I don’t think the "who wrote it" is the right lens for this. I suspect
that the big issue here is the likely huge increase in uncopyrightable
text in our documents. Sure we already have public domain/uncopyrighted
text, but from my limited viewpoint it seems minor and rare, whereas we
can foresee a future where it is major and commonplace. When that
happens, I don’t think our rules/processes will be adequate and an
evolution will be needed to address that. To give you a concrete
example, we prevent non-IETF derivatives of IETF standards for good
reasons, but if/when a standard has significant AI output incorporated
into it, even if that’s light editing, will we be able to continue with
that level of protection?
Having been looking at many of the AI lawsuits flying by, I can give you the
definitive answer: nobody has any idea.
It is true that the US Copyright Office has said that LLM output is not
copyrightable because authors have to be human. But there are lawsuits that
argue that at least some LLM output is a derivative work of the training
material. (If you look at the examples in the NY Times' complaint against
OpenAI, they have a point.) George R.R. Martin has sued claiming that LLM plot
summaries of his books are infringing, but the Authors' Alliance points out that
they look a lot like the summaries in Wikipedia which we hope are OK. So LLM
output isn't copyrightable except maybe sometimes it already was. None of these
cases are close to resolution, so it will be a long time until there are useful
precedents.
The question of how much human effort turns a PD work into a copyrightable
one is also a swamp. In the US it depends on whether there was some
creativity involved (the definitive case was about telephone white pages,
which didn't), while other countries also look to "sweat of the brow", the
amount of work involved. The creative effort can be quite small -- in a
2017 case the US court held that the decorations on generic looking
cheerleader uniforms were creative enough. So it may well be that the
amount of editing that a WG and the RPC do is enough to cross the
threshold, but unless there is a court case with similar facts, we're just
guessing. I wouldn't hold my breath.
So anyway, while I think it is a good idea to keep abreast of all this, it is
not a good idea to design processes based on our guesses about what the legal
treatment of LLM output will be because whatever our guess is, it will be wrong.
R's,
John
PS: you can see the cheerleader uniforms in an illustration in the middle of
this Wikipedia article:
https://en.wikipedia.org/wiki/Star_Athletica,_LLC_v._Varsity_Brands,_Inc.
--
rswg mailing list -- [email protected]
To unsubscribe send an email to [email protected]
