Comment by rossy | Hacker News Reader

Comment by rossy | Hacker News Reader

Parent Full thread

There are three things that have become very clear to me after the rise of generative AI:

1. The sheer amount of material on the internet that is "free to view but not free to use for any purpose" is the greatest resource of our time, and despite it being easy for individuals to take advantage of it (no one will take you to court for printing a newspaper comic and pinning it to your corkboard,) it's historically been difficult for corporations to exploit it (their best idea pre-AI is to encourage people to post it on social media walled-gardens where they can surround it with ads.)

2. The reason behind the impressive results of generative AI is because it exploits the above "free" resource, which is the greatest resource of our time. The reason behind the industry-wide push for AI and the insane amount of investment in it, is that they know it's their first real chance to exploit the greatest resource of our time. This is the gold rush.

3. Anthropomorphism is the wool that AI labs are pulling over legislators eyes so they can pull off this heist. If you see training and inference as a black box, a process that consumes a copyrighted work (among others) and produces something very similar to the original work that also competes directly with it, is clearly something that's against the spirit of copyright. But if you (afraid of being judged a luddite) see AI as a little man inside the computer who is "learning" and "creating," how could you deny him? Especially if it would deny your jurisdiction access to the above gold rush. A lot of scientific-sounding AI communication is propaganda for this way of thinking, like the Anthropic J-space stuff, which stops just short of claiming AI is conscious, despite leading the reader to that conclusion.

You don’t need anthroporphism to draw the conclusion that using copyrighted material in model training is fair use. It extends pretty directly from existing fair use doctrine surrounding transformative uses of technology, in particular from Authors Guild v. Google, where the 2nd circuit ruled that creating a searchable database of copyrighted material (i.e. Google Books) was transformative, as long as the material presented to the end user was limited snippets of text and not the entirety. Bartz v. Anthropic explicitly cites that case as precedent

Google: scans all books, people can search for them, find little snippets, see the title and author, buy the book to read it

AI: scans all art, people can ask it to produce art they're searching for, no reference to the original or its author, people no longer pay artists/designers/...

I think there's a very obvious distinction there. The whole purpose of copyright is ensuring the financial viability of creating works. AI slop is a direct market substitute for the originals it ate up for training, so it goes directly against that. Meanwhile, search actually improves the reach of works and, well, snippets and summaries are somewhere in between...

> The whole purpose of copyright is ensuring the financial viability of creating works

In the US it is "To promote the Progress of Science and useful Arts".

Of course, it was also supposed to be a limited-time monopoly rather than perpetual...

In this instance the training data was New Yorker cartoons signed by Brendan Loper, and the material presented to the end user is New Yorker-style cartoons signed by Brendan Loper.

To me that seems a exceedingly broad definition of fair use.

Maybe the reproduction of the signature itself is a copyright violation, I’m not sure what the smallest unit of an existing work you can reproduce legally is, but reproducing a style is in no way a copyright offense. Artists do that all the time

Perhaps we need to create a new form of intellectual property to protect innovations in style. But that’s not copyright, which protects existing works from unauthorized reproduction. It seems like it would be very difficult to adjudicate though; how do you assess whether a style is truly unique to an author or artist?

> Artists do that all the time

This is exactly the sort of anthropomorphism I'm talking about.

Yes, human artists reproduce styles all the time, but implying this equivalence between what a human does when they recreate a style, and what an AI model has done when its output is suspiciously similar to a single artist's "contributions" to the training set is reductive since the mechanism is obviously completely different. It doesn't say anything about whether one ought to be allowed, just because the other is, but the assumption that it should is doing the legwork for those who benefit from AI as a copyright washing machine.

My point in saying that artists do this all the time is to say that there is a very long legal tradition which says that artists doing this does not violate copyright, rather than tediously listing the relevant cases.

My point is not that the process is the same. That’s not relevant to the law. What’s relevant is whether or not the material is a reproduction that substantially infringes upon the original.

I’m also not making an ought argument. If you want to change the law, write your congressman. What I’m arguing is that existing legal doctrine on copyright law very clearly does not protect mere ideas or styles.

As I understand the process is very relevant to the law: why things like clean-room reverse engineering can be necessary.

> What I’m arguing is that existing legal doctrine on copyright law very clearly does not protect mere ideas or styles.

Like it or not, it is anthropomorphizing to claim that an algorithm can have ideas or use styles, rather than just saying that the apparent "ideas" and "styles" in its output are a mathematical derivation of the human ideas and styles in the copyrighted material in its training set.

Reply on news.ycombinator.com