Julian,

What I've been trying to illustrate multiple times now is that the
incentives you are creating through this policy are not protecting you, but
to the contrary make people think twice if they want to be honest with you
and declare LLM use (however tiny that might have been, or not).  You are
even opening a game of "lets. see if I can use LLMs, not declare them, and
have Julian review them without noticing." What kind of incentive is this?

If all you want to be protected is a labeling, the policy should simply
read:

*We prefer contributions with non-trivial LLM assistance to be labeled
LLM-assisted.*

If you really wanted to you could even add a fairly neutral: It helps
reviewers set expectations.

This above wouldn't be enough for me though, as it does not say that we
discourage drive-by-llm-contributions, and expect people to fully own their
contributions. That however is completely irrelevant to you, as you don't
want to interact with any of those, and as such will disregard them as soon
as they are LLM-assisted labeled.

And there is zero need for any form of judgement in either direction.  You
can at your discretion just not review PRs that have the LLM assisted
label.  If someone explicitly asks you to review a patch that has
LLM-assisted, you can let them know that you'd prefer them to find another
reviewer.  Again, what I said before, there is *no right to review*.

If you goal is to make sure people disclose their LLM use honestly, let's
not have a policy that is outright stigmatizing people who use LLMs
(because they did not share the "we strongly prefer hand crafted bytes"
sentiment). If on the other hand you want GHC to be a project that actively
pushes away (or makes people question if their contribution is welcome if
they use LLMs -- even a hint of doubt is enough to discourage people, It's
the UD part of FUD) a certain class of contributors with integrity, honesty
but also occasional LLM use, write a policy that not only states LLM
labeling, but also comes with a judgement call.

Hence, and why I keep coming back to this, I don't see this as a
compromise, even though I can understand why you see it as such. I see this
as a judgement (and classification) of people, which is what I reject.

What is most perplexing to me is, that you multiple times say you want more
collaboration, but the incentives you seem to lay out, seem to contradict
this. The incentives seem to push people away, make them question whether
or not they are truly welcome or just tolerated.  Maybe I'm fundamentally
misunderstanding what you consider collaboration.  To me this is
fundamentally building trust between people.  Throwing code over the wall,
however conceived, is not that.  Entering into a discussion, while
providing contributions is closer to that.  However putting up incentives
to make contributors start out in the defensive is not that, insinuating
distrust is not that, it all just leads to a corrupted interaction, not
collaboration.  If one party always has to assume they could at any point
be accused of something, that they simply can not prove to be incorrect, we
create a hostile environment.  There is a reason why we generally follow in
dubio pro reo and assume good-faith.  This policy moves us from good-faith
not outright to bad-faith, but into that direction; I object to _that_.

Best,
 Moritz

On Thu, 16 Jul 2026 at 12:00, Julian Ospald via ghc-devs <
[email protected]> wrote:

> > You clearly state that LLM usage for you is a signal of distrust.
>
> Yes, I personally distrust patches with heavy LLM assistance and I do not
> believe people who say "but I know what I am doing".
>
> This also means that I don't want to spend my free time interacting with
> such works. So who is protecting me?
>
> The disclosure does that: it allows me to just walk away and have someone
> else look at the patch. And it definitely does protect you, because my
> negative bias won't be able to drive decisions (unlike in my own projects,
> where I outright reject such patches).
>
> I'm confused why you don't see that this is literally the only compromise
> that is possible. I won't be contributing to projects, where people do not
> want to disclose their LLM use, even if it's "well, I use it all the time".
> That helps me set the expectations.
> _______________________________________________
> 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]

Reply via email to