A federal judge sides with Anthropic in lawsuit over training AI on books

So it totally isn't a warez streaming media server but AI?

I'm guessing since my net worth isn't a billion plus, the answer is no

Purposes which are fair use are very often not at all personal.

(Also, "personal use" that involves copying, creating a derivative work, or using any of the other exclusive rights of a copyright holder without a license or falling into either fair use or another explicit copyright exception are not, generally, allowed, they are just hard to detect and unlikely to be worth the copyright holder's time to litigate even if they somehow were detected.)

This will have the effect of empowering countries (and other entities) that don't respect copyright law, of course.

The copyright cartel cannot be allowed to yank the handbrake on AI. If they insist on a fight, they must lose.

If you can successfully demonstrate that then yes it is a copyright infringement and successfully doing that would be worthy of NeurIPS or ACL paper.

I'm not so sure about this one. In particular, presuming that it is found that models which can produce infringing material are themselves infringing material, the ability to distill models from older models seems to suggest that the older models can actually produce the new, infringing model. It seems like that should mean that all output from the older model is infringing because any and all of it can be used to make infringing material (the new model, distilled from the old).

I don't think it's really tenable for courts to treat any model as though it is, in itself, copyright-infringing material without treating every generative model like that and, thus, killing the GPT/diffusion generation business (that could happen but it seems very unlikely). They will probably stick to being critical of what people generate with them and/or how they distribute what they generate.

The amount of the source material encoded does not, alone, determine if it is infringing, so this noun phrase doesn't actually mean anything. I know there are some popular myths that contradict this (the commonly-believed "30-second rule" for music, for instance), but they are just that, myths.

This is still a weird language shift that actively promotes misunderstandings.

The weights are the LLM. When you say "model", that means the weights.

This is a bit distorted. This is a better summary: The primary purpose of copyright is to induce and reward authors to create new works and to make those works available to the public to enjoy.

The ultimate purpose is to foster the creation of new works that the public can read and written culture can thrive. The means to achieve this is by ensuring that the authors of said works can get financial incentives for writing.

The two are not in opposition but it's good to be clear about it. The main beneficiary is intended to be the public, not the writers' guild.

Therefore when some new factor enters the picture such as LLMs, we have to step back and see how the intent to benefit the reading public can be pursued in the new situation. It certainly has to take into account who and how will produce new written works, but it is not the main target, but can be an instrumental subgoal.

LLMs are models of languages, which are models of reality. If anyone deserves compensation, it's humanity as a whole, for example by nationalizing, or whatever the global equivalent is, LLMs.

Approximately none of the value of LLMs, for any user, is in recreating the text written by an author. Authors have only ever been entitled to (limited) ownership their expression, copyright has never given them ownership of facts.

If you train a LLM on harry potter and ask it to generate a story that isn't harry potter then it's not a replacement.

However, if you train a model on stock imagery and use it to generate stock imagery then I think you'll run into an issue from the Warhol case.

I think that's the conclusion of the judge. If Anthropic were to buy the books and train on them, without extra permission from the authors, it would be fair use, much like if you were to be inspired by it (though in that case, it may not even count as a derivative work at all, if the relationship is sufficiently loose). But that doesn't mean they are free to pirate it either, so they are likely to be liable for that (exactly how that interpretation works with copyright law I'm not entirely sure: I know in some places that downloading stuff is less of a problem than distributing it to others because the latter is the main thing that copyright is concerned with. And AFAIK most companies doing large model training are maintaining that fair use also extends to them gathering the data in the first place).

(Fair use isn't just for discussion. It covers a broad range of potential use cases, and they're not enumerated precisely in copyright law AFAIK, there's a complicated range of case law that forms the guidelines for it)

That's not what the ruling says.

It says that training a generative AI system not designed primarily as a direct replacement for a work on one or more works is fair use, and that print-to-digital destructive scanning for storage and searchability is fair use.

These are both independent of whether one person or a giant company or something in between is doing it, and independent of the number of works involved (there's maybe a weak practical relationship to the number of works involved, since a gen AI tool that is trained on exactly one work is probably somewhat less likely to have a real use beyond a replacement for that work.)

Worse, they’re using it for massive commercial gain, without paying a dime upstream to the supply chain that made it possible. If there is any purpose of copyright at all, it’s to prevent making money from someone’s else’s intellectual work. The entire thing is based on economic pragmatism, because just copying does obviously not deprive the creator of the work itself, so the only justification in the first place is to protect those who seek to sell immaterial goods, by allowing them to decide how it can be used.

Coming to the conclusion that you can ”fair use” yourself out of paying for the most critical part of your supply makes me upset for the victims of the biggest heist of the century. But in the long term it can have devastating chilling effects, where information silos will become the norm, and various forms of DRM will be even more draconian.

Plus, fair use bypasses any licensing, no? Meaning even if today you clearly specify in the license that your work cannot be used in training commercial AI, it isn’t legally enforceable?

I'm allowed to hear a copyrighted tune, and even whistle it later for my own enjoyment, but I can't perform it for others without license.

> This order doubts that any accused infringer could ever meet its burden of explaining why downloading source copies from pirate sites that it could have purchased or otherwise accessed lawfully was itself reasonably necessary to any subsequent fair use. There is no decision holding or requiring that pirating a book that could have been bought at a bookstore was reasonably necessary to writing a book review, conducting research on facts in the book, or creating an LLM. Such piracy of otherwise available copies is inherently, irredeemably infringing even if the pirated copies are immediately used for the transformative use and immediately discarded.

(But the judge continued that "this order need not decide this case on that rule": instead he made a more targeted ruling that Anthropic's specific conduct with respect to pirated copies wasn't fair use.)

If I buy a book, and as long as the product the book teaches me to build isnt a competing book, the original author should have no avenue for complaint.

People are really getting hung up on the computer reading the data and computing other data with it. It shouldnt even need to get to fair use. Its so obviously none of the authors business well before fair use.

I see elements of that here. Buying copyrighted works not to be exposed and be inspired, nor to utilize the aithor's talents, but to fuel a commercialization of sound-a-likes.

How many copies? They're not serving a single client.

Libraries need to have multiple e-book licenses, after all.

https://en.wikipedia.org/wiki/Mickey_Mouse#Walt_Disney_Produ...

I'm on the Air Pirates side for the case linked, by the way.

However, AI is not a parody. It's not adding to the cultural expression like a parody would.

Let's forget all the law stuff and these silly hypotheticals. Let's think of humanity instead:

Is AI contributing to education and/or culture _right now_, or is it trying to make money? I think they're trying to make money.

Human time is inherently valuable, computer time is not.

The issue with LLMs is that they allow doing things at a massive scale which would previously be prohibitively time consuming. (You could argue but them how much electricity is worth one human life?)

If I "write" a book by taking another and replacing every word with a synonym, that's obviously plagiarism and obviously copyright infringement. How about also changing the word order? How about rewording individual paragraphs while keeping the general structure? It's all still derivative work but as you make it less detectable, the time and effort required is growing to become uneconomical. An LLM can do it cheaply. It can mix and match parts of many works but it's all still a derivative of those works combined. After all, if it wasn't, it would produce equally good output with a tiny fraction of the training data.

The outcome is that a small group of people (those making LLMs and selling access to their output) get to make huge amounts of money off of the work of a group that is several orders of magnitude larger (essentially everyone who has written something on the internet) without compensating the larger group.

That is fundamentally exploitative, whether the current laws accounted for that situation or not.

https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....

Maybe there's another big Google Books lawsuit that Google ultimately lost, but I don't know which one you mean in that case.

Meta at least just downloaded ENGLISH_LANGUAGUE_BOOKS_ALL_MEGATORRENT.torrent and trained on that.

This ruling doesn't say anything about the enforceability of a "don't train AI on this" contract, so even if the logic of this ruling became binding prcecednet (trial court rulings aren't), such clauses would be as valid after as they are today. But contracts only affect people who are parties to the contract.

Also, the damages calculations for breach of contract are different than for copyright infringement; infringement allows actual damages and infringer's profits (or statutory damages, if greater than the provable amount of the others), but breach of contract would usually be limited to actual damages ("disgorgement" is possible, but unlike with infringer's profits in copyright, requires showing special circumstances.)

First, I dont think publishers of physical books in the US get the right to establish a contract. The book can be resold for instance and that right cannot be diminished. But secondly adding more cruft to the distribution of something that the end user has a right to transform, isn't going to diminish that right.

Humans, animals, hardware and software are treated differently by law because they have different constraints and capabilities.

And who gets the money? Not the original author.

LLMs may sometimes reproduce exact copies of chunks of text, but I would say it also matters that this is an irrelevant use case that is not the main value proposition that drives LLM company revenues, it's not the use case that's marketed and it's not the use case that people in real life use it for.

But for non-fiction and technical books, they sure as hell will. For example, if I ask Gemini "What does Bob Nystrom say about the Singleton pattern in Game Programming Patterns?", it says:

---

In "Game Programming Patterns," Bob Nystrom dedicates an entire chapter to the Singleton pattern, which is unusual because, as he states, "Every other chapter in this book shows you how to use a design pattern. This chapter shows you how not to use one."

Here's a summary of his key points regarding the Singleton pattern:

* More Harm Than Good: Nystrom argues that despite its noble intentions, the Singleton pattern often does more harm than good, especially in game development. He notes that the Gang of Four (GoF) stressed its sparing use, a message that was often lost in translation to the game industry.

* Global State in Disguise: He highlights that the Singleton is essentially a global variable in disguise. While it ensures a single instance, it often leads to problems associated with global state, such as tightly coupled code, difficulty in testing, and reduced flexibility.

* Encourages Tight Coupling: By providing a global access point, Singletons encourage various parts of the codebase to directly depend on them. This makes the code harder to change and refactor because modifications to the Singleton can have widespread, unpredictable effects.

* Hinders Testability: Testing code that relies on Singletons can be challenging. Because Singletons are globally accessible and often difficult to mock or replace, it can be hard to isolate units of code for testing, leading to brittle tests or a lack of testing altogether.

* Limits Flexibility: Nystrom points out that while a class might currently only need one instance (e.g., a FileSystem or AudioSystem), hardcoding this limitation with a Singleton makes it difficult to change in the future. If a game later needs multiple file systems (e.g., for different save slots) or different audio contexts, a Singleton makes this a non-trivial refactoring.

* Alternatives Often Better: He suggests that often, what developers think they need a Singleton for can be better achieved with other patterns or simpler solutions. For instance, if the only goal is to have a single instance, a simple static class might suffice if global access is truly desired and initialization order isn't a complex concern. For managing services, he often advocates for the Service Locator pattern as a less problematic alternative that still provides global access without enforcing a single instance.

* Misconceptions about "Singular Resources": While the GoF suggested Singletons for "physical devices," Nystrom argues that few physical devices are truly singular in a way that necessitates a Singleton. For example, while you might only have one keyboard, a game could theoretically support multiple input devices. Using a Singleton prematurely prevents this flexibility.

* Lazy Initialization: He acknowledges that a well-implemented Singleton using lazy initialization (creating the instance only when it's first accessed) can save memory and CPU cycles and resolve some issues with static initialization order. However, he emphasizes that these benefits don't outweigh the pattern's fundamental architectural drawbacks.

* In essence, Nystrom's strong stance is that while the Singleton pattern can ensure a single instance and provide global access, its downsides (coupling, testability, flexibility) often make it an anti-pattern in game development, and developers should carefully consider alternatives before reaching for it.

---

Is that summary as good as actually reading the book? Probably not. Will people rely on that and skip buying my book. Almost certainly so.

Such imperfect measures offer a compromise between "big tech can steal everything" and "LLMs trained on unpurchased books are illegal".

It's not just books but any tragedy-of-the-commons situation where a "feeder industry" for training can be fatally undermined by the very LLM that desires future training data from that industry.

IMHO, a better example would be the AI generated summaries provided by Google. Often these summaries have sufficient information and detail that people do not read the source article. The authors aren't getting paid (perhaps through on-page ads, which are not viewed) and then go out of business.

This strikes me as a good fit for the tax-on-cassette metaphor.

I doubt the exact replica stuff will stand, as technically it was only achievable via advanced prompt engineering (hacking), not simply asking for a replica. So their 2 other arguments boils down to scraping a news database = infringement and LLM output = derivative works.

Copyright infringement does not deprive the copyright owner of its property and is not criminal. So in this case only the lawsuit part applies. The owner is only entitled to the monetary damages, which is the lost sale. But in this case the sale price was paid to the owner 1 month later, so the only real damages will be the interest the publisher could have earned if they had got their money one month earlier.

Google literally scrapes pirated content all day every day. When they do that they have no idea if the content was legally placed on that website. Yet, they scan and index it anyway because there's actually no way to know (at all!). There's no great big database of all copyrighted works they can reference.

I'm not saying Meta and Anthropic didn't know they were pirating content. I'm saying that it should be moot since they never distributed it. You can't claim a violation of copyright for content that was never actually "copied" (aka distributed). The site/seeders that uploaded the content to Meta/Anthropic are the violators since copyright is all about distribution rights.

Its like saying it should be legal for me to have this Judges nudes obtained 100% illegally as long as I pixelate all the naughty bits.

You have to call it "Starcrash" (https://www.imdb.com/title/tt0079946/?ref_=ls_t_8). Then it's legal.