Bias disclaimer: Amazon is my current employer, but I don't work on AI or anything else mentioned in the article.

Yes, this is a result of copyright laws. The other commenters are wrong/uninformed.

If it was up to the companies training LLMs, they wouldn't destroy the books: It's a waste of company resources, it's needlessly destructive/evil, it generates bad PR, etc etc. There are essentially zero advantages, other than it is what is required under US copyright law (or at least, it is what their highly paid lawyers believe is required under US copyright law).

Isn’t it being destroyed because it makes the scanning process easier?

Perhaps partially? I assume that cutting the pages out of the spin makes them easier to scan at least partially. That said I have no insider knowledge of this type of operation so I don't know how much easier that actually makes it.

But ultimately, it's a moot point, because the legal requirement means the books must end up destroyed. Even if the people at Amazon wanted to scan the books in a way that required no destruction at all, it's not currently (legally) possible for them to do so, so they might as well take the easy way out today.

Non-destructive scans are at least 10x as much, and tend to lower quality.

https://software.annas-archive.gl/AnnaArchivist/annas-archiv...

Cutting the pages out makes them machinable. Non-destructive scans involves gently turning pages, and paying a lot of attention to the state of the spine. Destructive scans involve guillotine cutting the spine off, scanning the covers by hand, putting the pages into a hopper, clamping them in and hitting a button. While that book is scanning, you're already cutting the spine off the next book. If the machine jams, try to work the jam out gently, scan the pieces, and let the computer stitch it together.

Perhaps this is a use case that's worth investigating, to invent better, less-destructive scanning processes! Or some way to rebind them afterwards.

No.

A judge a while ago decided that as long as the physical copy is destroyed, and "transformed" into an electronic copy, you can do the upload. But if you preserve the physical copy after scanning it, you are in violation of copyright because you "copied" the book.

That's literally the only reason they are trashing them. It's a legal requirement.

Ah, so I can buy and scan a dvd, destroy it and then legally distribute the legal copy via torrents. Good to know, because that is what the LLM thieves are doing.

Distributing an exact copy of the text would be illegal; if you could get an LLM trained on a book to output the exact copy of the text from the book then that would be illegal as well.

Can you find a citation for this? I have heard this claimed rule recently from other people, and I haven't seen this decision (nor do I know what level of court or jurisdiction it might be). This is not a rule that I heard many years ago when working on and adjacent to copyright issues (including book scanning!), although of course the issue has been newly litigated again recently, so there may be new interpretations coming out.

Edit: Someone else linked to an order in Bartz v. Anthropic which appears to emphasize that destroying the original copies improved the defendant's position with respect to the fair use analysis. Is that the decision you're thinking of?

They could just not do the evil thing.

(This is why I will never be a billionaire)

Not to my understanding. To begin with, it's far from given that "rare" books are all covered by copyright. But if they are, it's at best murky: whether you destroy the original doesn't really have anything to do with what you're doing with scanned contents. The scanned contents themselves may be inherently a copyright issue, regardless of destroying the original. The actual trained model has separate arguments more in its favor, so if no scanned contents exist - IE the data is read once for training and not stored or saved, they have a better argument. But in that case the destruction is totally disconnected from copyright, as they'd be totally okay to rescan the material.

i would assume that would come into play if they were uploading scans of the books? there must be some gray area where training like this doesnt apply to that. or ya know, just do it and face the consequences later because you already have the data and know that ai obsessed government will just shrug their shoulders.

No. A copy is still a copy even if you destroy the original.

Until about a year ago this would have been a reasonable and respectable argument, but at least in California you are arguing against current legal precedent: https://fingfx.thomsonreuters.com/gfx/legaldocs/jnvwbgqlzpw/...

That's a bit of a wild interpretation of copyright law.

I mean, Anthropic isn't going to fight it because it lets them do the thing they want to do, so I can see how this never gets beyond the court that allows them to do the thing they want to do.

But would this argument would have flown in the past?

It wasn't even attempted in Sony v Universal. Or any copyright suit up until this point. That doesn't smell funny to you?

Sure, it's a bit wild, and no, the same outcome may not have been reached if the case came up earlier. And if it makes it to the Supreme Court, who knows what they will turn it into.

The flip side is that Alsup (the judge who wrote the opinion) is probably the smartest district court judge we have when it comes to technology, and one of the people I'd trust most to come up good decision.

He's a treasure, and my instinct is that he got it right: https://en.wikipedia.org/wiki/William_Alsup

Those laws are lobbied for by large corporations, these are not just laws that exist outside of that context. They can also be changed, or Amazon could just incur the fines.

Large corporations will move fast and break things when it’s convenient; they don’t care much about the law - just about profit.