Yes, from the linked paper:

"Books3 is a dataset of books derived from a copy of the contents of the Bibliotik private tracker made available by Shawn Presser (Presser, 2020). Bibliotik consists of a mix of fiction and nonfiction books and is almost an order of magnitude larger than our next largest book dataset (BookCorpus2). We included Bibliotik because books are invaluable for long-range context modeling research and coherent storytelling"

This is the most ridiculous legal hand wave I’ve ever seen.

“They’re not books, man, they’re a dataset!”

I believe it is. See https://www.wired.com/story/battle-over-books3/

Do they claim that none of their data came from copyrighted sources / is copyrighted?

The claim (which I don't personally agree with, but I'm trying to represent here in good faith) is that although the data is copyright, training models constitutes "fair use" under US copyright law and therefore you're entitled to use copyright material for this.

Fair to say that whether or not this is correct is pretty important to all the outstanding court cases on this matter.

That seems to fall apart quickly. Even if training could be considered fair use, surely just distributing the raw masses of copyrighted works can't be under any reasonable definition. Otherwise, why did TBP, KAT, and MegaUpload shut down if you could defeat copyright with sheer numbers?

Indeed. Also in the US, whether or not something is fair use involves a four factor test[1] and two of the factors are the amount and substantiality of what's taken and the effect on any market. In this case, the amount is "everything" and the effect on the market is potentially very large for authors/publishers.

[1] https://fairuse.stanford.edu/overview/fair-use/four-factors/

>two of the factors are the amount and substantiality of what's taken and the effect on any market

books.google.com has been allowed to copy all the books they can lay their hands on, so long as they don't regurgitate them in full, so it's not really the taking, but any subsequent reproductions. And the effect on the market is insubstantial if the alternative wasn't going to be the equivalent sales.

You can download the whole dataset, so they're certainly able to regurgitate them in full.

One thing that we did with distributing certain copyright-protected textual material was to scramble them at the paragraph level.

If you take every paragraph in the Harry Potter saga and sort the paragraphs in alphabetical order, it's just as good for training short-context-window models, but not a "harm to the market" leading to a lost sale for anyone who wants to read the books.

It absolutely could be a harm to the market if people use the resulting model to generate "Harry Potter" books instead of buying the real ones.

The resulting model doesn't have access to the information about what follows what, so it can recreate paragraphs but can't recreate their proper order for a chapter or book. Well, it can try to guess..

Totally get that but the law doesn't care about that as I understand it. For the four-fold test it matters whether the use is going to affect the market for the original work (not the technicalities of how the model works). If people generate pseudo-Harry Potter via a model that was trained on Harry Potter then the court may well decide that the market for real Harry Potter is affected. That doesn't seem an unreasonable conclusion to me.

I'm pretty sure that's what the lawyers will argue in the Silverman case for example. It's going to be interesting to see how the courts decide.

Megaupload et all went against the entertainment industry in a time when that industry had the money to pay the lawyers to convince the judges what the law means.

In the present moment on the other hand, it is the entities in the AI industry (e.g. MS) that have the money and can hire the lawyers to convince the judges. Realistically speaking, it's very likely that things will swing the way of AI companies, which will benefit, albeit indirectly, these guys, even though by themselves they're too small to push their agenda, they're just bit players.

Since when has TBP shut down?

I think they are referring to the many times the domain name has been seized, and shut down temporarily.

https://www.youtube.com/watch?v=eTOKXCEwo_8 for those that haven't seen it.

Some of the founders were convicted of crimes but the database and code are out there.

I think there is actually a good argument that an AI model is transformative, and that training a model is therefore not infringing of the copyright. (An analogy: if you rolled dice to select words randomly from the Lord of the Rings and rearranged them into a poem, it's not infringing the Lord of the Rings even if in a sense, every word was taken from that book.)

But you still have to get your hands on the copyrighted data legally. It might be legal to scan every book an institution owns, and train off it, so long as those scans are not distributed. But it is probably not legal to scrape copyrighted content off torrents - creating the copy to train with is infringing, even if the model's final product maybe isn't.

Yes agreed, and transformative use itself also has limitations. You don't have carte blanche to use something just because you think it's transformative, for example the Lynn Goldsmith vs Andy Warhol Foundation case over the "Orange Prince" work. https://copyrightalliance.org/warhol-decision-reins-transfor...

while there is a good argument that AI produces transformative outputs, it's refuted when the models are shown to regurgitate literal text, which they have. Then it just starts to look like a neural memorization agent, compressed storage algorithm, etc.

Definitely open to the idea, that couldn’t be the whole argument. I mean, my brain can output some quotes, but I’m not a compressed storage algorithm. Or at least I hope I’m not.

I've seen examples of this, but they're nearly always isolated, rather difficult to obtain, and not in fact exact copies. You need to specifically ask for an exact copy, and then attempt to defeat the safeguards the model has in place to prevent this, and hope that it was "memorized" - which for the record is considered to be a flaw in the model as it's a reduction in information density and capability, compared to if that "memory" was used for something else. Good models seek to reduce this as much as possible. With the size of the datasets involve (see OP) this feels more like an understandable and reasonable issue to have.

This very rarely happens, usually when trying hard to get it to regurgitate, and I don't think it has ever happened for anything longer than 2 paragraphs, or at most a short article. Certainly not something like a book or even the whole issue of a newspaper.

Yup, exactly. Passages from GoT appear because they're frequently referenced across the Internet, including the multiple wiki style sites for GoT fans, not necessarily because the LLM is regurgitating whole-book content it was trained on.

"Open source" implies that, no? A definition of open source which includes blatantly pirated material on the condition that the people who collated and released the pirated material did so for free is really stretching it past breaking point. By that standard everything on The Pirate Bay is open source.

Why do everyone assume "open source" imply legality?

(/s)

humans have the innate hunter-gatherer sense that generosity and sharing are good for us, and selfishness is bad. And that ethical should be legal and unethical should not be.

Open source is generous sharing, and ethical. Start to nibble away at those ideals and at what point do you slip into unethical? IMDB was a crowd-sourced database put together by a wide community pitching in small efforts which one guy was maintaining like an FAQ. Then the guy maintaining it said, "It's worth money, I own it, screw all of you." How would people react if this happened to wikipedia? But wikipedia is safe because it's a non-profit... you know, like OpenAI, right?

There's odd stuff in there. I just randomly downloaded a file,

https://the-eye.eu/public/Books/ThoseBooks/Puzzles.tar -- 20-Jan-2023 14:54 -- 6M

and it pretends to be a jigsaw puzzle, but is actually eISBN 9781594868573 - The South Beach diet cookbook / Arthur Agatston

this list [0] seems like a starting place to look into various legal actions.. not sure how often it is updated e.g. Silverman et al

[0] https://originality.ai/blog/openai-chatgpt-lawsuit-list

Pouring one out for the future litigators, jurors, and judges who will have to pore over this inextricable web of legal and technical details

They'll just let their ai do it over lunch.