Comment by tivert | Hacker News Reader

Comment by tivert | Hacker News Reader

> Importantly, the court focused on the “highly transformative purpose” of Google’s use of this material—that Google was transforming many diverse copyright-protected books into a new, useful search mechanism that didn’t compete with the underlying works.

It's pretty obvious that generative AI should fail that test; because the whole point is to compete with the underlying works, and the technology is marketed and understood that way. It doesn't have the same defense as Google Books, because that is so crippled with copyrighted works that it cannot be used as a substitute for the books it's searching.

I mean a lot of people are hyping ChatGPT as a replacement for Wikipedia (which is was obviously trained with), and image generators as replacements for stock imagery (and there are documented cases of those reproducing recognizable watermarks: https://photutorial.com/stable-diffusion-watermarks-investig...).

Also generative AI systems aren't people, so I think it's totally fair to not grant their operators rights by analogy to something a human would do.

> the whole point is to compete with the underlying works

AI doesn't compete just compete with "the underlying works" though; it competes with all works, even works that weren't in its training set. So the type of "competition" you're describing is incidental, and not connected to the AI's use of any particular copyrighted work it was trained on.

I don't know whether that makes a difference under existing law, but I'd argue that it should.

I'd also argue copyright/fair use shouldn't even come into play, since AI models do not (typically) contain copies of the works they're trained on in the first place (except in cases of over fitting), but that's a completely separate issue.

This is a weird argument. You can violate someone's copyright as long as you compete with other people too?

> because the whole point is to compete with the underlying works

I don't see how this is the case and if it hinged on that, how is it different from a book citing another book. That book competes with the book it's citing.

>> because the whole point is to compete with the underlying works

> I don't see how this is the case

If it wasn't the case, what's the point of these models getting billions of dollars thrown at them? What's the actual sales pitch for using them in the economy?

> and if it hinged on that, how is it different from a book citing another book. That book competes with the book it's citing.

One of the obvious differences is these models don't cite their sources. Even if you can get them to generate something that looks like a cite, it's not a real cite of the works the model actually used.

> If it wasn't the case, what's the point of these models getting billions of dollars thrown at them?

This is a digression. There are use cases other than creating competing books, do I need to actually list them out?

- Turning unstructured text into structured data.

- Summarizing news

- Sentiment analysis

- Semantic search

- Helping people write research papers

and more.

How do any of those compete with fiction or non-fiction books?

> One of the obvious differences is these models don't cite their sources.

Do I need to cite the math textbook that I used to learn algebra when I write a book on calculus?

Did you miss the "generative AI" in the headline? All the examples you list are general applications of NLP's, not of generative AI. Your argument would be similar to defending nuclear missiles by listing all the good things that a nuclear power plant can bring.

Do I need to cite the math textbook that I used to learn algebra when I write a book on calculus?

Yes, you do, if you put verbatim snippets of those textbooks in your own work.

> All the examples you list are general applications of NLP's, not of generative AI.

I think you are being purposefully obtuse.

My opinion means very little, but AI like to see a solution that preserves generative AI progress while giving incentives to use licensed content.

E.g.

"The burden of proof when republishing the output of AI generated content is on the person publishing the generated content. All unlicensed sources must be cited in a public searchable index."

With this, if you're a company and you accidentally use the output of Generative AI you have liability if the work is infringement.

This gives companies an incentive to have licenses for their content but doesn't strictly prohibit ~linking~ sourcing from publically available content.

This will also allow the industry to use publicly viewable content to push the boundaries of generative AI in a research capacity without fear of liability.

A person could tell a sufficiently advanced generative AI trained on licensed data to recreate something under copywrite by just describing it. You would not be able to publish that output.

> This is a digression. There are use cases other than creating competing books, do I need to actually list them out?

Are the generative models crippled to only be able to do those things?

And some of the things you list are also likely to be in direct competition with the original works (e.g. "Summarizing news.") I believe "AI" generated news sites are already a thing (e.g. plagiarizing/summarizing original news articles, so another party can generate ad-revenue from them).

> Do I need to cite the math textbook that I used to learn algebra when I write a book on calculus?

No, because I previously said: "...generative AI systems aren't people, so I think it's totally fair to not grant their operators rights by analogy to something a human [mind] would do."

I do not see how a book citing an other book is commercial competing. Citations in most cases enhances the original book in same way that research papers cite each other, which then is used as a measurement of the value of the original paper. Some citations that criticize the original book could be used in direct competition, but that follows a very distinct fair use exception that critique and reviews enjoys.

Could you provide an example of a book citing an other book which as a result competed with the book it cited that would not fall under critique or review?

But the AI model itself does not compete with those works though. You have to use the AI model with additional input and substantial compute resources to potentially generate a work that competes.

You don't read wizardlm.ggml instead of Hobbit.txt.

I think the reproduction of recognized watermarks is hilarious.

I could see some kind of trend emerging as a competition to replace the watermarks as closely as possible to the real thing without it actually being the real thing just for the joke, an argued original work. Kinda like Nathan For You's Dumb Starbucks becoming some viral hit just by changing one word, could see some bored anti-establishment dudes following a similar path.

Reply on news.ycombinator.com