] > > > To unsubscribe send an email to [email protected] > > _______________________________________________ > > ghc-devs mailing list -- [email protected] > > To unsubscribe send an email to [email protected] > _______________________________________________ > ghc-devs mailing list -- [email protected] > To unsubscribe send an email to [email protected] >
_______________________________________________ ghc-devs mailing list -- [email protected] To unsubscribe send an email to [email protected]
