Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

NY Times copyright suit wants OpenAI to delete all GPT instances

arstechnica.com

Post reply on HN
Print view View original

561–570 of 921 posts

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#561

Post by cycrutchfield »

Vegenoid wrote: post #499
cycrutchfield wrote: post #56

Earlier quoted context omitted.

So presumably when they fix that issue (which, if the text matches exactly, should be trivially easy) then would you accept that as a sufficient remedy?

Copyright infringement is not avoided by changing some text so it isn’t an exact clone of the source. Determining whether a work violates a copyright requires holistic consideration of the similarity of the work to the copyrighted material, the purpose of the work, and the work’s impact on the copyright holder. There is not an algorithm for this, cases are decided on by people. There are algorithms that could detect…

And you think that it would be impossible to train a model to avoid outputs that are substantially similar to training data?

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#562

Post by dwringer »

mattdesl wrote: post #505
bonzini wrote: post #468

Earlier quoted context omitted.

Yes, the screenshot in the article is clearly doing an Internet search. The exhibit in the lawsuit shows that you can complete an article by using GPT on the first sentence of the prompt, with low temperature to aid reproducibility, and obtaining the original except for a single word. That is another thing, and it shows that the LLM has basically recorded the original text into its weights in compressed form: https:/…

It would be curious to test this on a larger sample than just a few. It is hard to believe that a majority of NYT articles are verbatim stored in the weights of a web-wide LLM, but if that is the case it would be a pretty unbelievable revelation about their ability to compress an entire web’s worth of data. But, more likely, I assume it is a case of overfitting, or simply finding a prompt that happened to work well.…

I have attempted this sort of thing with GPT 3.5 many times and never been successful, although I've still never been taken off of the GPT4 waiting list that I signed up for months ago and I'm not going to subscribe without trying it first. I [and presumably many thousands of others] have tried things like this with many LLMs and image generating models, but to my knowledge we've come up rather short. I've never managed to recreate anything verbatim and have struggled to get anything resembling a copyright infringement out of stable diffusion with the sole exception of a meme image of Willy Wonka.

That said, the meme image of Willy Wonka comes out of stable diffusion 1.5 almost perfectly with surprising frequency. Then again, this is probably because it appeared hundreds or thousands of times in the training set in all sorts of contexts because it's such a popular meme. There is a tension between its status as an integral part of language and its nature as a copyrighted screen grab.

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#563

Post by vel0city »

Matticus_Rex wrote: post #521
eigenket wrote: post #494

Earlier quoted context omitted.

If it can and does reproduce a piece of text verbatim then the text is indisputably stored somehow in the model.

That's just not true. There's no search and retrieval involved. It just associates the words so strongly in that context because they were in the training data so often that next-token prediction can (sometimes, in some limited circumstances) reproduce chunks of it. It's like if a human had read pieces of an article so many times and knew NYT style so well that they could spit out chunks of an article verbatim, but u…

So it stores the words, and it stores the links between those words...

but somehow storing the words and their links is not storing the actual text? What is text but words and their links?

If I had a database of a billion words, and I had a list of pointers to words in a particular order, and following that list of pointers reproduces a copyright text exactly, isn't the list of pointers + the database of words just an obfuscated recreation of that copyrighted work?

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#564

Post by kromem »

mihaic wrote: post #476
logicchains wrote: post #204

Earlier quoted context omitted.

> it would be as if I would copy parts of other propriety code and copy paste it into my own codebase. It's not copy-pasted; it's compressed in a lossy manner. Even GPT4 has nowhere near enough memory to store the entirety of its training data in a non-lossy compression format. Just likes how humans compress the information we read.

If you have a copyrighted photo that I simply put through jpeg compression, am I legally allowed to use that? Software programs are not humans, and need to be treated differently. Anthropomorphization is one of the slipperiest paths to argue anything.

It depends on how much is reproducible and what the use is.

If only small patches of the original image can be reproduced then it becomes much more murky.

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#565

Post by octacat »

jalapenos wrote: post #517
octacat wrote: post #387

Earlier quoted context omitted.

So, if someone decompiles a program and compiles it again, it would look different. "It is not copying", we just did some data laundering. Feeding someone else data into your system is usually a violation of copyright. Even if you have a very "smart" system, trying to transform and obfuscate the original data.

I'm regularly feeding other people's data into my "system" (brain) in order to produce my outputs. So I'm a living breathing copyright violator. As a person I should be banned. Fortunately, copyright is a bullshit fictitious right with no basis in natural law. So I don't lose much sleep over it.

Computers are deterministic. Giving the same inputs training would produce the same model. The comparison with brain is incorrect. You could add noise on input data during the training - it would more of less reproduce the real learning. Still, it could produce less useable models as a result.

The court could ask to show the training dataset.

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#566

Post by newyorkgritty »

cesarb wrote: post #352
munchinator wrote: post #180

It's interesting to me the ambiguous attitude people have to reproducing news content. Whenever there is a story from NYT on HN (or any other large media outlet), the top comment is almost always a link to an archived version which reproduces the text verbatim. And this seems to be tolerated as the norm. And yet, whenever there is a submission about a book, a TV show, a movie, a video game, an album, a comic book, or…

> Whenever there is a story from NYT on HN (or any other large media outlet), the top comment is almost always a link to an archived version which reproduces the text verbatim. [...] And yet, whenever there is a submission about a book, a TV show, a movie, a video game, an album, a comic book, or any other form of IP, it is in fact very much _not_ the norm for the top-rated comment to be a Pirate Bay link. If the sto…

Much of this is incorrect

> the archive link has the exact same content as the link submitted

No, articles are updated as new information comes in, retractions are made, etc. Especially breaking news (the type that would reach the top of HN). The archived versions are outdated.

> others randomly got a login request

It's not random, you get a number of free articles before the paywall appears ("soft" paywall).

The paywall is removed entirely for some topics/stories, especially matters of public health (common during the pandemic).

> the only reason the archive site has the content is that it didn't get the login screen

No, it's because they don't block archive crawlers, and prefer people bypassing the paywall and reading news at NYT. Hopefully users find the content valuable, and some of them subscribe as a result.

(opinions are my own)

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#567

Post by bonzini »

mattdesl wrote: post #505
bonzini wrote: post #468

Earlier quoted context omitted.

Yes, the screenshot in the article is clearly doing an Internet search. The exhibit in the lawsuit shows that you can complete an article by using GPT on the first sentence of the prompt, with low temperature to aid reproducibility, and obtaining the original except for a single word. That is another thing, and it shows that the LLM has basically recorded the original text into its weights in compressed form: https:/…

It would be curious to test this on a larger sample than just a few. It is hard to believe that a majority of NYT articles are verbatim stored in the weights of a web-wide LLM, but if that is the case it would be a pretty unbelievable revelation about their ability to compress an entire web’s worth of data. But, more likely, I assume it is a case of overfitting, or simply finding a prompt that happened to work well.…

You can't reproduce on the web interface, because the temperature settings are higher than what's required to compress the text. You need to use the API.

However, I had good luck reproducing poems on GPT 3.5, both copyrighted and not copyrighted, because the choice of words is a lot more "specific" so to speak, and therefore higher temperature isn't enough to prevent complete reproduction of the originals. See https://chat.openai.com/share/f6dbfb78-7c55-4d89-a92e-f4da23... (Italian; the second example is entirely hallucinated even though a poem with that title exists, while the first and third are recalled perfectly).

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#568

Post by kromem »

makeitdouble wrote: post #360
logicchains wrote: post #115

NYT's perspective is going to look so stupid in future when we put LLMs into mechanical bodies with the ability to interact with the physical world, and to learn/update their weights live. It would make it completely illegal for such a robot to read/watch/listen to any copyrighted material; no watching TV, no reading library books, no browsing the internet, because in doing so it could memorise some copyrighted conte…

Memorising isn't the issue. It's providing it back verbatim and/or cutting access to the source. You'd get the same problem with someone with a photographic memory who a group of people would turn to recite them the news instead of buying the newspaper. As of now public performance of copyrighted material is infringement.

That's not the case, as they aren't trying to get a ruling on the forced reproduction by prompt as infringement, but rather to get a ruling that training is infringement.

I fully agree with the perspective that infringement in usage needs to be limited even if I strongly disagree that training is infringement.

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#569

Post by cwmma »

munchinator wrote: post #180

It's interesting to me the ambiguous attitude people have to reproducing news content. Whenever there is a story from NYT on HN (or any other large media outlet), the top comment is almost always a link to an archived version which reproduces the text verbatim. And this seems to be tolerated as the norm. And yet, whenever there is a submission about a book, a TV show, a movie, a video game, an album, a comic book, or…

I think one of the key differences is something pointed out in the article, in that what the Open AI is doing is a substitute for reading the new york times and possibly a rival to it.

On the other hand having an archive link to a times article in order to discus it is not really a substitute for a times subscription as a news paper has to walk a line of letting some of it's articles be read while requiring payment for others (the times actually allows you to create a "gift link" to do exactly what the archive links do).

Re: NY Times copyright suit wants OpenAI to delete all GPT instances

#570

Post by pcmaffey »

billywhizz wrote: post #414
munchinator wrote: post #180

It's interesting to me the ambiguous attitude people have to reproducing news content. Whenever there is a story from NYT on HN (or any other large media outlet), the top comment is almost always a link to an archived version which reproduces the text verbatim. And this seems to be tolerated as the norm. And yet, whenever there is a submission about a book, a TV show, a movie, a video game, an album, a comic book, or…

there's quite a big difference between "pirating" digital content and making it available to anyone for free and taking that content and building a for-profit service on top of it, which is what OpenAI are doing, no?

I was just going to post this. Seems quite an obvious and significant distinction, that doesn’t need to provoke all the existential hand wringing. Making money off someone else’s content is a totally different moral and legal case.
Post reply on HN