Should there be a clause for AI?

[Aaron Wolf]

I like the idea of addressing AI issues within copyleft-next. AI is the 
most important dilemma we all face here.

That "Training Public License" seems to have a potential flaw by my 
first read. It says " you must release all resulting models, weights, 
and related code under the GPLv3 or later" but what does "release" mean? 
There's no Affero-type clause, so keeping the AI running on private 
servers even if giving others access might not count. And without 
anything specific about "release", it doesn't say to whom. It doesn't 
say that it must be published and accessible to anyone in the general 
public. On the opposite end, "release" mandates could suggest that every 
single iteration of the AI must be published, which is infeasible.

In order to deal with the issue of copyleft code used in AI training, I 
imagine a need for more clarity, as in how GPLv3 family describes 
"convey" but furthermore also clarifying when what is conveyed to whom.

I will share a related perspective on the big-picture in that other new 
thread on copyleft overall

Cheers,

Aaron

On 7/11/25 7:31, Ben Cotton wrote:
> On Fri, Jul 11, 2025 at 4:01 AM Vasileios Valatsos<me at aethrvmn.gr> wrote:
>> I apologize in advance for opening what is, in my regard, a giant can of
>> worms.
> It's a conversation that we can't avoid. Credit to you for standing up. :-)
>
>> I wonder if (a) this would make sense in copyleft-next, (b) if it even
>> belongs in the conversation, (c) if there is a better way to tackle this.
> Would this be field of use discrimination? I think it would, and would
> therefore violate FSF Freedom 0 and OSD criteria 6. That would render
> copyleft-next neither Open Source nor Free in the capital-letter sense
> of the terms.
>
> More theoretically, I'm a big fan of "don't make rules you can't/won't
> enforce." If you release something under the TPL and I decide to use
> it to train an AI model, how will you know? bkuhn will correct me if
> I'm wrong, but that would seem to be one challenge in GPL enforcement
> in the traditional software arena. At least with traditional software,
> there's typically some external evidence that a project was used in
> violation of its license. I'm not sure there's a good way to detect it
> in an LLM unless I make my training data public (which I would not if
> I were intending to violate the license).
>
> I'm not suggesting we stick our heads in the sand, but large language
> models are a complex problem from a variety of legal (I say as
> not-a-lawyer), ethical, practical, etc standpoints. I don't see a
> clear way forward to addressing them in a license like copyleft-next
> at this point.
>
> --
> Ben Cotton (he/him)
> TZ=America/Indiana/Indianapolis
> _______________________________________________
> next mailing list
> next at lists.copyleft.org
> https://lists.copyleft.org/mailman/listinfo/next
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.copyleft.org/pipermail/next/attachments/20250711/843b78e8/attachment.html>