This is really not so clear cut as "fair use" might cover 99% of all data scrapping; you are not reproducing the originals just use them to estimate probabilistic distribution of tokens in pre-training. You are never going to get the exact book word-for-word using LLMs.
>You are never going to get the exact book word-for-word using LLM.
This is pretty much the exact claim of a NYT lawsuit against OpenAI.
"One example: Bing Chat copied all but two of the first 396 words of its 2023 article “The Secrets Hamas knew about Israel’s Military.” An exhibit showed 100 other situations in which OpenAI’s GPT was trained on and memorized articles from The Times, with word-for-word copying in red and differences in black."
https://www.hollywoodreporter.com/business/business-news/cou...
Yes, LLMs fundamentally operate as a lossy compression scheme for their training data. There's been countless examples of them reproducing their training data with very high accuracy
People claim that the data isn't stored, but clearly a representation of it is encoded and reproducible. I saw chatgpt word for word plagiarise a stack overflow comment just two days ago
Does this actually imply a representation of it has been stored or simply that the model is sort of over-fit?
https://arxiv.org/html/2510.25941v1
You can get it to reproduce content but it’s a game of cat and mouse. Were it not for the alignment to avoid direct reproduction it would taken far more often.
> RECAP consistently outperforms all other methods; as an illustration, it extracted ≈3,000 passages from the first "Harry Potter" book with Claude-3.7, compared to the 75 passages identified by the best baseline.
I don’t buy this argument. The tokens are useless without their context, which provides the probability distributions needed to make them useful. Sure you MIGHT not be able to get the book word for word, but it’s impossible to make a useful model without the whole book and all of the artistry that went into it, to guide the tokens in their expected output.
Fair use generally does not cover commercial use, which this clearly is, and is dependent on the amount of the original content present in the derived work, which I would contend in this case is “all of it”
"Commercial Use" is only one part of the four prongs of the fair use test. For example, commercial Parody is generally considered Fair Use. Look at Space Balls, which is a direct transformation from Star Wars.
This is all new territory. We don't have court-settled law yet.
It's more complicated than that. Quite a bit more.
Commercial use counts _against_ a fair use defense, but is not dispositive: it's not accurate at all to say it "generally does not cover" commercial use. This is the "purpose and character" test, one of four in contemporary (United States) fair use doctrine.
Purpose and character also includes the degree to which a use is _transformative_. It's clear that the degree to which a training run mulching texts "transforms" them is very high. This counts toward a fair use finding for purpose and character.
> is dependent on the amount of the original content present in the derived work, which I would contend in this case is “all of it”
The "amount and substantiality" test. Your case for "all of it" can't possibly be sustained: the models aren't big enough. It's amount _and_ substantiality: this has come up in the publication of concordances, where a relatively large amount of a copyrighted work appears, but it's chopped up and ordered in a way which is no longer substantially the same. Courts have ruled that this kind of text is fair use, pretty consistently. It's not an LLM, of course, but those have yet to be ruled on.
Also worth knowing that courts have never accepted reading or studying a work as incorporation, and are unlikely to change course on the question. It's taken for granted that anyone is allowed to read a copyrighted work in as much detail as they wish, in the course of producing another one. Model training isn't reading either, but the question is to what degree it resembles study. I'd say, more than not.
Specifically:
> it’s impossible to make a useful model without the whole book and all of the artistry that went into it
Courts have never once accepted "it would be impossible for defendant to write his biography without reading plaintiff's" as valid, and it's been tried. The standard for plagiarism is higher than that.
"Effect upon the work's value" is probably the most interesting one. For some things, extreme, for others, negligible. I suspect this is the one courts are going to spend the most time on as all of these questions are litigated.
Ultimately, model training is highly out-of-distribution for the common law questions involving fair use. It was not anticipated by statute, to put it mildly. The best solution to that kind of dilemma is more statute, and we'll probably see that, but, I don't think you'll be happy with the result, given what I'm replying to. Just a guess on my part.
It is of course true that it is unsettled law, and that fair use is more complicated than my offhand comment suggested.
> Courts have never once accepted "it would be impossible for defendant to write his biography without reading plaintiff's" as valid, and it's been tried. The standard for plagiarism is higher than that.
This I think misses the thrust of my argument, though. Its hard to find an exact human analogy, because neither the technology nor the scale at which it operates is remotely human.
I see it less as “writing his biography without reading the plaintiff’s” and it’s more “using the same style and metaphors to make thousands of copies of very similar biographies, with certain bits tweaked,” like turning an existing work into mad lib.
I don’t know how the courts will eventually rule on it, but it certainly feels like theft to me.
It's fascinating how intuitions differ. To me, it doesn't feel like theft at all. For one thing, theft is depriving another of something, and has therefore never been a good metaphor for infringement; hackers used to be the most insistent about this principle, and it's weird to see a doctrine which was cooked up in a literal AI lab get thrown out the window for literal AI.
But pretending you said "infringement", for me it comes all the way back to the Constitution: "To promote the Progress of Science and useful Arts". I cannot possibly twist the development of large language models into something which violates the spirit of that purpose. I don't see how anyone can.
Your point about the scale is valid, and the alienness of it, sure. But you haven't made the case that the vastness of the scale should affect the conclusion.
Something I left out in the first post is that copyright is meant to protect expression, and not ideas: this is the deciding factor in the 'nature of the copyrighted work' test for fair use. More expression, more protection: more ideas, less.
I think the visual arts have a strong case that image generators directly infringe expression: I'm not convinced that authors do, and I think software should never have been protected under copyright because the ideas-to-expression ratio is all wrong for the legal structure. There's clearly no scale case to be made for ideas: "but what if it's _all_ the ideas" fails, because the ideas are not protected at all. Nor should they be, that's what patents are for, and why patents are very different from copyright.
LLMs are remarkably good at 'the facts of the matter', hallucination not withstanding. They're very poor at authorial 'voice transfer', something image generators are far too good at. It's when I start asking myself "well what even _is_ this 'expression' thing anyway?" that I conclude that we're out over our skis on the LLMs-and-IP question: precedent can't tell us enough, and that leaves legislation.
Try prompting Claude to create a drop-in replacement for an existing library, testing against that library's test suite to validate functionality.
It will pretty much plagiarize the library verbatim from memory, sans comments.
When I was in school, writing "in my own words" was never an excuse to not cite a source. It was actually something that took me a little while to understand, it's the source of the information that needs to be cited, and that's not limited to literal quotations of someone else's writing.
That's more an argument for why you can't just use LLMs as a source of truth. Conveniently, LLMs like ChatGPT do often cite their sources, especially if you prompt them to.
Maybe a nit: LLMs do not and cannot cite their sources (at least scraped sources for the purpose of training)
It’s kind of the harness that is doing the citing (or providing the context for the model to).
But an LLM sans search can reproduce some copyrighted work with minor variations and there’s no way to know exactly where it came from.
> You are never going to get the exact book word-for-word using LLMs
You could say the same about MP3 encoders but I don't think that would convince any judge
Come up with obscure topic that has few relevant results, post about to Reddit on your profile page, wait a few hours and then query Gemini/ChatGPT about that exact thing and tell me you still feel this way.
This confuses input and output.
A copy made for the purposes of training is still a copy.
Even if you throw the text away after training, you've still made a copy.
Fair use was built around human limitations. The mass scraping campaigns done by the AI giants were clearly an overreach in spirit, if not letter. Most people's intuition is that these massive operations that are valued in the trillions can't have been drawn from some untapped common resource, and they're correct. Someone, somewhere is not being properly compensated.
I have no problem with taxing AI companies so that their profit is marginal, or forcing them to provide compute for free. That seems like the correct balance of what they're harvesting from the "commons" (which is really just the totality of private IP that was exposed to their crawlers).