9k authors say AI firms exploited books to train chatbots | Hacker News Reader

9k authors say AI firms exploited books to train chatbots

latimes.com

30 pointswebmaven12 comments

"compensating the authors for using their works would be the fair move"

Fundamentally their asking to renegotiate the sale price of their work depending on what the purchaser decides to do with the book.

I'm looking forward to the time when I must declare if I'm purchasing a book for:

  - personal enjoyment
  - to produce a parody
  - burn it

Note that one of the claims regarding the books in these corpora is that they were not purchased in the first place, but were pirated, so the First Sale doctrine (or Fair Dealing, depending on your jurisdiction) doesn't really apply.

You could devise a scenario whereby an artist deliberately used pirated works in order to make a satirical[0] point about copyright itself (or the global publisher oligopoly, etc.) that might be accepted by a judge as a fair use argument, but that is not what's happening here.

[0] Parody is when you change a work to make fun of it (eg. a twilight parody where Edward and Bella's dialog is in Valspeak). Satire is when you change a work to make fun of something *else* (eg. a version of Twilight where Edward and Bella are replaced with Putin and Trump). The general reason Satire isn't as protected by Fair Use as Parody is because you can usually replace the (ab)used work with something else while still making a similar point about the actual subject of your critique (eg. make a Putin/Trump version of Pride and Prejudice instead of using Twilight).

AI is fundamentally about redistributing the content it trains on.

I don't think this is a reasonable argument, but I do think it might still prevail in court, which from the POV of the authors an their lawyers makes it a worthwhile filing. Personally I hope it doesn't prevail, the implications for how people learn and synthesize based on an existing corpus is too profound.

Doesn't 'fair use' come into play at some point? I know this is all new, but isn't this approximately what fair use is about? Even a little modification of prior art is enough to qualify. Embedding it into an AI corpus is quite a bit of modification.

Or am I way behind the curve on this one.

One of the claims is that the book corpora weren't even sourced legally. It is harder to make a fair use claim if the copy the model was trained on was pirated to begin with. Contemporary in-copyright ebooks are a bit different in this regard than images found on the public web.

The article itself says "Not only does the recent Supreme Court decision in Warhol v. Goldsmith make clear that the high commerciality of your use argues against fair use, but no court would excuse copying illegally sourced works as fair use."

Fair use isn’t straightforward to begin with, but basically there’s room for a Judge to narrow its applicability in this instance.

Here’s the four factor test courts use:

> In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include:

> 1. the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;

> 2. the nature of the copyrighted work;

> 3. the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and

> 4. the effect of the use upon the potential market for or value of the copyrighted work.

I think points 1 and 3 are the least debatable: commercial LLMs should easily fail 1 and all LLMs should pass 3. EDIT: actually changed my mind on this. If they’re using the whole work, then LLMs should actually fail #3. It’s not just how much of the model it makes which is insignificant relative to all the rest of the text in the model, but also how much of each individual work is used. If that’s 100% then that’s actually an easy fail.

We could have a debate about #2 till the end of time and I’m not really here for that, not today anyway, but it is probably worth debating.

#4 is also absolutely debatable, but I think there is a large potential for LLMs to harm the market for the original source material that they were trained on.

Considering the AI community it would be very easy to portray them as: portraying the derivatives as a new/original work, in a commercial gold rush (these first two brutally violate #1), violating the entirety of the work (#3), and a massive chilling effect on the market (#4 and how hard writer/actor strikes are anti-AI with studios simultaneously refusing to budge in any way on AI).

They can portray it however they want but if they lose in court because they can’t keep kosher then they lose in court. At which point their only recourse is Congress.

Thank you that is quite useful.

Three is confusing to me. When somebody make a Mona Lisa with a mustache, they've used the entirety. It's how much they modified that's important? Whether it can be mistaken for the original.

Nobody is going to mistake an AI-ingested corpus for the original. It seems quite debatable.

It’s a four factor test. You can lose on one of the factors and still have it be determined to be fair use by a Judge, but if the Mona Lisa were still under copyright (shouldn’t be, but we’ll use it for this example anyway), then per Andy Warhol Foundation for the Visual Arts v. Goldsmith, it would likely turn a lot on whether you commercialized it and whether it affected the market for the original.

Reply on news.ycombinator.com