Comment by karaterobot | Hacker News Reader

Comment by karaterobot | Hacker News Reader

I'm the kind of creative whose job is threatened by AI, but if I'm being honest and cold-blooded, I'm not actually sure that I'd call what AI training does stealing—apart from any illicit acquisition of the data, that is. I'm open to being persuaded. I just have not read a really persuasive argument that when a model is trained, what it's doing is closer to stealing than it is to being influenced and then recapitulating the general form of those influences.

The latter is something we have no problem with in our culture. Every generation of artists do this. Every artist does this, in fact. It may be how creativity itself works, for all I know. But we never call it stealing, or when we do we say things like "great artists steal" and acknowledge it as a process rather than a problem.

It feels to me like when people make the "AI is stealing" argument, they're really arguing about something else and using this as a proxy for it. Not exactly sure what that other thing is. Fear of losing work and being displaced is the obvious one, and it's a valid fear. I share it! But, if that's what it is, I think we should be having that discussion directly, rather than talking around it. We always do the thing where we piss around talking about proxies, rather than going right to the agon.

AIs are not creative, that's anthropomorphism. They use lossy compression to save training data. Later on, the compressed data is used to generate things that are potentially copyright violations (because they are very close to the original).

I don’t understand, generating (and then monetizing) things that are copyright violations is already illegal.

What we’re taking about is suing the tool that could potentially be used to generate copyright violations. Which seems silly?

Steelmanning your argument a bit more: OpenAI is selling the fact that you could potentially generate copyrightable material as a subscription, because the artwork is somewhere in the compressed data?

Abstracting the underlying tech a bit: If we imagined a printer that could faithfully represent whatever you told it to and you told it to reproduce Da Vinci’s that you later sold as originals while others told it to reproduce original forms of art (successfully), who’s at fault here, you or the printer?

Seems obvious to me it’s the user of the printer, not the printer itself, even if the printermaker is making money.

In a strict reading, the tool itself is a copyright violation. In that an LLM is the code and its weights, and the weights contain a lossy compressed copy of a copyrighted work.

Now you might say that's an overly rigid reading of copyright law, and I might agree, but that is nevertheless the reading of it applied to other domains - music, film and so on - where incorporating even a fractional, lossy and heavily processed portion of another work is enough to trigger copyright law. There is no blanket de-minimis that says "well, it's only 5%, so it's fair use".

The legality of Copyright material being present in the latent space of a commercialized foundation model would definitely be an interesting case!

Though most of these cases aren't litigating this. They are charging that the ability to recreate copyrightable material is the fault of the tool and not the tool user[1].

[1]: https://www.darrow.ai/resources/ai-copyright-law

If using the work as training data was lawful (which hasn't been always the case as we know today) I tend to agree with you, but it might depend: creating an alteration of the work might be prohibited altogether or you might need to give proper credit to the original authors and that might already be violated when you present certain output to a user. In general, the user is responsible to obey the license of the original work. The problem is that the LLM doed not know on which works its output (mostly) is based on. So in theory, no output for which copyright laws might apply should be used whatsoever.

It's not cut and dried in that way ("certain output"), in many models, any one piece of training data can influence any number of weights and potential outputs, and it's difficult or impossible to say which output was influenced to what extent by which input.

A drop of blood falls in a lake.. which bit of lake now contains blood?

> the user is responsible to obey the license of the original work

The user is in this case the "AI" company.

They already violated law almost certainly by copying stuff from the net in the first place. (Almost certainly as it by now does not look like this was fair use, and there is no other means to make legal what they did.)

Copyright law has never clearly and adequately defined the boundary between copying and measuring.

If I measure the length of a song? Not copyright infringement. If I take even a very short sequence of PCM samples, copyright infringement.

But the practical thresholds appear to be different, there's a bunch of ways you can extract kbs' worth of information from a song, some are considered copying (ripping the melody to midi with a converter, say), some are not.

And in the case of "AI", depending on the model it could be extracting and encoding almost anything, because it's a catch-all term encompassing dozens of architectures and methods.

That's not true.

Some people (like the NYT¹) are in fact demanding before curt the destruction of the trained models as they obviously contain copyrighted content.

¹ https://pressgazette.co.uk/media_law/new-york-times-open-ai-...

> OpenAI is selling the fact that you could potentially generate copyrightable material as a subscription, because the artwork is somewhere in the compressed data?

Yeah, that's the express and sole purpose of the tool. You use OpenAI image generation because you want something that looks like a stock image for your website. But you don't want to pay for a stock image because you're cheap. So you'll get a close enough approximation for free. Or, uh, less.

"express and sole purpose of the tool"

Categorically untrue. 100% of all my usage of image generation is for artistic purposes that have nothing to do with the reproduction of or even the mimicry of existing art.

This is a gaping hole in the legal case if the lawyers need this to be true.

> 100% of all my usage of image generation is for artistic purposes that have nothing to do with the reproduction of or even the mimicry of existing art.

How? You're creating artwork that is based off of the training data and is meant to replicate a subset of it.

I think, just because you cannot pinpoint the exact art piece(s) that you are recreating doesn't mean that that isn't what you're doing.

Sorry when you said "Sole purpose", I assumed you meant the ChatGPT USERS sole purpose and OpenAI's expectation of their users' Sole purpose. My sole purpose as a ChatGPT user is not to create/recreate stock photos.

"just because you cannot pinpoint the exact art piece(s)"

Precisely, and neither can the owner of the copyright. Unless of course I rub into classical copyright issues ("these two art pieces look too similar") which goes back to my original question, why isn't just enforcing existing copyright law on bad ChatGPT users enough?

If I "use" your artwork, and generate another artwork that looks nothing like yours (either from my brain or by automating it) why do you care?

> If I "use" your artwork, and generate another artwork that looks nothing like yours (either from my brain or by automating it) why do you care?

I wouldn't, but presumably the person who you are immitating would.

The fallacy here is that you think an LLM is taking one or two art pieces and then turning around and creating something brand new, like a human brain might do.

No. It's taking billions of art pieces and then sampling them at various frequencies to create your art work. We can find your art work somewhere in the training data, it's just hard.

In essence, we're smuggling the copyright violation through convolution. Since these models are truly very large, we can pretend that there isn't copyright violation. But there is, we just can't really find it because LLMs cannot work backwards.

" taking one or two art pieces and then turning around and creating something brand new, like a human brain might do."

I don't think the human brain is doing this at all (except maybe if you're writing a satire or homage). Your writing is the sum total of all your experiences which includes everything you've ever read, a number that might (depending on your age) at least approach millions of tokens.

And just like an LLM you'd find it difficult to piece together your influences.

The difference is I'm a person - I get special privileges. A computer program does not, we don't need to cater to them because, well, they don't have feelings. Or rights.

Ultimately these are tools, slaves to humans, meant for one thing only - to serve us. If they're not serving us, they should be eradicated. Because that's what we do.

Stealing from humans, no matter how convoluted, is bad. We shouldn't do that, and luckily that's completely optional and we don't have to stand for it. The computer programs can't fight back, it's okay.

Don't care if a computer program has special privileges or not, that's irrelevant. It's a tool, and it has the potential to make a whole host of processes and people's lives easier, and no one, not you, not an internet mob, not a council of elders has the right to preemptively stop those real-world tests based on vibes.

So getting to the bottom of what is actually stealing or not will be important (we can't just unilaterally declare it so), and AFAICT if a user misuses it to regenerate copyrighted material that particular user is at fault, not the tool itself.

People don’t like honesty because they want to be special snowflakes, but sorry, humans aren’t that creative. The entire copyright and patent system is predicated on that fact, that you can only make money by making a thing exclusive and proprietary. That is directly caused by the fact that not only is it NOT original, but that others absolutely will invent/create the same thing eventually even without exposure to the original material, which is why people want to protect their exclusivity.

“I’m original because I add the little curls to the lines, make the lines softer, and make faces pointy rather than round” isn’t originality, nor is any of the art that people claim copyright for.

Additionally, it is a fundamental misunderstanding of AI modes if you think the models are simply lossy compression of data. That is like saying the alphabet is just lossy compression of The Odyssey.

But why should things work differently for "AI"?

One could of course say that "IP" ("intellectual property") is an ill concept all in all and abolish it. But than without exceptions! If it's not a valid concept in regard to "AI" it's not a valid concept in general. But if we want to keep that concept "AI" needs to bend to the same rules that apply to anybody else. Simple as that. Law is universal. (At least it should be like that everybody is equal before the law.)

It feels to me like you are the one arguing about something else. It doesn't matter if AI doesn't copy work wholesale or it can be viewed as how artists are inspired.

AI companies use work that they do not have the rights for, to construct their products.

You don't need rights to view a work and caption it. You can walk into a library or gallery today and take your own notes on everything in there.

But this is very obviously not what AI is doing. In essence, an LLM is taking works and compressing them and storing them in a big database, to later be reproduced with a touch of non-determinism.

It's closer to archival and compression than captioning. But they're selling it. That's a big problem - you can't just take someone else's work, sprinkle some god knows what on it, throw it in a big pot with other shit you stole, and then reproduce it later. Even if your reproduction is super convoluted and not 100% accurate.

Reply on news.ycombinator.com