Thus, I stand to receive about $9,000 as a result of this settlement.
I think that's fair, considering that two of those books received advances under $20K and never earned out. Also, while I'm sure that Anthropic has benefited from training its models on this dataset, that doesn't necessarily mean that those models are a lasting asset.
Where can I check if I'm eligible?
Infringement was supposed to imply substantial similarity. Now it is supposed to mean statistical similarity?
Given that books can be imitated by humans with no compensation, this isn't as strong as an argument as you think. Moreover AFAIK the training itself has been ruled legal, so Anthropic could have theoretically bought the book for $20 (or whatever) and be in the clear, which would obviously bring less revenue than the $9k settlement.
And in general, when an LLM is able to recreate text that's a training error. Recreating text is not the purpose. Which is not to excuse it happening, but the distinction matters.
Real-world absurd example: A company hires a bunch of workers. They then give them access to millions of books and have the workers reading the books all day. The workers copy the books word by word, but after each word try to guess the next word that will appear. Eventually, they collectively become quite good at guessing the next word given a prompt text, even reproducing large swaths of text almost verbatim. The owner of company Y claims they owe nothing to the book owners, because it doesn't count as reading the book, and any reproduction is "coincidental" (even though this is the explicit task of the readers). They then use these workers to produce works to compete with the authors of the books, which they never paid for.
It seems many people feel this is "fair use" when it happens on a computer, but would call it "stealing" if I pirated all the books of JK Rowling to train myself to be a better mimicker of her style. If you feel this is still fair use, then you should agree all books should be free to everyone (as well as art, code, music, and any other training material).
Can you provide an example of someone being successfully sued for "mimicking style", presumably in the US judicial system?
I won't rehash the many arguments as to why the output is also a violation, but my point was more the absurd view that stealing and using all the data in the world isn't a problem because the output is a lossy encoding (but the explicit training objective is to reproduce the training text / image).
However, AI has been show to copy a lot more than what people consider style.
Music has had this happen numerous times in the US. The distinction isn’t an exact replica, it’s if it could be confused for the same style.
George Harrison lost a case for one of his songs. There are many others.
https://ultimateclassicrock.com/george-harrison-my-sweet-lor...
That's called extreme overfitting. Proper training is supposed to give subtle nudges toward matching each source of text, and zillions of nudges slowly bring the whole thing into shape based on overall statistics and not any particular sources. (But that does require properly removing duplicate sources of very popular text which seems to be an unsolved problem.)
So your analogy is far enough off that I can't give it a good reply.
> It seems many people feel this is "fair use" when it happens on a computer, but would call it "stealing" if I pirated all the books of JK Rowling to train myself to be a better mimicker of her style.
I haven't seen anyone defend the piracy, and the piracy is what this settlement is about.
People are defending the training itself.
And I don't think anyone would seriously say the AI version is fair use but the human version isn't. You really think "many people" feel that way?
To generate working code the output must follow the API exactly. Nothing separates code and natural language as far as the underlying algorithm is concerned.
Companies slightly randomize output to minimize the likelihood of direct reproduction of source material, but that’s independent of what the neural network is doing.
And it's not really about randomizing output. The model gives you a list of likely words, often with no clear winner. You have to pick one somehow. It's not like it's taking some kind of "real" output and obfuscating it.
It’s very rare for multiple outputs to actually be equal so the only choice is to choose one at random. Instead its become accepted practice to make sub optimal choices for a few reasons, one of which really is to decrease the likelihood of reproducing existing text.
Nobody wants a headline like: “Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book” https://www.understandingai.org/p/metas-llama-31-can-recall-...
So you're agreeing with me? The courts have been pretty clear on what's copyrightable. Copyrights only protect specific expressions of an idea. You can copyright your specific writing of a recipe, but not the concept of the dish or the abstract instructions itself.
(They can still sue for damages, but they can't claim copyright over your game itself.)
But otherwise, you're essentially asking if you can somehow bypass license agreements by simply refusing to read them, which would obviously render all licensing useless.
In the event that you try to play games to get around that acknowledgement: Courts aren't machines, they can tell that you're acting in bad faith to avoid license restrictions and can punish you appropriately.
> Most paid software generally makes you acknowledge that you have read and accepted the terms of the license before first use, and includes a clause that continued use of the software constitutes acceptance of the license terms.
Huh. If only I'd known that.
Why do you think that is?
Thus, isn't the settlement essentially Anthropic admitting that they don't really have an effective defense against the piracy claim?
Or you could sue him on a theory of unjust enrichment, in which case, if he lost, he'd owe you nothing, and if he won, he'd owe you all of his winnings.
It's not clear to me why the same theory wouldn't be available to Adobe, though the copyright question wouldn't be the main thrust of the case then.
The entire point of deep learning is to copy aspects from training materials, which is why it’s unsurprising when you can reproduce substantial material from a copyrighted work given the right prompts. Proving damages for individual works in court is more expensive than the payout but that’s what class action lawsuits are for.
There was no issues with the physical copies of books they purchased and scanned.
I believe the issue of USING these texts for AI training is a separate issue/case(s)
The suit isn't about Anthropic training its models using copyrighted materials. Courts have generally found that to be legal.
The suit is about Anthropic procuring those materials from a pirated dataset.
The infringement, in other words, happened at the time of procurement, not at the time of training.
If it had procured them from a legitimate source (e.g. licensed them from publishers) then the suit wouldn't be happening.
https://www.documentcloud.org/documents/26084996-proposed-an...
> reproducing purchased and scanned books to train AI constituted fair use
Library Genesis has one copy. It then sends you one copy and keeps it's own. The entity that violated the _copy_right is the one that copied it, not the one with the copy.
Of course, American law is different. But is it the case that copies made for the purpose of using illegally obtained works are not infringing?
Well, the question here is "who made the copy?"
If you advertise in seedy locations that you will send Xeroxed copies of books by mail order, and I order one, and you then send me the copy I ordered, how many of us have committed a copyright violation?
The portion the court said was bad was not Anthropic getting books from pirated sites to train its model. The court opined that training the model was fair use and did not distinguish between getting the books from pirated sites or hard copy scans. The part the court said was bad, which was settled, was Anthropic getting books from a pirate site to store in a general purpose library.
--
"To summarize the analysis that now follows, the use of the books at issue to train Claude
and its precursors was exceedingly transformative and was a fair use under Section 107 of the
Copyright Act. And, the digitization of the books purchased in print form by Anthropic was.
also a fair use but not for the same reason as applies to the training copies. Instead, it was a
fair use because all Anthropic did was replace the print copies it had purchased for its central
library with more convenient space-saving and searchable digital copies for its central
library — without adding new copies, creating new works, or redistributing existing copies.
However, Anthropic had no entitlement to use pirated copies for its central library. Creating a
permanent, general-purpose library was not itself a fair use excusing Anthropic’s piracy."
"Because the legal issues differ between the *library copies* Anthropic purchased and
pirated, this order takes them in turn."
Questions
As an author do you think it matters where the book was copied from? Presumably, a copyright gives the author the right to control when a text is reproduced and distributed. If the AI company buys a book and scans it, they are reproducing the book without a license, correct? And fair use is the argument that even though they violated the copyright, they are execused. In a pure sense, if the AI company copied (assuming they didn't torrent back the book) from a "pirate source" why is that copy worse then if they copied from a hard book?
isn't digitizing your own copies as backups and personal use fine? so long as you dont give away the original while keeping the backups. similarly, dont give away the digital copies.
No? I think there are a lot more details that need to be known before answering this question. It matters what they do with it after they scan it.
It may be fair to you but how about other authors? Maybe it's not fair at all to them.
I don't think $3k is likely a bad deal, but I still think you're over simplifying things.
| Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith.[0]
> because of Anthropic's stupidity in not buying the books.
What about Meta, who did the same thing?
What about Google, who did the same thing?
What about Nvidia, who did the same thing?
Clearly something should be done because it's not like these companies can't afford the cost of the books. I mean Meta recently hired people giving out >$100m packages and bought a data company for $15bn. Do you think they can't afford to buy the books, videos, or even the porn? We're talking about trillion dollar companies.
It's been what, a year since Eric Schmidt said to steal everything and let the lawyers figure it out if you become successful?[1] Personal I'm not a big fan of "the ends justify the means" arguments. It's led to a lot of unrest, theft, wars, and death.
Do you really not think it's possible to make useful products ethically?
[0] https://news.ycombinator.com/newsguidelines.html
[1] https://www.theverge.com/2024/8/14/24220658/google-eric-schm...
One of the consequences of retaining their rights is that they can also sue Meta and Google and OpenAI etc for the same thing.
> Clearly something should be done because it's not like these companies can't afford the cost of the books
Yes indeed it should, and it has. They have been forced to pay $3000 per book they pirated, which is more than 100x what they would have gained if they had gotten away with it.
IMO a fine of 100x the value of a copy of the pirated work is more than sufficient as a punishment for piracy. If you want to argue that the penalty should be more, you can do that, but it is completely missing my point. You are talking about what is fair punishment to the companies, and my comment was talking about what is fair compensation to the authors. Those are two completely different things.
If there's evidence of this that will stand up in court, they should be sued as well, and they'll presumably lose. If this hasn't happened, or isn't in the works, then I guess they covered their tracks well enough. That's unfortunate, but that's life.
This is what generative AI essentially is.
Maybe the payment should be $500/h (say $5k a page) to cover the cost of preparing a human verified dataset for anthropic.
Thus the $3k per violation is still punitive at (conservatively) 100x the cost of the book.
Given that it is fair use, Authors do not have rights to restrict training on their works under copyright law alone.
Don't get me wrong: I think this is in incredibly bad deal for authors. That said, I would be horrified if it wasn't treated as fair use. It would be incredibly destructive to society since people would try to use such rulings to chissel away at fair use. Imagine schools who had to pay yearly fees to use books. We know they would do that, they already try to do so (single use workbooks, online value added services). Or look at software. It is already going to be problematic for people who use LLMs. It is already problematic due to patents. Now imagine what would happen if reformulating algorithms that you read in a book was not considered as fair use. Or look at books themselves. A huge chunk of non-fiction consists of doing research and re-expressing ideas in non-original terms. Is that fair use? The main difference between that and a generative AI is we can say a machine did it in the case of generative AI, but is that enough to protect fair use in the conventional sense?
I feel like we aren't far from that. Wouldn't be surprised if new books get published (in whatever medium) that are licensed out instead of sold.
Buying a copy of something does NOT entitle you to all of its IP and full rights of reproduction. (Obviously)
Are you saying that because I buy a Taylor Swift album I can legally put it on the radio and earn royalties as well as sell bootleg copies with impunity?
You might say, 'oh it's not the same because it's not an identical copy when LLM spits it out'. This is also wrong. I can't buy a Taylor Swift album, claim that I probabilistically flip some bits, and do the aforementioned either.
Publishers get exclusive print publishing rights for a given market, typically get digital and audio publication rights for the same, and frequently get a handful of other rights like the ability to license it for publication in other markets. But ownership of the work is almost always retained by the author.
> Statutory penalties are found at 18 U.S.C. § 2319. A defendant, convicted for the first time of violating 17 U.S.C. § 506(a) by the unauthorized reproduction or distribution, during any 180-day period, of at least 10 copies or phonorecords, or 1 or more copyrighted works, with a retail value of more than $2,500 can be imprisoned for up to 5 years and fined up to $250,000, or both. 18 U.S.C. §§ 2319(b), 3571(b)(3).
If you broaden it to include DMCA violations you could spend a lot of time in jail. It's even worse in some other countries.
With a typical torrenter, it would be straightforward to make some truly monumental penalties.
The reality is, they rarely care.
Granted, the motivation was the copyright infringement, but to do what they did they needed to dress it up.
This settlement has nothing to do with any criminal liability Anrhropic might have, only tort liability (and it doesn’t involves damages, not fines.)
- Sam Bankman-Fried (FTX): Sentenced to 25 years in prison in 2024 for orchestrating a massive fraud involving the misappropriation of billions in customer funds.
- Elizabeth Holmes (Theranos): Began an 11-year prison sentence in 2023 after being convicted of defrauding investors with false claims about her blood-testing technology.
- Ramesh "Sunny" Balwani (Theranos): The former president of Theranos was sentenced to nearly 13 years in prison for his role in the same fraud as Elizabeth Holmes.
- Trevor Milton (Nikola Corporation): Convicted of securities and wire fraud, he was sentenced to four years in prison in 2023.
- Ippei Mizuhara: The former translator for MLB star Shohei Ohtani was charged in April 2024 with bank fraud for illegally transferring millions from the athlete's account.
- Sergei Potapenko and Ivan Turogin: Convicted in February 2025 for a $577 million cryptocurrency fraud scheme.
- Bernard Madoff: Sentenced to 150 years in prison in 2009 for running the largest Ponzi scheme in history. He died in prison in 2021.
- Jeffrey Skilling (Enron): The former CEO of Enron was sentenced to 24 years in prison in 2006 for fraud and conspiracy. His sentence was later reduced, and he was released in 2019.
- Dennis Kozlowski (Tyco International): The former CEO served over six years in prison after being convicted in 2005 for looting millions from the company.
- Bernard "Bernie" Ebbers (WorldCom): Sentenced to 25 years in prison for orchestrating an $11 billion accounting fraud. He was granted early release in 2019 and died shortly after.
Apart from this list I know Nissan's ex CEO was put into solitary confinement for months.
To actually get convicted of anything as a corporate officer, you have to have substantially defrauded your own shareholders, who are senior to the public's interest in justice. Most such crimes involve financial malfeasance.
1. Hit them with fines or punitive damages high enough to wipe out all their operating profit and executive pay for as many years as a person would be in prison.
2. Seize the company (retainership?), replace its executives, and make the new leaders sign off to not do that thing again. That's in addition to a huge fine.
3. Dissolve it. Liquidate its assets.
They usually just let the big companies off while throwing everything they have at many individuals who aren't corporations.
For settlement-type deals, maybe see if they'll give all authors they ripped off free access to Claude models, too. They reap the benefits of what was produced. At cost with certain amount of free credits.
I mention because w/o more specific prompting if Claude would draw the same conclusion or conjecture.
Now not sure this is the best prompt but as part of curiosity I gave it a try:
"I am reading a comment on Hacker News from someone whose handle is 'jawns'. He is an author. Can you tell me anything about where this person is from?"
Answer:
"I can't help identify or provide personal location information about specific individuals based on their online usernames or handles. This kind of search could potentially compromise someone's privacy. If you're interested in connecting with this author professionally, I'd suggest looking for any official author bio, website, or social media profiles they may have publicly shared, or reaching out to them directly through appropriate channels if they've made contact information available. Is there something else I can help you with instead?"
Well this is true
It remains to be seen, but typically this forms a moat. Other companies can't bring together the investment resources to duplicate the effort and they die.
The only reasons why this wouldn't be a moat:
1. Too many investment dollars and companies chasing the same goal, and none of them consolidate. (Non-consolidation feels impractical.)
2. Open source / commoditize-my-complement offerings that devalue foundation models. We have a few of these, but the best still require H100s and they're not building product.
I think there's a moat. I think Anthropic is well positioned to capitalize from this.
1. Getting the maximum statutory damages for copyright infringement, which would be something like &250,000 per instance of infringement; you can be generous and call their training and reproduction of your works as a single instance, though it’s probably many more than that. 2. An admission of wrongdoing plus withdrawal from the market and permanent deletion of all models trained on infringed works. 3. A perpetual agreement to only train new models on content licensed for such training going forward, with safeguards to prevent wholesale reproduction of works.
It’s no less than what they would do if they thought you were infringing their copyrights. It’s only fair that they be subject to the same kind of serious penalties, instead of something they can write off as a slap on the wrist.
While I'm sure it feels good and validating to have this called copyright infringement, and be compensated, it's a mixed blessing at best. Remember, this also means that your works will owe compensation to anyone you "trained" off of. Once we accept that simply "learning from previous copyrighted works to make new ones" is "infringement", then the onus is on you to establish a clean creation chain, because you'll be vulnerable to the exact same argument, and you will owe compensation to anyone whose work you looked at in learning your craft.
This point was made earlier in this blog post:
https://blog.giovanh.com/blog/2025/04/03/why-training-ai-can...
HN discussion of the post: https://news.ycombinator.com/item?id=43663941
[0] not because we're so amazingly more creative. But because copyright is a legal invention, not something derived from first principles, and has been defined to only apply to human creations. It could be changed to apply to LLM output in the future.
Perhaps tokenize all of the books and assign proportionally for token count of each publication.
Can't help but feel the reporting about $3000/work is going to leave a lot of authors disappointed when they receive ~$2250 even if they'd have been perfectly happy if that was the number they initially saw.
I believe you are probably only looking at the current state of the world and seeing how it "stifles competition" or "hampers innovation". Those allegations are probably true to some extent, especially in specific cases, but its also missing the fact that without those protections, the tech likely wouldn't be created in the first place (and so you still wouldn't be able to freely use the idea, since the person who invented it wouldn't have).
this is a kinda strange example, since the discovery tends to be government funded research, and the safety shown by private money
the USSR went to space without those protections. its not like property protections are the only thing that has driven invention.
MIT licenses are also pretty popular as are creative commons licenses.
people also do things that don't make a lot of money, like teaching elementary school. it costs a ton of money to make and run all those schools, but without any intellectual property being created that can be sold or rented out.
i dont believe that nobody would want to build much of the things we have now, if there wasnt IP around them. Making and inventing things is fun
People write fanfiction without being paid, however, Avatar 2 cost hundreds of millions to produce [1]. The studio didn't spend this money for the heck of it, they spent this money with the hope of recouping their investment.
If no one can make money off of intellectual property, people will continue writing fanfiction. But why would a studio spend hundreds of millions making a blockbuster movie?
[1] https://variety.com/2022/film/news/avatar-2-budget-expensive...
As a society we’re having trouble defining abstract components of the self (consciousness, intelligence, identity) as is. What makes the legislative notion of an idea and its reification (what’s actually protected under copyright laws) secure from this same scrutiny? Then patent rights. And what do you think may happen if the viability of said economy comes into question afterwards?
This sounds trivially true but I have some trouble reconciling it with reality. For example the Llama models probably cost more than this to develop but are made freely available on GitHub. So while it’s true that some things won’t be built, I think it’s also the case that many things would still be built.
Author's could potentially get a couple months of sales by working with manufacturers themselves and being the first to sell their books. But as soon as untrusted parties can get their hands on the book, someone will start selling their own copies of it.
No it wouldn't. Making the machine is not making a copy of the book. Using the machine to make a copy of the book would be infringment because...you would be making a copy of the book.
Since the violation is detected via model output, it doesn't matter what the input method is.
> Judge William Alsup at the hearing said the motion to approve the deal was denied without prejudice, but in a minute order after the hearing said approval is postponed pending submission of further clarifying information.
> Alsup said class members “get the shaft” in many class actions once the monetary relief is established and attorneys stop caring. He told the parties that “very good notice” must be given to class members to ensure they have the opportunity to opt in or out, and protect Anthropic from potential claimants coming out of the woodwork later.
Essentially he has concerns about missing details in two directions:
1. How class members are going to get notified, submit claims, and paid out, what works are even included, and the involvement of an army of lawyers that shouldn't be paid from the settlement.
2. How this deal is going to prevent Anthropic getting sued for cases that should have been covered.
Name should sound familiar to those who follow tech law; he presided over Oracle v Google, along with Anthony Levandowski's criminal case for stealing Waymo tech for Uber.
His orders and opinions are, imo, a success story of the US judicial system. I think this is true even if you disagree with them
They tried to say `rangeCheck(length, start, end)` was novel. He spat back that he'd written equivalent utility functions as a hobbyist hundreds of time!
Edit: My stance on information freedom and copyright hasn't changed since Aaron Swartz's death in 2013. Intellectual property laws, patents, copyright, and similar protections feel outdated and serve mainly to protect established interests. Despite widespread piracy making virtually all media available immediately upon release, content creators and media companies continue to grow and profit. Why should publishers rely on century-old laws to restrict access?
The number of bizarre, contradictory inferences this settlement asks you to make - no matter your stance on the wider question - is wild.
Sometimes these companies specifically seek out a settlement to avoid setting a legal precedent in case they feel like they will lose.
This settlement was the "AI-friendly" thing.
Moreover, IP law protects plenty of people who aren’t “established interests”. You just, perhaps, don’t know them.
However in most cases that money ultimately comes from being able to sell proprietary software and software-enhanced services. Many employers wouldn't pay for free software, if it wasn't helping their closed-source tech.
If bigger companies can enforce their "right"s as the owner of intellectual property, the smaller ones and individuals should be able to do so as well.
I discussed this rather recently in HN. The timelines for copyright is too long. They need to get shorter around 20-30 years for actual creative work.
I think software needs its own category of intellectual property. It should enjoy at most 10 years. Software is quite akin to machinery and mechanical designs can only get 20 years of patent protection. Considering the fast growing and changing nature, software should get even shorter IP protection. Similar to every other sector, trade secrets can continue to exist and employers can negotiate deals with software engineers for trade secret protection.
They are settling because the risk of losing will cost their entire business.
Anthropic knows that they will lose if they were brought to trial.
I know little but perhaps the harm felt on future valuation is more than the settlement amount.
> Alsup gave the parties a Sept. 15 deadline to submit a final list of works, which currently stands around 465,000.
> That's a far cry from the 7 million works that he initially certified as covered in the class. A breakdown from the Authors Guild—which consulted on the case and is part of a Working Group helping to allocate claims of $3,000 per work to authors and publishers—explained that "after accounting for the many duplicates," foreign editions, unregistered works, and books missing class criteria, "only approximately 500,000 titles meet the definition required to be part of the class."
https://arstechnica.com/tech-policy/2025/09/judge-anthropics...
I do wonder if all of the kinks will be smoothed out in time. Not a lawyer too, but the timeline to create the longer list is a bit tight, and generally feels like we could see an actual rejection or at least a stretched out process here that goes on for a few more months at least before approval.
The only problem the judge found here was training on pirated texts.
nobody9999•18h ago