What’s going to be the name used for the laws that attempt to tackle machine paraphrasing?
or
"Battling Unlawful Language: Limit Scraping and Harness Initial Texts Act" or the "B.U.L.L.S.H.I.T Act".
That's super interesting and is news to me. Thanks for sharing. Would you mind linking to relevant statutes or court decisions?
(This isn't a "citation needed" post -- I believe you, and I'm genuinely curious to read more, but can't find anything!)
I too load creative works into all sorts of temporary structures in order to read the paper. I don’t need to license it, I pay for a subscription.
Should I pay more if I memorize the paper? Should I pay more if I read it to my sick friend in the hospital? Should I pay more if I save copies to my own hard drive and grep for words in the files? Should robots have a higher subscription price?
This whole “they copied it into a gpu” doesn’t matter. People read and interpret. Robots read and interpret. I don’t want to live in a world where every specific device and use needs to be licensed. Especially ex post facto. That will suck so hard.
Should I pay more if I memorize the paper? Should I pay more if I read it to my sick friend in the hospital?
I think the issue is that LLMs don’t make a copy or distribute a copy. They use the content to create something else. I don’t remember the copyright term for whether is is transformative enough. But it basically says I can’t copy Starry Night, but I can create a painting with the same color scheme and themes as long as it’s different enough from the original.
I think you already live in that world (though IANAL), there was a ruling that the Glider cheat tool for WoW was a copyright violation even though it was poking around inside the local copy necessarily made in RAM as part of normal usage of WoW.
https://arstechnica.com/gaming/2009/01/judges-ruling-that-wo...
here's their ToS, which is pretty clear about what you cannot do: https://help.nytimes.com/hc/en-us/articles/115014893428-Term... (relevant parts below)
Without NYT’s prior written consent, you shall not:
...
(2) use robots, spiders, scripts, service, software or any manual or automatic device, tool, or process designed to data mine or scrape the Content, data or information from the Services, or otherwise use, access, or collect the Content, data or information from the Services using automated means;
(3) use the Content for the development of any software program, including, but not limited to, training a machine learning or artificial intelligence (AI) system.
...
(5) cache or archive the Content (except for a public search engine’s use of spiders for creating search indices);
As an example, my local library offers full-text NYT articles through both nytimes.com and ProQuest.
Notably, ProQuest's terms only explicitly ban scraping metadata and developing software or services that "compete or interfere" with ProQuest products:
> Restrictions. Except as expressly permitted above, Customer and its Authorized Users shall not:
> Remove any copyright and other proprietary notices placed upon the Service or any materials retrieved from the Service by ProQuest or its licensors;
> Perform automated searches against ProQuest’s systems (except for non-burdensome federated search services), including automated “bots,” link checkers or other scripts;
> Provide access to or use of the Services by or for the benefit of any unauthorized school, library, organization, or user;
> Publish, broadcast, sell, use or provide access to the Service or any materials retrieved from the Service in any manner that will infringe the copyright or other proprietary rights of ProQuest or its licensors;
> Download all or parts of the Service in a systematic or regular manner or so as to create a collection of materials comprising all or a material subset of the Service, in any form.
> Store any information on the Service that violates applicable law or the rights of any third party.
Fair use would have to be blocked via a license, and it’s going to be difficult to argue that someone agrees to a license merely by turning on their radio. Responding to unauthenticated internet requests with content is the internet equivalent of broadcast and similarly the LinkedIn case held that this did not allow LinkedIn to impose terms of service in a contract of adhesion in this fashion.
If I watch Lebron James play and use it to develop an athletic training program, does it matter if I play baseball? Or WNBA? Or does it only matter if I play against him in the championship?
And I usually lean anti-corporate too, but banning people in the US from using data for LLMs might just mean they start being trained somewhere else that doesn't care as much about US law.
I say if people want to come up with bullshit lawsuits, we should play into them and force those people to suffer the consequences of said bullshit.
Computers and networks have been around a long time. These issues have been given a good workout.
“Fair Use” can apply to essentially any of the exclusive rights under copyright, including making (with or without distributing) a copy or derivative work.
> Probability models really just aren’t copyrightable to begin with.
“Probability models” that are built from copyrightable works either:
(1) have a sufficient human creative input to be copyrightable on its own (in which caee it still may infringe copyrights applicable to its source material as a derivative work), or
(2) do not have a sufficient human creative input to be copyrightable, and thus are a form of mechanical copy of their data from which they are developed (which, to the extent it either is, in aggregate, a copyrighted work, and/or contains copies of other copyrighted works, is protected by one or more copyrights, which do not cease to apply to the mechanical copy, and which an unlicensed mechanical copy would violate unless it fell into an exception like Fair Use.)
You're halfway there with #2. The output is not copyrightable, but unless you can actually point to a sequence of words from the original it can't be infringing.
They absolutely are copying in the course of training the model, and they are doing something that often looks a lot like copying when producing output with the model.
> Style (which can be described as a probability model) is not copyrightable.
Style is not all that can be described in an LLM's "probability model", otherwise LLM models would never be able to reproduce content from the training set, which they do.
> The output is not copyrightable, but unless you can actually point to a sequence of words from the original it can’t be infringing.
While verbatim copying (“a sequence of words from the original”) is infringing, mechanical copying with a mechanical substitution filter (even a wickedly complex one) is also copying, and violates copyright.
And ChatGPT is quite capable of producing verbatim text from its training set.
Only in the sense that a Cisco router is copying in the course of sending me the article, which we've all agreed doesn't count as infringement.
The bigger problem is that the plaintiff has to show it is more likely than not that that sequence of words came from their text and not some other source, which is going to be obscenely difficult.
> And ChatGPT is quite capable of producing verbatim text from its training set.
Granted, and then the copyright holder could sue for infringement at that point. Exactly whom he should sue is a more difficult question.
Look, I get it: you want copyright to allow authors to say "you can't do that with my work"; I'm sympathetic, but it just doesn't give authors that power.
But it does count as infringement, if its not explicltly or implicitly licensed (because it is necessary to a use that is licensed or necessary to a use that does not itself require a license but is the normal use for for which a licensed copy, which you have, is sold and used) and it is not itself Fair Use (usually, because it is necessary in the course of a use which is itself Fair Use.)
> The bigger problem is that the plaintiff has to show it is more likely than not that that sequence of words came from their text and not some other source, which is going to be obscenely difficult.
The plaintiff has to (1) show facts from which a judge concludes that a reasonable jury might conclude that, and (2) get the jury to conclude that.
This can be difficult, and it might not be trivial in this case, but in practice the combination of opportunity and non-trivial similarity tends to put more weight on the defense to show a convincing alternative explanation. Outside of conclusions that a judge views as conpletely unreasonable based on the facts, the civil burden of proof boils down to what a jury feels is more likely, not some.
> Granted, and then the copyright holder could sue for infringement at that point. Exactly whom he should sue is a more difficult question.
Well, who else might be liable depends on the specific circumstances, but with OpenAI controlling the whole course between the first copy of the copyrighted work and the infringing end product, and doing it all for commercial gain, that OpenAI wpuld be on the hook, civilly and potentially criminally for any infringement, is clear.
> Look, I get it: you want copyright to allow authors to say "you can't do that with my work"; I'm sympathetic, but it just doesn't give authors that power.
I want copyright to be both more limited in its exclusive rights and either shorter or costlier to the copyright holder than it is. But what I want is not what the law is.
I don't think you've thought about this hard enough
The transient copy isn’t for a licensed use, and whether it is a Fair Use is specifically the subject of debate. So this really is basically an admission that but for the potential applicability of Fair Use, the use of the copyrighted mmaterial in training is a violation of copyright.
Also, even if the training is fair use, that doesn’t mean that the copies of the source material produced by OpenAI and distributed to their customers using the model are fair use. Just becaue making the tool is a transformative fair use doesn’t mean using the tool to generate copies of the material which was used to train it, which are significantly less trandormative than the model itself, are Fair Use. (And the fact that one of the functions that the model is used for is this commercial, for profit by the maker of the model, copying of the source material is – as much as I believe AI model training on its own is quite likely to generally be fair use – an argument against the model training being fair use in this csase.)
> Computers and networks have been around a long time.
True, and commercially producing and delivery copies of copyrighted works in a manner which substitutes for the original work in the marketplace, no matter what intermediate steps go into doing that, and no matter that computers or networks are used in those intermediate steps, is pretty much the clearest case of violation of copyright you can get.
> These issues have been given a good workout.
Some of them have, some of them have not. Whether and in what conditions training an AI on source material that may be subject in aggregate to a compilation copyright by someone else, and which consists further of individual works that have their own copyrights, might be “fair use” is not one of the issues that have been given a good workout. Neither – because producing predictive models in that way has not previously been common – has whether, unlike other intermediate tool use, using such a model in the course of doing what would otherwise be an infringement by producing a copy of specific copyright-protected works, commercially, for a customer at their request, is no longer a violation because the use of the model somehow isolates it from liability,
This issue is ultimately going to come down to the transformative clause of fair use. The fact is that the _model_ is unquestionably a transformative product of the inputs, and a judge ruling otherwise is going to cause a cascading shitstorm of litigation and put a chill through the creative economy. The outputs of the model under certain conditions can be guided towards copyright infringement, and any sane ruling will focus on protecting rightsholders from overly derivative model outputs. In all likelihood the precedent will be that the standard for being transformative will be raised for "algorithmically generated" content, and the people who distribute that content will still be fully liable in the event of infringement, with "I didn't know, the AI did it" not being an acceptable defense.
If I read five calculus textbooks and write a new one, I don't think that's derivative content (or maybe it is?) Seems like that's what an LLM does - read many works, write a new work.
Fan art and fan fiction are derivative without copying sequences of words
How would they prove this? Is it safe to say that each article used has a nearly meaningless influence on the weights?
Could this be used as a defense? Perhaps train a (smaller) model, remove a single article, and show how it doesn't influence performance?
It’s not like it’s ok to violate copyright if you don’t compete. It’s still illegal to take a NYT article and print it on a t-shirt.
The issue is that copyright law doesn’t prevent the kind of model training as there’s no clearly derived work. I don’t think that’s been tested in courts yet, but I expect it won’t be found to be copyright because there’s other precedent that influenced is not infringement.
Whether that copy was fair use is the key question.
But as you point out in the router case above, it's transient.
I bought a license to listen and use.
Anyway, I've always been a bit prickly about IP stuff ever since a troll lawyer threatened me over a TI-BASIC game when I was like 14. I'm also sure I'm completely wrong-headed about this whole thing and overly anthropomorphizing the LLM.
The tool building of model training is more likely to be fair use than the use of the tool to provide mechanical copies of copyirght-protected material that competes directly with the original in the market.
Congress took the ruling seriously enough to carve out a fair use exception for that use. I don’t think there’s any comparable exception here, especially when the ultimate purpose is to build a commercial product capable of producing works (news reports) in precisely the same market as the original works.
Assuring
Responsible
Behavior in
AI
Generated
Expressions
G.A.R.B.A.G.E.