Julian, > If we treat LLM assisted contributions differently from non-LLM assisted contributions, then this is a somewhat implicit signal of distrust: you used an LLM, therefore we will penalize you with stricter reviews, require additional proof, etc. [...] > collaboration in large projects is very often trust based [...] > We can't force this disclosure anyway, so we simply assume that patch authors are truthful.
You see the discrepancy in this? This leads to the following bizarre situation (Conduct => Review): Truthfully discloses LLM generation => Additional scrutiny or stricter review, because we distrust you Does not disclose it => Receives ordinary review unless discovered Did not use an LLM => Receives ordinary review, may still be accused of LLM use. This creates an incentive structure in which honest contributors that comply with the honor system are automatically placed in the less-trusted category. You also hold the following four views concurrently: 1. Responsible LLM users should not be stigmatized. 2. How the code was produced is private and should not matter. 3. LLM generation should be disclosed because it justifies stricter review. 4. Such disclosure is not a weakness and does not make contributors second-class. However two conflicts with three, and three makes four difficult to sustain. And while the first is aspirational, the proposed review practices is the mechanism by which the stigma would arise. This is then justified by operational constraints (I think given infinite resources, this would be an ok position). Best, Moritz On Wed, 15 Jul 2026 at 23:20, Julian Ospald via ghc-devs < [email protected]> wrote: > >From what I understand you are worried that LLM users will be stigmatized > for their LLM use, even if their use is entirely reasonable, adheres to the > policy and the code is well understood, free of bugs and so on. > > If we treat LLM assisted contributions differently from non-LLM assisted > contributions, then this is a somewhat implicit signal of distrust: you > used an LLM, therefore we will penalize you with stricter reviews, require > additional proof, etc. > That may make contributors feel excluded or like second-class citizens, > even if they did all the work necessary and used LLMs responsibly. > > How code was "conceived" is entirely private and is no ones business: I > could have had a vision during my dream or found the code in a drift > bottle, deciphered it and are now submitting it to you. All that matters is > that I affirm: **I own the code, intellectually and legally... I can > explain it and I can reproduce it manually if need be.** > > I think given infinite resources, this would be an ok position. We would > process all MRs in the same fashion, mechanically and not assume anything. > But the reality is that collaboration in large projects is very often trust > based. Everyone knows that Simon likes to rewrite the simplifier and that > he's intimately familiar with it and so people would adjust their reviews > based on the context. If he would start delegating such rewrites to an > LLM... the situation would be quite different now, wouldn't it? Now it's > not about "does Simon know what he is doing?"... we know he knows the > subsystem, but it's suddenly harder to tell how much *effort* went into a > particular part of a patch! This is significant and we can't really ignore > this. > > But I am sure Simon would write that down in his MR happily: "I had claude > help me with this humongous task, can you people please look more carefully > over this part of the patch?". > > What this policy is saying in spirit (I think) is exactly that: please > help us understand which part of the patch needs more egregious review. > Which part are you very familiar with and confident about? So we must rely > on self-disclosure. We can't force this disclosure anyway, so we simply > assume that patch authors are truthful. > > And yet, I still think that the wording of the policy is largely fine: > > - P1 mentions that effort matters and reviewers time needs to be respected > - P2 says explicitly "We value the contribution itself, regardless of how > it was produced" > - P3 says we prefer human authorship, simply because we're more familiar > with the dynamics and the process... I think this is totally fair > - P4 asks for self-disclosure, but clearly says it's "an acknowledgement, > not a weakness" > > 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. > _______________________________________________ > 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]
