Hacker News Reader

So, if we remove her name and call this style #27b-6, there’s no issue.

She does not own the style. She owns her name and can get fussy about how it’s used, but you can’t stop someone from learning your style.

The relevant part is probably

> ...works of the author, for her own benefit without the author's permission

which did happen, in the training and fine-tuning process.

Training on copyrighted material is legal. As it should be, IMO.

Is that established law (in what jurisdiction?), or just your opinion?

I am pretty sure that it is not established law, but I am pretty sure that that is how it will work out. US provisions for fair use make training models likely OK, and the EU is carving out exemptions for it. See https://valohai.com/blog/copyright-laws-and-machine-learning... for more.

The question of whether the output of the model itself counts as a derivative work, though, is rather more complex. In the case of Github Copilot it has proven very adept at spitting out large chunks of clearly copyrighted code with no warning that it has done so. And lawsuits are being filed over this.

But in the case of the visual artwork, I'm pretty sure that it is going to be ruled not derivative. Because while it is imitative, you cannot produce anything that anyone can say is a copy of X.

But as ML continues to operate, we'll get cases that are ever closer to the arbitrary line we are trying to maintain about what is and is not a copyright violation. And I'm sure that any criteria that the courts try to put down is not going to age well.

In the music industry, even the tiniest sound sample used in a work entitles the original creator to compensation. It might come to pass that using any portion of an author's work in your training data will confer certain rights to the author over the AI generated product. You'll have to perpetually keep records of your training data for commercially available work, lest you be sued. A whole bureaucracy will evolve, a Getty Curated Training Data, public domain training data sets. Basically the same sorts of issues that we've had over the past 30 years, except replacing "internet" with "AI."

And if the past is any guide, the forces of capital will prevail commercially, but after aborted attempts to rein them in with lawsuits, hobbyists and kids on social media will be mostly ignored by rights holders.

That would be silly.

It’s about outputs not inputs. If you create copyrighted material you must compensate.

One day a robot is going to be walking around the world. Will it have to pay someone every time it glimpses a video / book / logo / car design / etc?

> One day a robot is going to be walking around the world. Will it have to pay someone every time it glimpses a video / book / logo / car design / etc?

Looking at all of the people happily deciding to interpret copyright as widely as possible is kinda horrifying in this context: https://www.youtube.com/watch?v=IFe9wiDfb0E

> Your stored mind contains sections from 124,564 copyrighted works. In order to continue remembering these copyrighted works, a licensing fee of $18,000 per month is required.

> Would you like to continue remembering these works?

> [you have insufficient funds to pay this licencing fee]

> Thank you. Please stand by.

> [Copyrighted works are being deleted]

> Welcome to Life. Do you wish to continue?

Yes exactly :)

Somehow I trust the law is going to work out ok though, despite all the hot takes from people who don’t really understand ML or copyright.

It’s scary to see so many people not understanding the distinction between an input and an output.

I’m the EU it’s established law. In the US it’s basically true but it will be playing out in the courts a little in the future.

We don’t know that yet.

Things algorithmically generated from a copyrighted work constitute derived work.

Obviously if the thing is as complex as an AI model it might be hard to prove that a copyrighted work was among the inputs.

This is certainly the legal question of the hour, and it may allow courts to sidestep the equivalence problem to apply different rules to humans than to AIs. History is replete with courts deciding that the rights of some minority don't count because they are deemed different (or less) in some way. Might as well get started with AIs right away. Might as well make sure that when AI is eventually demonstrably sentient there are already tons of established laws and billions of dollars that says they are not.

It seems to me that you cannot take someone's art and plug it into an empty model: we have to start with a model that has been trained on a huge corpus of art, and only then can you show it some pictures and ask it to imitate. This is no different than a human. But my opinion does not matter at all. All it requires is that a jury says "This is an algorithm, therefore it is derivative."

It's a derived work but also transformative and thus almost certainly fair use.

>Things algorithmically generated from a copyrighted work constitute derived work.

This is a very strong statement that is not, at least to me, obviously true. I am curious what your argument is - learning from a copyrighted work does not automatically make a work derivative.

These models do not store copies of images they learned from, or attempt to replicate these images. They learn about constituent parts and assemble them based on the prompt, which is not conceptually all that different from how humans do the same thing.

There are obviously moral questions and legal questions around AI art, and I expect that we will see more, but I'm not sure that this statement is accurate.

The legal right that's been violated is copyright: copies of her work were taken and fed into the ML model. This was done without the picture owner's approval. Moral right may be weak in the USA, but copyright ain't. Most likely making the owners whole is going to be an expensive proposition at some point.

Training on copyrighted material is legal, and rightly so.

Are you sure? Has this been tested in court? As you write: "rightly so", may I ask if you are a judge, lawyer? Has this decision of yours been appealed, gone to a higher court? Does the word "training" as applied to a computing system not simply mean "copying"? Who decides this?

Copyright is still about outputs instead of inputs, just like it is with human learners.

Training isn’t copying. Its an input. It’s akin to “seeing” or “reading”.

Any judge needs to think in terms of outputs. If a system outputs something that violates copyright, there’s a problem.

But attempting to regulate inputs can’t work. For instance, it will be impractical/impossible when we have agents moving around in the real or virtual world, how are they supposed to know when they should turn their sensors off so they don’t see copyrighted material.

Copyright law is the right to control a set of specific kinds of actions. A complete list of which ones is at https://www.copyright.gov/what-is-copyright/. "Inputing the work to a computer program" is not in the list of actions. Whether or not the result of running that program may violate copyright is a question of fact based on what the program does.

As for what it does, my non-lawyerly opinion is that it is no different than a human artist looking at paintings then imitating the style. Which is very much legal.

That said, a case can be made for derivative works. I don't think it is a very good case, but a case can be made for it.

> The legal right that's been violated is copyright: copies of her work were taken and fed into the ML model.

If copies of her work are fed into my brain via my eyes, and I then make an artwork in a style similar to hers (but not close to any particular piece she's created), there is no violation of copyright.

What's different?

The difference is that brains don't get copies of files as input, as you wrote: "via my eyes". Looking is not copying. ML models do get actual copies of files as input.

The legal right that's been violated is copyright: copies of her work were taken and fed into the ML model...

You've stated this very confidently but I see no legal standing for it. Generally copyright infringement concerns itself with the reproduction aspect. Feeding a bunch of images into a model which generates a set of mathematical weights is at worst transformative in nature.

Now whether or not the images produced by this model constitute copyright infringement is a different matter.

As always with generative AI, the legalality is far less interesting than the morality.

The problem with discussing morality is that we each bring our own moral systems to bear, then use emotionally charged words like "right" and "wrong". This leads to people at first disagreeing, and then arguing past each other. Doubly so because you feel like you have made a real point when you agree with yourself, while the other person feels like you haven't. And vice versa.

As a result I only want to discuss morality with people who either largely agree with me, or who are able to take a step back from their own moral system to discuss what someone else from their moral system might think.

By contrast laws and human behavior are external and hence easier for people to agree on. Yes, they are dissatisfying because none of us entirely agree with the law, and laws are deliberately vague in certain things. But I find that discussions of them tend to actually work out better.