Hacker News Reader

> If they were humans, a court could deem the outputs copyright infringement

I'm not sure I understand how this is self-evident. The closest equivalent I can see would be a human who looks at many pieces of art to understand:

- What is art and what is just scribbles or splatter?

- What is good and what isn't?

- What different styles are possible?

Then the human goes and creates their own piece.

It turns out, the legal solution is to evaluate each piece individually rather than the process. And, within that, the court has settled on "if it looks like a duck and it quacks like a duck..." which is where the subconscious copying presumably comes in.

I don't know where courts will go. The new challenge is AI can generate "potentially infringing" work at a much higher rate than humans, but that's really about it. I'd be surprised if it gets treated materially different than human-created works.

> The new challenge is AI can generate "potentially infringing" work at a much higher rate than humans, but that's really about it

The other challenges are: (i) the model isn't a human that can defend themself by explaining their creative process, it's a literal mathematical transformation of the inputs including the copyrighted work. (And I'm not sure "actually the human brain is just computation" defences offered by lawyers are ever likely to prevail in court, because if they do that opens much bigger cans of worms in virtually every legal field...) (ii) the representatives of OpenAI Inc who do have to explain themselves are going to have to talk about their approach to licenses for use of the material (which in this case appears to have been to disregard them altogether). That could be a serious issue for them even if the court agrees with the general principle that diffusion models or GANs are not plagiarism.

And possibly also (iii) the AI has ridiculous failure modes like implementing the Getty watermark which makes the model look far more closely derived from its source data than it actually is

It's worth pointing out that the problem, in this scenario, is for the creator (ie, the human running the algorithm). They will need to determine whether a piece might violate copyright before using it or selling it. That seems like a very hard problem, and could be the justification for more [new] blanket rules on the AI process.

Proving artwork you created is free from all copyright issues is similarly impossible, but in practice isn’t an issue. So, I don’t see any AI specific justification being relevant.

How common is it for an artist to accidentally generate a work that resembles an existing work?

I can't speak for digital art because I'm not an artist but I can say it's extremely common for original music to (often accidentally/subconsciously) include melodies or pieces of melodies from other music.

That's more like two paintings sharing the same color scheme or using the same brand+color of paint but still being unique. The performance of those melodies and structures is what makes a song unique and creative.

Adam Neely the YouTuber has produced several videos about recent legal cases where musicians sue because of some superficial similarities. In most cases the higher courts in the US recognize that copying is part of the normal creative process and is not infringement unless it is a blatant rip-off.

This is exactly my thinking. If the court finds somebody guilty of infringing on a human-made piece of digital art the response is to punish the human, not to ban or impose limits on photoshop.

At risk of stretching the analogy, you don’t charge the gun with murder…

Except, of course, the human has very little control over what the AI outputs in TXT2TXT scenarios, at least in terms of whether the output would match the definition of copyright infringement of someone else's work. IMG2TXT is kinda different -- I think you could make a much stronger case for derivative work there. So you have a tool that can randomly create massive liability for you, and you can't know if its done so or not until someone sues you.

The humans are the cause of it happening in the first place. Maybe the gun analogy is not such a stretch: if you pull the trigger, you own the consequences.

Reasonable fair use principles could distinguish personal and R&D use from commercial use.

I think it's more common with music. Some musician goes to a foreign country and heard an obscure local song. 20 years later the musician has forgotten completely about the song and the trip. One day a catchy melody appears in the head of the musician out of the blue, and the musician complete the song and add a lyric. The song get famous, and later reach the foreign country, and everyone acuse the musician of plagiarism.

Reminds me of a recent scandal involving Adele, where she is accused of plagiarizing a Brazilian composer: https://english.elpais.com/usa/2021-10-19/toninho-geraes-vs-...

>I'm not sure I understand how this is self-evident. The closest equivalent I can see would be a human who looks at many pieces of art

...and then gets told "Hey, go and paint me a copy of that Andy Warhol piece from memory".

The model might not violate the copyright, but its output is derivative work if the copyrighted works are included in the training set.

That would be over-fitting which is certainly a failure mode of ML. But it's still a failure mode, not an inherent property.

Note that a derivative work doesn't instantly make it 'fair use' for the purposes of copyright. You typically still need 'adaptation' permission from the copyright holder to made a derivative work of it, so you can't make 'Breaking Bad: The Musical' by recreating major scenes in a play format, at least not without substantially changing it[0].

For the purpose of fair use, copyright.gov has an informative section titled "About Fair Use" which details what sort of modifications and usage of a copyrighted work would be legal without any permission from the copyright holder https://www.copyright.gov/fair-use/#:~:text=a)(3).-,About%20...

0: https://en.wikipedia.org/wiki/Say_My_Name!_(Musical)

It’s the opposite of what you’re suggesting in the first sentence. Something being a derivative work is a fairly clear sign that it’s not fair use.

It's kind of a combination of these two positions: something being a derivative work means that it requires an excuse under copyright law, whether that be a license, or a fair use defense (which may be fact-specific), or some other justification. Without such a justification, the derivative work is treated as a copyright infringement.

This is a result of 17 USC §106(2), which says that one of the things that "the owner of a copyright [...] has the exclusive rights to do and to authorize" is "to prepare derivative works based upon the copyrighted work", unless an exception (including fair use) applies.

https://www.law.cornell.edu/uscode/text/17/106

> If they were humans, a court could deem the outputs copyright infringement

Why? I am a human, I can learn the composition of stock photos and use the idea in my work. There is no copyright infringement, or there is simply no way to prove it.

The connection needs to be really strong to consider something as derivative work.

> consider the case of translating a literary work between languages: this is a classic example of a derivative work

This is a classic example of "you need a license on the original material to sell it" (unless it is public domain)

You need the rights also if you are translating from a pre existing translation in a different language (i.e. you're translating a Japanese book from the French authorized translation).

Translating an opera is not automatically derivative work and does not fall under the umbrella of fair use.

source: my sister is a professional translator

EDIT: technically you'd need to acquire the rights even if you translate it by yourself, for fun, and show it to someone else.

> The challenge with these models is that they’ve clearly been trained on (exposed to) copyrighted material, and can also demonstrably reproduce elements of copyrighted works on demand. If they were humans, a court could deem the outputs copyright infringement, perhaps invoking the subconscious copying doctrine (https://www.americanbar.org/groups/intellectual_property_law...).

Every single human has been exposed to copyrighted material, and probably can reproduce fragments of copyrighted material on demand. Nobody ever writes a book or paints a picture without reading a lot of books and looking at a lot of paintings first. For a "subconscious copying" suit to apply, you need to demonstrate "probative similarity" - that is, similarity to copyrighted material that is unlikely to be coincidental.

In other words - it's not clear to me that the situation with AI is any different than with a human, or that it presents new legal challenges. If it looks new, it is new.

> Blurring a work but incorporating it would nonetheless still be derivative, I think.

you have to show that the derivative work is a substantial part of the new work.

If your background is a small portion of the new image, and the blue makes it difficult to see that it was the original, and that any other blurred image would've done the same job, then i would argue that the final new painting does not constitute a derivative work.

A similar argument could be made for AI models. The model consists of billions of training images. None of these images are individually substantial in the final output, despite that on some part of the output, you can trace a derivative work. For example, if an author used 1 million books, and copied every nth word from each book and merged it, to produce a final book (which happens to produce a coherent book), i would argue that the author did not infringe copyright on any of the original 1million source books.

>> Translating a work would absolutely be derivative (consider the case of translating a literary work between languages: this is a classic example of a derivative work)

True for literature, which can be copyrighted. Not true for logic. Code that performs the same function but is written independently can't, by definition, violate copyright. It could violate a patent, but it's much harder to patent code.

However, in the AI case, the models themselves do nothing even approaching such translation. They fully incorporate what are essentially just compressed representations of existing works into their weights.

I'd guess if any single copyright holder could show that their work could be substantially retrieved from the model with the right query string, they'd have a reasonable claim on royalties from the entire model, or else the whole model would have to be thrown out.

Yes, I expect that if you ask the model for "Getty images photo of [famous person] doing [thing Getty Images has only one photo of that person doing]" you might well get the original photo out.

Should be easy to try. My guess is that it most likely won't.

You won't get the exact image, because the model doesn't perfectly memorize all inputs, but you'll likely get something so close that everyone would consider it a derivative work.

This is very similar to getting Copilot to spit out Carnack's fast inverse square root code with the right prompt.

There's some interesting thought experiments around this:

1. You do the same thing but don't use "getty" in the prompt.

2. You do the same thing with "getty" in the prompt on a model NOT trained with Getty images

3. You ask a human photographer to do the same thing and you show him a Getty image

4. You ask a human photographer to do the same thing but without showing him a Getty image (but he's presumably seen many in the past).

If 1 and 2 produce images that are also similar to a Getty image, where do we stand? I imagine it's likely that a trained model can learn "getty-like" without any actual Getty images to make 2 happen.

And currently IP law treats humans and AIs the same (in the sense that infringment rules don't distinguish between them) so would you consider 3 and 4 to be similar to 1 and 2?

For 1, there's probably a case where a picture that exactly matches the prompt is very famous, so maybe the right prompt would manage to effectively select an image licensed by Getty. The GPLed fast inverse square root routine was an example of this kind of thing: one good match and the model finds it. Finding such a case might or might not be possible.

For 2, you can't find what isn't there, so something random would come out.

For 3 or 4, I suppose you could pay paparazzi to stalk the celebrity and try to produce a similar original shot, and this might be impossible depending on the prompt (for example, photo of [person] at [event] on [date]"), but if it's possible it has absolutely no resemblance to 1 or 2. The photographer would produce an original work, unless your #3 contractor tries to remove the watermark and pass off the Getty image as their own, which would be idiotic.

There is no particular "Getty Images" style, other than their quality requirements, they are a huge company that acquires and licenses a ton of pro photography. There's no such thing as "getty-like".

So, no, only option 1 might possibly produce a problematic derivative work.

> For 2, you can't find what isn't there, so something random would come out.

It's not a search engine. It can synthesise things that it hasn't seen to some degree (assuming it knows the elements that make up the request.

The tricky bit would to teach it "Gettyness" without showing it Getty images but I think that's entirely possible. Getty images aren't astonishing examples of unprecedented originality. So it just needs to know a) the celebrity b) the action and c) what people mean when they ask for something that looks like a Getty image.

EDIT - I answered in a rush and realise you made a similar point to me about "Gettyness" - which makes it even harder for me to understand why you think it would be a violation.

How many different ways can George Clooney eat a burrito in Times Square?

Suppose that there's a really iconic pic of George Clooney eating a burrito in Times Square, with a "Getty Images" watermark on it, in the training set. It's perfectly framed, has a somewhat comic expression, and a bit of the contents spilled on his shirt. It seems quite possible that the model could produce an image that isn't identical to this but so much so that people assume it's a copy, maybe produced by an artist based on the picture. Worse, it will have most of the "Getty Images" watermark right on the image.

If that doesn't convince you that there might be an issue I'll just stop here.