Anthropic destroyed millions of physical books to train its AI, court documents reveal

[Daniel Sims]

WTF?! Generative AI has already faced sharp criticism for its well-known issues with reliability, its massive energy consumption, and the unauthorized use of copyrighted material. Now, a recent court case reveals that training these AI models has also involved the large-scale destruction of physical books.

Buried in the details of a recent split ruling against Anthropic is a surprising revelation: the generative AI company destroyed millions of physical books by cutting off their bindings and discarding the remains, all to train its AI assistant. Notably, this destruction was cited as a factor that tipped the court's decision in Anthropic's favor.

To build Claude, its language model and ChatGPT competitor, Anthropic trained on as many books as it could acquire. The company purchased millions of physical volumes and digitized them by tearing out and scanning the pages, permanently destroying the books in the process.

Furthermore, Anthropic has no plans to make the resulting digital copies publicly available. This detail helped convince the judge that digitizing and scraping the books constituted sufficient transformation to qualify under fair use. While Claude presumably uses the digitized library to generate unique content, critics have shown that large language models can sometimes reproduce verbatim material from their training data.

Anthropic's partial legal victory now allows it to train AI models on copyrighted books without notifying the original publishers or authors, potentially removing one of the biggest hurdles facing the generative AI industry. A former Metal executive recently admitted that AI would die overnight if required to comply with copyright law, likely because developers wouldn't have access to the vast data troves needed to train large language models.

Still, ongoing copyright battles continue to pose a major threat to the technology. Earlier this month, the CEO of Getty Images acknowledged the company couldn't afford to fight every AI-related copyright violation. Meanwhile, Disney's lawsuit against Midjourney – where the company demonstrated the image generator's ability to replicate copyrighted content – could have significant consequences for the broader generative AI ecosystem.

That said, the judge in the Anthropic case did rule against the company for partially relying on libraries of pirated books to train Claude. Anthropic must still face a copyright trial in December, where it could be ordered to pay up to $150,000 per pirated work.

Permalink to story:

Anthropic destroyed millions of physical books to train its AI, court documents reveal

 

[candle86]

Makes sense, if they bought they can use it.

 

[Kotters]

Destroying books makes for a sensational headline, to be sure.

 

[DrSuess]

Not sure about that as the the book was converted from a physical format to a digital one during the training process. In theory a book can be used by one person at a time, an AI trained on a book can be used by how ever many users the server can handle (millions of users). I'm not sure fair usage applies, or is in fact fair to the copyright holder. It sounds like their should be some licensing type agreements like artist have with Spotify.

One book sold and destroyed doesn't seem to equal potentially million upon millions of request for information on a book that provide 0 revenue for the copyright holder.

 

[m4a4]

Not sure about that as the the book was converted from a physical format to a digital one during the training process. In theory a book can be used by one person at a time, an AI trained on a book can be used by how ever many users the server can handle (millions of users). I'm not sure fair usage applies, or is in fact fair to the copyright holder. It sounds like their should be some licensing type agreements like artist have with Spotify.

One book sold and destroyed doesn't seem to equal potentially million upon millions of request for information on a book that provide 0 revenue for the copyright holder.

You're looking at it wrong.
If a person was to read and learn from a book (and that book was only read by them), said person could theoretically talk about that book to millions of people.
It's the same concept here. Relaying the information learned from a source is fair use for humans (as long as it's "transformative"), and would be the same for AI.

 

[stewi0001]

I have a bunch of old books they can have.

 

[Squid Surprise]

It’s “new” tech and old laws don’t apply - without reinterpretation.

That reinterpretation simply depends on the agendas of those making the rulings. If we value AI over copyright owners, than the AI companies win… if we decide that AI is no longer useful/desirable, you’ll see the copyright holders win.

I’d be betting on AI…

 

[Kirby1]

It's never going to be illegal for AI to train in the same ways humans can just because it can do it faster and better. You wouldn't see a person put on trial because they trained themself on too many books.


They bought the books. They can do literally whatever they want with them.

 

[BorisMcKennedy]

It's never going to be illegal for AI to train in the same ways humans can just because it can do it faster and better. You wouldn't see a person put on trial because they trained themself on too many books.


They bought the books. They can do literally whatever they want with them.

That’s not strictly true. Books are copyrighted and can’t be reproduced, or profit made without compensating the copyright owner.

 

[Kirby1]

That’s not strictly true. Books are copyrighted and can’t be reproduced, or profit made without compensating the copyright owner.

Yes, you can't buy a book and make copies of it for sale. But you can learn from it and quote it freely as much as you want. The information within the book is free to use.

 

[techdev]

Yes, you can't buy a book and make copies of it for sale. But you can learn from it and quote it freely as much as you want. The information within the book is free to use.

Yes, the information within the book is free to use, as long you don't make money out of it.
Claude AI has different tiers besides the free one. Pro is 15 € per month and Max starts from 90 € per month.

Using other analogy, it's fine for me to learn the guitar, learn how to play "Stairway to Heaven" by Led Zeppelin, and play a cover of it as long as I play it in front of an audience free of charge.
The moment I am starting to charge for listening to a Led Zeppelin cover, I will get sued because I am infringing copyright laws and not compensating Led Zeppelin as the legitimate copyright owners.
Same applies to movies using copyrighted songs. Permission must be asked and rights paid to use a song.

Their plan was to NEVER compensate any copyright owner. I don't believe part of the money you would pay Anthropic for would actually go to copyright owners.

 

[DrSuess]

You're looking at it wrong.
If a person was to read and learn from a book (and that book was only read by them), said person could theoretically talk about that book to millions of people.
It's the same concept here. Relaying the information learned from a source is fair use for humans (as long as it's "transformative"), and would be the same for AI.

I don't think I am, no one I know would be able to regurgitate the entire book, not just once but an unlimited amount of times to an unlimited number of people over an infinite period of time. Your point of view simple doesn't add up to me.

 

[BuckarooBonzaii]

I think the word is "Fair Use" that makes it complicated and depending on situations. Simply put and in general, it allows you to use copyrighted material for teaching, research, commenting, criticism and more, all which makes it a balancing act. Since AI is still being developed, is it considered research and in a learning process? (AI reading the books).

When you buy the physical copies of books you can lend it, resell it, burn it, even use the pages for toilet paper. But again Copyright and Fair Use is a balancing act.

 

[Porkous]

So how does it work ? A Circuit investigation into LLM database won't allow to copy the data it used ? In other words, If I ask the model to show a text from a book, maybe few pages, even the entire book... Won't work ? LLM encoding is an entire universe of its own language?

 

[pnntmp]

So what?
It's like revealing someone used the toilet paper they bought, but I don't remember articles about the latter.

When you buy something, you can do whatever you like with it - this includes destroying it.
Of course, it's utterly ridiculous to frame this as 'destruction' - each 'destroyed' book becomes accessible to billions of people. This is a good thing.

 

[mcnabney]

They are literally making derivative works from IP owned by others. That is specifically prohibited by copyright law.

 

[Squid Surprise]

They are literally making derivative works from IP owned by others. That is specifically prohibited by copyright law.

What do you mean by "derivative works"?

 

[onebignerd]

It's never going to be illegal for AI to train in the same ways humans can just because it can do it faster and better. You wouldn't see a person put on trial because they trained themself on too many books.


They bought the books. They can do literally whatever they want with them.

Wrong! Buying the books you are still bound by copyright laws protecting the content. A person reading the books to "train themselves" is quite different to digitally scanning books, which is copying the content and using copyrighted content to train AI. Fair use does not include replicating the content into an AI database to be republished and used by millions of users free of charge. That is no different than photocopying the books and distributing them to millions.

 

[onebignerd]

It's never going to be illegal for AI to train in the same ways humans can just because it can do it faster and better. You wouldn't see a person put on trial because they trained themself on too many books.


They bought the books. They can do literally whatever they want with them.

Duplicate posting removed

 

[mcnabney]

What do you mean by "derivative works"?

Additional content based upon existing IP that you don't own.

For example, you can't sell or make money off of anything that you don't have the rights to. Like writing a Harry Potter book and selling it, or making Harry Potter toys and selling them. You can make stuff for yourself, but you can't make money off of it. That is where Anthropic (and Facebook and others) is doing. They are selling a product or service that is based upon IP that they don't own.

 

[Squid Surprise]

Additional content based upon existing IP that you don't own.

For example, you can't sell or make money off of anything that you don't have the rights to. Like writing a Harry Potter book and selling it, or making Harry Potter toys and selling them. You can make stuff for yourself, but you can't make money off of it. That is where Anthropic (and Facebook and others) is doing. They are selling a product or service that is based upon IP that they don't own.

You can’t do it based on ONE work… but deriving content from MULTIPLE sources is perfectly fine… it’s the difference between plagiarism and research

 

[New Day]

Humans at large don't remember all in detail, which would end this ordeal by not needing more and more books, nor mega storage facilities. Genetics to the rescue!

 

[Squid Surprise]

Why destroyed millions of physical books to train AI? Why not scan those books into the computer?

They did…. But when you’re scanning multiple books, the fastest way is to remove the spine and bindings…. That leaves you with a bunch of loose leaf paper which would be costly (and time intensive) to re-bind…

 

[Kotters]

Why destroyed millions of physical books to train AI? Why not scan those books into the computer?

... to scan them, you have to destroy them. Debind the book and scan the pages.

 

[bclaymiles]

... to scan them, you have to destroy them. Debind the book and scan the pages.

False, at least in the sense of "have to". The Internet archive has custom rigs that scan the pages from the book directly, without tearing it apart.

https://archive.org/scanning

When a relative died a few years ago, we donated their books to IA. They sent a van around to pick them up, very convenient (of course, the relative lived only a few dozen miles from the headquarters; others might not have quite the same experience).