https://www.marble.onl/posts/general_technology_doesnt_viola...
What? You mean I have to "express" this idea by typing it into a production system? So.. if I ask a production system, to "think" about any IPR character it likes. And construct a scenario, have I just committed IPR theft against all the IPR holders at one remove? How about if I chain the production systems and make one prompt the others, to prompt the others.. but at no time do I identify the IPR.
How can this be all on me? How not? Maybe.. the IPR is broken in the expression not the instruction. In which case deciding most of it is in the prompt not the production seems a bit odd.
Note, I do not think these llms and gpt systems have personhood, or own any IPR they create, independently of the entities which operate them.
Copyright on characters is strange to me. Length is not very useful with copied characters, and can be useful with copied plots. If I write a fanfic with an original plot but someone else's character then how much expression would I have copied? What kind of use of the character would be substantial or not substantial? Most of the expression in the fanfic would be my own, because I decide what the characters do. What characters do is part of their... characters too. Fair use can prohibit minimal copying or permit a large amount of copying depending on the situation [1]:
> That said, some courts have found use of an entire work to be fair under certain circumstances. And in other contexts, using even a small amount of a copyrighted work was determined not to be fair because the selection was an important part—or the “heart”—of the work.
Taken to a hypothetical extreme, this "heart" doctrine (originating from Campbell v. Acuff-Rose Music, Inc. (1994) [2]) would make pretty much all fan fiction using existing copyrighted characters fail the third fair use prong, "amount and substantiality" used of the copyrighted work [3]. On the other extreme, if the court finds no substantial similarity [4] between the copying work and the copyrighted work, then the fanfic does not infringe.
> And if one of you writes it down and lodges it online.. Anything beyond fair use length. There. Either, you, or me, of both of us just broke IPR.
Never mind publishing the fanfic, even writing it without sharing it might be infringement. Copyright law doesn't consider common sense. ASCAP tried to argue - in a lawsuit against cell service providers, rather than phone users - that phone users with ringtones set to copyrighted music were infringing on copyright [5]. ASCAP lost because of the definition of "public performance", not because setting ringtones to non-default music is an utterly normal activity nor because expressing personal preferences should be protected by the First Amendment.
So here's a fake fair use simulation. Fair use prong one is "purpose and character of the use". Transformative and non-commercial use are a strong but imperfect combination. Fair use prong two is "nature of the copyrighted work". Fanfic writers lose that one. Fair use prong three is "amount and substantiality". Heart of the work and character copyright get in the way; on the other hand, there might be no substantial similarity in the plots. Fair use prong four is "effect upon work's value". No meaningful effect if you never share the work; if you publish a fanfic then who knows. The prongs don't have established weights, so fair use is very case-by-case. If there's no substantial similarity then the fanfic writer might win. (I say win, but winning with the fair use defense would still leave most people in financial crisis from the attorney and court fees.)
> How about if I chain the production systems and make one prompt the others, to prompt the others.. but at no time do I identify the IPR.
> How can this be all on me? How not? Maybe.. the IPR is broken in the expression not the instruction. In which case deciding most of it is in the prompt not the production seems a bit odd.
You would still be the person who decides what to do with the output. As long as the ML model service provider isn't foolish enough to automatically store/publish the outputs, then the user can decide what to do with an output. If the user recognizes an output that contains a partial copy of an existing work then the user can avoid publishing/using the output.
The proposed "Training And Output" (TAO) doctrine in the article doesn't consider the case of when a user unfamiliar with X franchise doesn't recognize that the model outputs sometimes depict characters from X franchise. The service provider could get the bulk of the liability in such a case. Additionally, courts would still have to decide how big of a burden an ML model service provider has to prevent vague prompts from producing outputs that resemble copyrighted expression. Even many innocent, vague prompts can mention characteristics/features strongly associated with certain existing characters, works, and authors. For example, "spider superhero" is strongly associated with Marvel's Spider-man, so courts might well place a functionally impossible burden on ML model service providers to filter out possibly infringing results and thereby indirectly prohibit ML model creators from training on any visual Marvel works. On the other hand, the article emphasizes that ML models have significant uninfringing uses.
[1] https://www.copyright.gov/fair-use/
[2] https://en.wikipedia.org/wiki/Campbell_v._Acuff-Rose_Music%2....
[3] https://en.wikipedia.org/wiki/Fair_use
[4] https://en.wikipedia.org/wiki/Substantial_similarity#Substan...