Re: Coding assistants (LLM) policy

Olivier Dion <[email protected]> writes:

> I have made a tentative policy regarding the usage of LLM for the Guile
> project [0].  You can find a verbatim copy of this policy at the end of
> this email along with the Developer Certificate of Origin (DCO) that
> accompany it.

The comments about honor system are strange.  While I realize that
whenever there are rules, people might behave contrary to rules and not
disclose it, the honor system comment says more than is typical about
people that think LLMs are ok, or perhaps about what people who think
LLMs aren't ok think about people that use them :-)

I would therefore sugggest amending:

  Require disclosure/attribution of any text/code that came from an LLM.
  (Implicitly, consider violating this equivalent to submitting code
  written by others while claiming to have written it.)

while understanding that pro-LLM culture seems to at times lead to lying
about it.

Secondarily, the text talks about the DCO, and signing the DCO indicates
that the signer is the sole author or that non-signer content is
properly licensed (compatible open-source license) and that any
license-required attribution is in place.  It's an affirmative statement
that things are ok license wise.

With LLMs, it's an unsettled legal question as to whether LLM-generated
text is a derived work.  If it is, the contribution lacks a valid
license and the DCO statements are untrue.

By saying this is even perhaps ok, the project is deciding:

  Notwithstanding that whether LLM output is a derived work of training
  data is not a settled legal question, the project chooses to behave as
  it LLM output is *not* a derived work, and will accept that into the
  codebase.  Further, we do not have any requirements for tracking this,
  meaning that we will not have an obvious recovery plan if LLM output
  is rules to be a derived work of training data.


(I sort of thought guile was a "GNU project", and if so thought that the
entire GNU project was, at least for now, declining to accept LLM
contributions.)

> The above policy applies to code contributions.  However, I will take
> the time here to specify another policy which I will personally apply to
> the mailing lists and the Codeberg interface (issues and pull-requests).
>
> Codeberg and mailing lists are meant for human interactions.  They are
> not meant for statically generated texts nor for AI if you believe in
> that.  The internet is becoming full of these generated texts/images.
> Whether it is a cooking blog post or literally every pictures that my
> small (500 persons) town now post.  I found myself not liking surfing
> the web like I used to.  It is mostly all fake and gray and I think it
> will just become worse overtime.  However, there are still some places
> where it is good to be and I believe that the Guile community is such a
> place.  I would like to keep it that way.  That being said, people do
> what they want, but so do I.  Therefore, I will refrain to answer
> anything I think is LLM generated in emails and on Codeberg.

I think it's great you are taking that stance, and I'd urge the project
to adopt it as policy.

> Also.  Please out of respect to others and for your own, don't be a
> human proxy.  If I want a LLM generated interaction, I too can open a
> LLM prompt.  So please please please, don't copy/paste the interaction
> you just had with XYZ model.  I personally found this to be an insult.
> I am much more interested in chatting with you than reading LLM outputs.

It should be policy that in all project forums, representing LLM text
asd one's own (by declining to identify it) should be prohibited.
I have seen this in other project environments, and I consider it a
serious offense.