Comment by jrockway | Hacker News Reader

Comment by jrockway | Hacker News Reader

Parent Full thread

I read about this in the Times today (and am surprised that it wasn't on HN already).

My guess is that the court will likely find in the Times favor, because the legal system won't be able to understand how training works and because people are "scared" of AI. To me, reading a book, putting it in some storage system, and then recalling it to form future thoughts is fair use. It's what we all do all the time, and I think that's exactly what training is. I might say something like "I, for one, welcome our new LLM overlords". Am I infringing the copyright of The Simpsons? No.

I am guessing some technicality like a terms-of-use violation of the website (avoidable if you go to the library and type in back issues of the Times), or storing the text between training sessions is what will do OpenAI in here. The legal system has never been particularly comfortable with how computers work; for example, the only reason EULAs work is because you "copy" software when your OS reads the program off of disk into memory (and from memory into cache, and from cache into registers). That would be copyright infringement according to courts, so you have to agree to a license to get that permission.

I think the precedent on copyright law is way off base, granting too much power to authors and too little to user. But because it's so favorable towards "rightsholders", I expect the Times to prevail here.

It's extremely speculative to claim that LLM models are basically doing what humans do. There is very clearly something that isn't right about that because in order for a human to learn to speak and converse and they don't need to imbibe the entire corpus of all written text in human history - which is basically what we're doing with these LLMs. What we're giving them is vast amounts of data which is totally unlike how humans work. There's very clearly some gap here between what a LLM is doing and what a human is doing. So you can't use that as a basis to justify why it's ok for OpenAI to operate like this.

To put it another way, let's say I turn the dial all the way the other way, I train the worlds crappest LLM on NYT material, it massively massively overfits and all it will ever return is verbatim snippets of the NYT. Is that copyright infringement?

The core part of the argument here is actually just that OpenAI doesn't want to adhere to what the current standard is for using copyrighted material, if you want to use it and create something new with it you need to license the material. Since OpenAI's LLM isn't actually like a human it needs to license such a vast dataset that it would be uneconomical to run the business without stealing all the content.

My hard drive can - bit for bit - recall video files. If I serve them to other people on the internet without permission of the copyright holder, that’s called piracy.

Yeah, but the LLMs can't. They aren't big enough to contain every byte of every NYT article, even with the best-known compression algorithms. Rather, they pick up and remember the same patterns that humans do when they write. Authors of the articles also did that, and so the two algorithms (human writer, LLM inference) end up with the same result. (That doesn't preclude large chunks of text that are actually remembered, though. We humans have large chunks of verbatim text floating around in our brains. Passwords, phone numbers, "I pledge allegiance to the flag...", etc.)

Anyway, like I said, I don't think OpenAI will win this. Someone will produce one verbatim article and the court will make OpenAI pay a bunch of money as though every article could be reproduced verbatim, and AI in the US will be set back that many billion dollars. It probably doesn't matter in the long run; it preserves the status quo for as long as the judge is judging and the newspaper exec is newspaper exec-ing. That's all they need. The next generation will have to figure out how to deal with AI-induced job loss... and climate change. Have fun, next generation!

But is it still piracy if you compress them and serve only a likeness of the original?

If 20% of a NYT article is recalled correctly, does that mean I can publish 20% of a movie if surrounded by junk? What if I do that 5 times over?

Yes.

In general, if you perform copyrighted works you are doing copyright infringement. There are certain exceptions (personal use, education, very small fragments with proper attribution, maybe a few others) but whether you are reading it aloud from a book or performing it from memory makes no difference.

So, if you setup a service like ChatGPT but powered by humans responding real time to queries, and these humans would occasionally reproduce large chunks of NYT articles, they and the service itself would be liable for copyright infringement. Even if they were all reproducing these from memory.

Now, this is somewhat different from the discussion of whether training the model on the copyrighted data, even if it had effective protections from returning copies of it, constitutes copyright infringement in itself. I believe this is a somewhat novel legal question and I can think of no direct corollaries.

I certainly don't think we can just handwave and say "at some level, when a human reads a copyrighted work, they are doing the same thing", because we really don't know if that is true. Artifical neural networks certainly have no direct similarity with the neural networks in the brain as far as we can tell. And, even if they did, there is no reason to give a machine the same rights that a human has - certainly not until that machine can prove sentience.

I don't agree that an LLM is doing what we are doing.

"Its what we do all the time" is a major assumption

Reply on news.ycombinator.com