Re: Ballot option: Allow AI-Assisted Contributions

--00000000000028546906583e43e9
Content-Type: text/plain; charset="UTF-8"

Hi Lucas,

On 04/08/26 at 14:55 +0000, Nilesh Patra wrote:
> > On Wed, Jul 22, 2026 at 09:31:10PM +0200, Lucas Nussbaum wrote:
> > > 1. **Legal Compatibility:** Contributors should ensure that the terms
> and
> > >    conditions of the generative AI tool do not impose contractual
> > >    restrictions that conflict with the distribution, modification, or
> use of
> > >    the output in the context of Debian.
> >
> > I am not sure how one fully could ascertain that. Is code generated by
> > ChatGPT/claude/codex/cursor/gemini (which most of the LLM users make use
> > of these days) satisfy this criteria?
> >
> > Similar to https://wiki.debian.org/DFSGLicenses would it be possible to
> > have a list of such LLMs which are acceptable?
>
> AFAIK, none of the commonly used tools or APIs make claims over
> copyright or other aspects that would be a problem (so they are all OK
> AFAIK).
>

Lawyers that I have spoken with are of the opinion that AI providers
(either model weight providers or service providers) can not *really*
legally claim any kind of copyright on the outputs of the AI models, so
they are very explicitly NOT doing that. The outputs are quite clearly
derived from your input so *if* any copyright applies at all, it will be
related to input and context (like the project being analysed to be
modified). Barring major legal decisions on this front, this is the
(relatively undefined) status quo.

However, both AI model and AI service providers can and do force users to
agree to terms of service to be able to use their models or services.
Conditions in those TOS documents varies widly and a few *can* be
problematic for free software development, especially when sensitive
information is concerned, like pre-announcement security voulnerabilities.
For example, Grok wants to have a full "irrevocable, perpetual,
transferable, sublicensable, royalty-free, and worldwide" license to all
content generated by your queries ...

https://www.termsfeed.com/blog/ai-terms-service-fine-print/ is a pretty
interesting write-up on some of these terms.

It is important to understand that TOS work very differently from copyright
and copyleft licensing - TOS is *only* binding on the user directly using
the service. It is *not* binding (and can not be made binding) to any third
parties that only get access to the *output* of the models and are not
using the model itself. On the other hand TOS can and do contain clauses
that allow the issuer of TOS to modify it at any time. This does noting to
the outputs already generated, but may restrict future interactions of the
user with the service or even with a downloaded model. Usage of a
downloaded model with a one-sided change clause *can* become in violation
of its TOS later on.

So, IMHO - yes, if a proposal allowing AI use in Debian in any form does
pass, it will be useful to create and maintain a reference guide on which
AI models and services have "Debian-compliant TOS" and which don't an
why/how specifically.

-- 
Best regards,
    Aigars Mahinovs

--00000000000028546906583e43e9
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr">Hi Lucas,<br><br><div class=3D"gmail_quote gmail_quote_con=
tainer"><blockquote class=3D"gmail_quote" style=3D"margin:0px 0px 0px 0.8ex=
;border-left:1px solid rgb(204,204,204);padding-left:1ex">On 04/08/26 at 14=
:55 +0000, Nilesh Patra wrote:<br>&gt; On Wed, Jul 22, 2026 at 09:31:10PM +=
0200, Lucas Nussbaum wrote:<br>&gt; &gt; 1. **Legal Compatibility:** Contri=
butors should ensure that the terms and<br>
&gt; &gt;=C2=A0 =C2=A0 conditions of the generative AI tool do not impose c=
ontractual<br>
&gt; &gt;=C2=A0 =C2=A0 restrictions that conflict with the distribution, mo=
dification, or use of<br>
&gt; &gt;=C2=A0 =C2=A0 the output in the context of Debian.<br>
&gt; <br>
&gt; I am not sure how one fully could ascertain that. Is code generated by=
<br>
&gt; ChatGPT/claude/codex/cursor/gemini (which most of the LLM users make u=
se<br>
&gt; of these days) satisfy this criteria?<br>
&gt; <br>
&gt; Similar to <a href=3D"https://wiki.debian.org/DFSGLicenses" rel=3D"nor=
eferrer" target=3D"_blank">https://wiki.debian.org/DFSGLicenses</a> would i=
t be possible to<br>
&gt; have a list of such LLMs which are acceptable?<br>
<br>
AFAIK, none of the commonly used tools or APIs make claims over<br>
copyright or other aspects that would be a problem (so they are all OK<br>
AFAIK).<br></blockquote></div><div><br></div><div>Lawyers that I have spoke=
n with are of the opinion that AI providers (either model weight providers =
or service providers) can not *really* legally claim any kind of copyright =
on the outputs of the AI models, so they are very explicitly NOT doing that=
. The outputs are quite clearly derived from your input so *if* any copyrig=
ht applies at all, it will be related to input and context (like the projec=
t being analysed to be modified). Barring major legal decisions on this fro=
nt, this is the (relatively undefined) status quo.</div><div><br></div><div=
>However, both AI model and AI service providers can and do force users to =
agree to terms of service to be able to use their models or services. Condi=
tions in those TOS documents varies widly and a few *can* be problematic fo=
r free software development, especially when sensitive information is conce=
rned, like pre-announcement security voulnerabilities. For example, Grok wa=
nts to have a full=C2=A0&quot;irrevocable, perpetual, transferable, sublice=
nsable, royalty-free, and worldwide&quot; license to all content generated =
by your queries ...</div><div><br></div><div><a href=3D"https://www.termsfe=
ed.com/blog/ai-terms-service-fine-print/">https://www.termsfeed.com/blog/ai=
-terms-service-fine-print/</a> is a pretty interesting write-up on some of =
these terms.</div><div><br></div><div>It is important to understand that TO=
S work very differently from copyright and copyleft licensing - TOS is *onl=
y* binding on the user directly using the service. It is *not* binding (and=
 can not be made binding) to any third parties that only get access to the =
*output* of the models and are not using the model itself. On the other han=
d TOS can and do contain clauses that allow the issuer of TOS to modify it =
at any time. This does noting to the outputs already generated, but may res=
trict future interactions of the user with the service or even with a downl=
oaded model. Usage of a downloaded model with a one-sided change clause *ca=
n* become in violation of its TOS later on.</div><div><br></div><div>So, IM=
HO - yes, if a proposal allowing AI use in Debian in any form does pass, it=
 will be useful to create and maintain a reference guide on which AI models=
 and services have &quot;Debian-compliant TOS&quot; and which don&#39;t an =
why/how specifically.</div><div><br></div><span class=3D"gmail_signature_pr=
efix">-- </span><br><div dir=3D"ltr" class=3D"gmail_signature"><div dir=3D"=
ltr">Best regards,<br>=C2=A0 =C2=A0 Aigars Mahinovs</div></div></div>

--00000000000028546906583e43e9--