Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I personally believe that both training and generating images through AI is fair use. It's transformative. You're creating completely new ideas. Consider that plus the fact that styles aren't copyrightable.

Also keep in mind that, at least in the US, you can't legislate your way out of fair use since it's a constitutional right. (It's the result of the tension that exists between the copyright clause and the first amendment)



The output is transformative. But in order to use it as an input to a training set, you're effectively making a copy that may or may not be authorized, not to mention at some level the intermediate representation might well present a copyright issue. "Style" isn't the issue.

I think on an ethical level we're just going to need the expectation of separate agreements / legal conceptions for having ones work go into a training set. It's novel territory, providing not only work-for-hire but meta-work-for-hire. Might even be that there's no true compensation without equity.


I can only potentially see an issue with that if you're selling the model which is essentially a collection of copyrighted works (organized in a new way).

However, the output of that model is the epitome of derivative work.

The trained model has to be viewed as an artistic tool used by the artist to create art, imo.


Derivative work isn't permitted under current copyright though, you know that?

Like I can't release a sequel to a Star Wars movie without getting a license to do so by Disney, because they own the original work's copyright, I'm not allowed to produce a derivative of it.

Similarly this is why there needs to be an explicit open source license that allows you to produce derivative work from it.

That's the crux of it, the trained model is a derivative work, literally it's mathematically derived almost deterministically (ignoring SDG being stochastic).

The question is, even though it's derivative, is it still fair use to use it without permission?


These models wouldn't produce a sequel to Star Wars unless you only fed it Star Wars. The output works would be original works, derived from many sources. It'd be more like it outputted Star Wars because you trained it on Hidden Fortress and Buck Rogers.


> derived from many sources

You can't say it's derived from many sources and then say it's original work haha.

The output work would appear original in most cases (if you filtered out for left over watermarks and too much memorization, and avoided it copied any style or visibly recognizable characters, etc.). That said, even though it appeared original, it would be known to be derived unless you'd hide the fact you used a ML model for it.

I'm not sure you can argue it's not derived when it's literally being mathematically derived lol.

I think "fair use" is really the crux of it. Like do we feel this is a valuable derivativion to allow, because it benefits society in other ways that are worth it.


C'mon; we all know the 50 Shades of Grey was Twilight fan fiction, but it is also original work... Likewise all of fantasy, which is footnotes on Tolkien. You're suggesting a much higher bar than we apply for humans, which leads to bad places... Do we raise copyright to the point that no one can produce new work, or accept a world where generative art lowers the bar for making art?


> While Meyer has not granted permission for James to create a derivative work from Twlight, she has not yet initiated any public litigation directed at the 50 Shades of Grey series. Still, because Meyer's copyright lasts for her lifetime plus 70 years, James and her heirs may face litigation down the road.

This is in fact an unknown, it could appear that Meyer would have at least a case against it if she wanted too.

We know for sure people are capable of original thoughts. If there were no prior movies to learn from, a human could figure out how to make one, just as the pioneers in film making did.

On the other hand, models cannot, they need large quantities of existing source material, so it's unclear if they're capable of original work, and likely it's derivative.

> Do we raise copyright to the point that no one can produce new work

You don't actually have to raise copyright at all, it only takes for a court to say that ML training isn't fair use, that's all. It would only exclude ML training from fair use, meaning the people training the model would need to get a license approval from the copyright owners.

> or accept a world where generative art lowers the bar for making art?

Yes, that's the question of Fair Use. We might consider the infringement to be a net positive and want to include it as Fair Use.

Normally that's what Fair Use looks at, it's the balance between what's right by the authors, and what's best by society.

I think it's totally fair to say it be best here to allow this use because the benefit to society of that technology are worth it.

I also think it be fair to say that the benefits are not worth it compared to the loss inflicted on artists, or to think that most companies investing in AI would actually have the mean to compensate the artists anyways so it wouldn't even stifle ML innovation while also being fair to the artists.

Personally I'm still debating it with myself and not sure where I stand.

But I think it's very misleading and disingenuous to pretend that the ML model isn't even using copyrighted material or deriving work from it, and to pretend like it's similar to an artist simply looking at other people's work or being inspired by other people's work. It's very different in many ways, the process, the mechanisms, the nature, the function, the ethics, the financial applications, the capabilities, there's just so many differences here.


>it would be known to be derived unless you'd hide the fact you used a ML model for it.

If you can't tell without knowing it's ML generated, I don't see how you could argue that it's infringement.


>Derivative work isn't permitted under current copyright though, you know that?

Yes it is? There's nothing that blanket bans derivative works in the west. Parody is often derivative, and explicitly protected by the fair use doctrine in the US, for example.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: