Re: Ballot option: Allow AI-Assisted Contributions

--00000000000054fa5106584aec42
Content-Type: text/plain; charset="UTF-8"

On Wed, 5 Aug 2026 at 11:15, Gerardo Ballabio <[email protected]>
wrote:

> Aigars Mahinovs wrote:
> > 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.
>
> As I understand it, the problem with copyright isn't that AI providers
> might claim copyright. It's that *someone else* might claim copyright
> because the AI scraped and regurgitated their code. That's still an
> open legal question AFAIK.


If that were ruled to be true, then the *entire* AI landscape would
collapse as there is no way to gather enough data with *compatible*
licenses to produce a coherent AI model. Even if the pure-legal
interpretation were in favor of that outcome, there are strong social and
commercial incentives against the law moving that way.
https://www.wipo.int/en/web/frontier-technologies has big discussions about
that. Even largest pro-IP organizations are very careful to balance between
interests of existing copyright holders and users of AI systems.

So the current consensus seems to be to operate on the presumtion that
various legal loopholes (like fair-use in the USA and data mining exception
in EU) are sufficient to decouple the copyright of the model from the
copyrights of the training data (if the actual acquiring of the training
data is done without violating the copyrights). At this point there also
seems to be an opintion that the model weights and their pure output (as in
replies to trivial, non-copyrightable queries) are not *actually*
copyrigtable at all and thus stand in public domain. Not even with database
aggregation copyright. The reasoning behind that is not really clear to me,
but it seems that many legal departments have arrived at the same
conclusion - based on all the TOS for AI services and models being
formulated in a way *not* to invoke or rely on copyright protection.

As they sain Germany - clarity is something different than this.
-- 
Best regards,
    Aigars Mahinovs

--00000000000054fa5106584aec42
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div class=3D"gmail_quote gmail_quote_container"><div dir=
=3D"ltr" class=3D"gmail_attr">On Wed, 5 Aug 2026 at 11:15, Gerardo Ballabio=
 &lt;<a href=3D"mailto:[email protected]">[email protected]=
om</a>&gt; wrote:<br></div><blockquote class=3D"gmail_quote" style=3D"margi=
n:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex=
">Aigars Mahinovs wrote:<br>
&gt; Lawyers that I have spoken with are of the opinion that AI providers (=
either model weight providers or service providers) can not *really* legall=
y claim any kind of copyright on the outputs of the AI models, so they are =
very explicitly NOT doing that.<br>
<br>
As I understand it, the problem with copyright isn&#39;t that AI providers<=
br>
might claim copyright. It&#39;s that *someone else* might claim copyright<b=
r>
because the AI scraped and regurgitated their code. That&#39;s still an<br>
open legal question AFAIK.</blockquote><div><br></div><div>If that were rul=
ed to be true, then the *entire* AI landscape would collapse as there is no=
 way to gather enough data with *compatible* licenses to produce a coherent=
 AI model. Even if the pure-legal interpretation were in favor of that outc=
ome, there are strong social and commercial=C2=A0incentives against the law=
 moving that way.=C2=A0<a href=3D"https://www.wipo.int/en/web/frontier-tech=
nologies">https://www.wipo.int/en/web/frontier-technologies</a> has big dis=
cussions about that. Even largest pro-IP organizations are very careful to =
balance between interests of existing copyright holders and users of AI sys=
tems.</div><div><br></div><div>So the current consensus=C2=A0seems to be to=
 operate on the presumtion that various legal loopholes (like fair-use in t=
he USA and data mining exception in EU) are sufficient to decouple the copy=
right of the model from the copyrights of the training data (if the actual =
acquiring of the training data is done without violating the copyrights). A=
t this point there also seems to be an opintion that the model weights and =
their pure output (as in replies to trivial, non-copyrightable queries) are=
 not *actually* copyrigtable at all and thus stand in public domain. Not ev=
en with database aggregation copyright. The reasoning behind that is not re=
ally clear to me, but it seems that many legal departments have arrived at =
the same conclusion - based on all the TOS for AI services and models being=
 formulated in a way *not* to invoke or rely on copyright protection.</div>=
</div><div><br></div><div>As they sain Germany - clarity is something diffe=
rent than this.</div><span class=3D"gmail_signature_prefix">-- </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>

--00000000000054fa5106584aec42--