Anthropic judge rejects $1.5B AI copyright settlement

I'm an author, and I've confirmed that 3 of my books are in the 500K dataset.

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.

How is it fair? Do you expect 9,000 from Google, Meta, OpenAI, and everyone else? Were your books imitated by AI?

Infringement was supposed to imply substantial similarity. Now it is supposed to mean statistical similarity?

>Were your books imitated by AI?

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.

Copyright should be about copying rights, not statistical similarities. Similarity vs causal link - a different standard all together.

The entire purpose of training materials is to copy aspects of them. That’s the causal link.

The aspect it's supposed to copy is the statistics of how words work.

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.

In training, the model is trained to predict the exact sequence of words of a text. In other words, it is reproducing the text repeatedly for its own trainings. The by-product of this training is that it influences model weights to make the text more likely to be produced by the model -- that is its explicit goal. A perfect model would be able to reproduce the text perfectly (0 loss).

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).

> In training, the model is trained to predict the exact sequence of words of a text. In other words, it is reproducing the text repeatedly for its own trainings.

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?

There isn’t a clear line for extreme overfitting here.

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.

You want different levels of fitting for different things, which is difficult. Tight fighting on grammar and APIs and idioms, loose fitting on creative text, and it's hard to classify it all up front. But still, if it can recite harry potter that's not on purpose, and it's never trained to predict a specific source losslessly.

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.

> 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-...

I will say that picking the most likely word every single time isn't optimal.