Hacker News Reader

Comment by weinzierl | Hacker News Reader

I find this a super interesting topic: My prediction is that we will eventually see a huge backlash against this copyright whitewashing. I think the current status quo is unacceptable but I fear the upcoming legislation will throw out the baby with the bathwater.

View on HN

One reading of the latest US Supreme Court ruling on the Andy Warhol estate's case is that the current legal frameworks already prevent this.

The ruling focused on what was the commercial use case, not on what was the degree of transformation in Andy Warhol's artwork. The opinion appears to say that as the Prince painting was not explicitly licensed for the commercial use case of reproduction in magazine story and could conceivably have been substituted for the purpose of printing in the magazine to accompany a story by the original artist's photograph, it violated copyright. There was even an explicit distinction drawn between this work and the Campbell Soup paintings which were characterized as a social commentary and therefore fair use.

There are likely class action lawsuits brewing based on this ruling because today when commercial artists in agencies create layouts, they generally license images & artwork.

The way I read that, it was only ruled in favor because they were directed to recreate x photo with some changes. If the Scream painting were still copyrighted, it sounds like telling an AI to "create Edvard Munch's The Scream just with lighter colors" would be a copyright violation - but something like "create a new painting in the style of Edvard Munch" would still fall under "artists can't copyright a style".

The output isn't the only infringement, arguably. You could also argue (and I expect good lawyers will), that the numeric representation of the artist's works inside the model is already an infringing copy. (Just like the JPEG bytes stored on a server, even without them being blitted to a screen.)

Models don't store any artist's works. They are way too small to do that.

Getty has alleged that Stable Diffusion is sometimes returning some of their copyright images[1]. Even if the model seems too small to directly store the images, it seems at least plausible to me that the parameters can act as a compression such that the model could just output an almost direct copy of an original. I have certainly seen stable diffusion emit images which look like a getty watermark has just been blurred out.

[1] https://www.reuters.com/legal/getty-images-lawsuit-says-stab...

It doesn't store the original images, but it has learned how getty images watermark looks like and where it's located, because it has been repeated millions of times. So it sometimes can return that.

This is why it's important to clean up the training dataset. To remove duplicates, images containing watermarks, images that are too similar to each other and so on.

Models don't store any artist's works. They are way too small to do that.

I have close to none knowledge on this subject but I find it very curious and I'd like to know more on that because it seems to me they don't store it but only in the traditional sense. For example, if you could procure a quote, for instance, I asked (chatgpt):

- "In Game of Thrones what did Jon Snow say to Arya when he gave her the sword named 'needle' ?",

- and it answers: "[...] "Stick 'em with the pointy end. [...]"

Then it indicates to me that the information is there. Maybe we should consider that the model actually stores the information but the information is compressed ? Could you ask midjourney to recreate the Mona Lisa of Mickey Mouse ? The right information can't just appear out of thin air. If I recall correctly, someone had some success in identifying and modifying the right neurons or weights of some LLM which changed it "opinion" on where Rome is located ?

This is not precisely true. It's been shown that image models can reproduce certain works almost exactly (up to very minor differences). It takes some effort to find such pieces but they exist.

It takes a lot of effort and the results aren't that great. (low resolution, bad hands) It's way easier to just find the original image and use that instead.

Technically, the DMCA does not limit people from storing copyrighted works, only downloading, distributing and/or reproducing.

Except this is not transformative at all - LLMs are literally trained to be able to output their training data.

That's why it produced those substantially plagiarized CNET articles.

It's more complex than that, they are trained to model their language and understanding of a concept to match the author. There is not enough memory in the network to remember the actual text.

It's like giving a talented an artist a day with a painting, and then a day later ask them to precisely copy the painting from memory. Would they come close enough for it to be considered a forgery, or will it be a transformative reinterpretation? It will probably depend on the skill of the artist, and I feel it might go either way.

Whether a human artist can do this is largely irrelevant to the question of copyright, however. Selling a close copy of another artists’ work (unless they’re long dead) would likely be infringement.

I think the distinction is that you don't arrest someone for merely having the capability of reproducing a work, but after they've reproduced it and attempted to use it in an infringing manner. I don't know how applicable the analogy is given that LLMs don't "think", of course.

you might say that, but the literal benchmark of LLMs (or any supervised learning algorithm for that matter) is loss, or how much 'distance' is between their output, and the validation set, when seeded with the training set. With a loss of well below 1%, which is typical, it means it can pretty much recreate the training data.

I've not seen that number before, what paper is that where there's a loss of less than 1% on the validation/test set?

For people who are unaware of this case, here[1] is a writeup.

TL;DR: Andy Warhol used a photograph of Prince by Lynn Goldsmith to create a series of original silk screen artworks called "Orange Prince". The supreme court ruled (as I understand it) that Warhol's use was not sufficiently transformative to negate the right of the original copyrightholder.

[1] https://www.theguardian.com/artanddesign/2023/may/18/andy-wa...

Edit to add: Here’s a side-by-side https://www.mondaq.com/images/article_images/1324682a.jpg . I mean extreme right second from the bottom he hasn’t even changed the colours.

Do you think people should be allowed to learn from books we read, music we hear, etc?

If we’re extending copyright to be not only about reproduction of the work, but also providing part of a large knowledge base for future creators to build upon… why does it matter if it’s a neural network or human being doing the learning?

I personally hope the copyright maximalists fail, because if they succeed in making “learning from a copyrighted work” without paying illegal, we end up in a dark place indeed.

I think AI aren’t people, and that they’re not even AI: it’s machine learning, and ML is a subset of AI that we haven’t really broken out of. ML is a tool, and asking whether ML should be allowed to learn on copyrighted work is equivalent to asking whether your computer is allowed to contain a copyrighted work: “it depends.”

I’m in favor of getting rid of copyright entirely, since it’s become a far cry from what it was intended to be, but we have to acknowledge both sides of the arguments in order to proceed.

    I’m in favor of getting rid of copyright entirely

This is extreme to me. What is your proposed replacement?

https://en.wikipedia.org/wiki/Copyright_alternatives

I agree we should either get rid of, or very least overhaul copyright.

However, the line in the sand I draw is giving any kind of human-based rights to ML or AI. People already are thinking ChatGPT is close to human intelligence, and we know we haven’t even seen what’s truly possible.

My fear is the typical oversimplification of complex technology will lend us to giving human rights to programs, and lead to a whole host of issues.

We need to stop anthropomorphizing technology until we really hit a philosophical quandary. Giving in too early will just give power to less-than-polished tech and either slow down research, or further drain a country’s resources

> the sand I draw is giving any kind of human-based rights to ML or AI.

Ok, but this isn't about giving rights to AI.

Instead this is about given rights to humans to use AI.

If you make a law about this stuff, at the end of the day, it all trickles down to allowing or disallow human actions.

The rules are already different for humans and machines. If I memorise a copyrighted book then that’s not copyright infringement, but if I store a copy on a hard drive then that is. You ask why distinguishing between humans and machines matters and then answer your own question: “we end up in a dark place”.

IANAL, but as I understand it, the difference is because copyright is about the transmission of the work.

If you memorize a work, you have used the work in accordance with copyright law: you were authorized to read the work. If you write out a copy of the work you memorized and give it to someone, you've broken copyright law.

If you store a copy of a work you were authorized to have, you have not broken copyright. However, if you download a book that you have no authorization for, that is a violation of copyright.

In other words, it's not about where the work is "stored" or whether or not humans were involved, it's about the transmission of that work between parties.

Actually, if you memorize that book and then type it in to your own computer, you're fine. So maybe a bit of a rework on that line of argument.

It still seems odd to me that the standard advice for "how do I avoid copyright issues with example code" is "rewrite it in your own style", and we're happy with this, but now that we have an algorithm that can do this (using a stupendous amount of calculation, it's not like these things learn for free... yet...) this is suddenly somehow cheating.

I think the protestations about copyright aren't entirely honest. People aren't mad that LLMs are "stealing their work", they're scared that LLMs are copying their ability to produce such work.

Its some of both. Stable Diffusion will rarely blatently copy a work, as it has no distinction between creative reconstruction and plagiarism. I think this applies to LLMs too, its just much harder for human brains to detect.

At the same time, the artists and such who claim any of its output is illegal are being hypocritical. Nothing is made in a vacuum.

In terms of actions, the US has gone as far as to kneecap a competitor's chip industry and restrict their GPU access.

I reckon that behind closed doors there will be a strong sense that any kind of regulation that impedes AI development would not be in the national interest.

I would argue if the fair use argument holds up - that the most substantial value of these models lies in their training dataset.

It would be morally reprehensible for anyone to block training open models based on the data provided by the public at large. Such legalization would be totally unacceptable.

Why would that be morally reprehensible?

Is it not as "morally reprehensible" that I have to restrict my use of the internet now in order to avoid having my works used in that way? That reduces the amount of information available to everybody, after all.

What is “copyright whitewashing”? I’ve never heard that phrase before.

they mean laundering IP theft

Exactly, laundering was what I meant, I mistranslated from German.

Imagine a machine that takes in copyrighted works and, with the press of a button, outputs works which are substantially similar but not copyrighted.

(I.e. pointing to your black box and claiming it's a clean room)

But to reiterate existing talking points, using copyrighted works for AI weight generation is exactly the same thing as how human artists learn skills just because some people refer to the process as “learning” or “training”, and any counterarguments to this are meaningless fearmongering outdated Luddite rages, besides such uses are fair uses and transformative uses and explicitly allowed in some jurisdictions under specific terms not allowing blanket rights to regurgitation so it’s all lawful and financially fully exploitable.

forwardslash sierra.

So, like Terry Brooks then?

> which are substantially similar but not copyrighted.

if it's considered substantially similar, then why is it not copyrighted? The difference must also be substantial. Therefore, it should be acceptable as a new works, rather than a derivative works.

Actually, I only realize now, that it is a mistranslation from my mother tongue. What I meant to say is probably better expressed by 'content laundering' analogous to 'money laundering'.

Or maybe, maybe the people in charge will wake up, do their job, accept the challenge and make modern cc laws? Oh forget about it we will forever live in the 90'

> I find this a super interesting topic: My prediction is that we will eventually see a huge backlash against this copyright whitewashing. I think the current status quo is unacceptable but I fear the upcoming legislation will throw out the baby with the bathwater.

Worse, they have the potential to push development toward countries that do not have such laws.

Reply on news.ycombinator.com