On Thu, 6 Aug 2026 at 16:03, Simon Richter <[email protected]> wrote:
> On 8/5/26 5:59 PM, Gerardo Ballabio wrote:
>
> > 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.
>
> That, and whether the person prompting AI can claim copyright on the
> generated code.
>
> If we accept the premise that the code that went into the training data
> has been so thoroughly processed that no traces of the original
> copyrights remain and the generated code is therefore not subject to the
> intersection of the original licenses, that still leaves the context
> going into the inference phase.
>
> That context is a mix of
>
> - system prompt (provided by the service)
> - project context (when modifying existing code)
> - user prompt
>
> Now, there are multiple possible interpretations:
>
> 1. the output is not a derivative work of the context. The prompter
> therefore is not the author, and asserting that they own the copyright
> is therefore fraudulent.
>
> 2. the output is a derivative work of the context. We therefore need a
> licence from the service provider to be able to freely choose a licence
> for the generated output (I suppose we have that, but we need to make
> sure), and the generated code is largely attributable to the authors of
> the existing code, not the person implementing the latest change.
>
> I don't see any interpretation where a thorough review is sufficient to
> establish authorship, regardless of whether the generated code is
> copyrightable or not.
It is an interesting debate and it would be nice if there was a definitive
answer at some point.
IMHO it is pretty clear that AI output is a derrived from from the direct
user input (prompt) and the direct context (the project being modified)
because of the close causal relationship between those objects. The user
controls those inputs and the related copyrights and licenses in open
source development context are pretty clearly defined. The more complex and
novel the output the more complex discussion between user and the AI
eventually led to that output, so that is where the user contributed their
creative input. So the output should have compatible legal status to the
project. Whether output becomes copyrightable directly or the user takes a
non-copyrightable (de facto public domain) code and modifies it to apply
enough copyright to release under a project-compatible license is IMHO just
a question of legal formalities.
System prompt and its is usage and non-claiming of copyright is covered by
the TOS of the service or model. The service provider can not allow or
forbid the user to license the output if they are not claiming any
copyright on the output - having a license depends on having the copyright
law to apply in the first place. If the copyright law does not apply to the
output at all, then it has no license conditions and can be used freely as
public domain code.
Can public domain code be included into, for example, a GPL projects? AFAIK
- yes.
--
Best regards,
Aigars Mahinovs
