AI-generated content: Legality, ethics, and the nature of art

See I’d argue that the AI isn’t regurgitate training data (which suggests data being stored) so much as it’s operating like a human with a really good memory. If I watch Infinity War and then go home and draw by best recreation of a scene then if my drawing is good enough I can’t sell it without violating copyright.

You can absolutely violate copyright using your own memory as a medium, which is what a number of A.I.s are pretty close to doing.

What a lot of these companies are doing is that they’re not really doing enough to separate themselves from the liability, and their explanations of how things work are obviously not adequate, because these programs are doing a thing they said doesn’t happen.

Like, sure, people need to be careful about this but like…how many enterprise level potential customers are being told this is an actual risk? Because this now means that there’s a possibility of it barfing out something that is copyright infringing, the user not realizing it because they don’t recognize the source material and using it. Sure, it’s on that company, but like…I don’t think Midjourney is going to be very attractive if they can’t get a handle on this.

It also creates a weird range of liability where if it does produce something copyright infringing, you could legit run into a situation where no one at the company realized it. That’s never fun, and now they’re stuck in a position where *no one* in the chain can be blamed for it. Companies don’t like that.

Edited by Zendervai on Mar 19th 2024 at 4:06:41 AM

I think that would be on the marketing firm.

My own position on the matter is that the question of whether a work violates copyright should be mostly determined by the output rather than the input.

An AI can be used to violate copyright, but it doesn't intrinsically, not even if it's trained on copyrighted artworks. In fact, even if an AI isn't trained on copyrighted artworks, it can still make copyright-infringing art.

Sounds like those music lawsuits where a tune is just a *little* too close to someone else's previously released music for comfort. Even if the musician didn't mean to go, "I'll take the melody from (that other song)".

It also actually legit means that, since some of these programs actually do sometimes regurgitate stuff from the training data verbatim or close enough as to be considered verbatim in a legal context, that it is actually reasonable to want to restrict what can be used as training data. Especially unregulated training data. Because if there's a chance that anything in the training data can be regenerated on the other end...doesn't that make the question of what exactly went into it relevant again?

Like, what if ChatGPT accidentally included a bunch of leaked login information when it was scraping the internet for data? It's very unlikely that any of it will get spit out in a useable way, but it can't be guaranteed that it won't be, and OpenAI can't control when/if it will happen. That's a problem. This is not helped by these companies making it pretty clear that they aren't actually totally sure what's in the training data. That's not good either. Yeah, it's not even remotely likely that it'll spit out something really classified or whatever, but it's not impossible and these companies have been arguing that it is.

Not really. Eventually, you're going to hit a point where you'll be able to produce something violating without including the thing you're violating in the training (or we'll come up with a slightly different technology and so on), so you might as well work out a long-term solution rather than jumping to "okay, ban inclusion of anything copyright infringing!"

... which should be focusing on the person trying to do something with this obviously copyrighted image and making sure that nobody is claiming you can get a blanket copyright on anything produced.

This still isn’t quite correct.

Nothing is being “pulled from” or “arranged”. Once the training is complete that data is entirely discarded. Overfitting of the training data can come across like memorization, but the model itself still isn’t retaining any of the original data. Overfitting can also be fixed fairly easily.

Even if a model did produce copyrighted output though, you’re still overlooking the obvious flaw in your argument here: liability would be on the person who prompted the AI to create copyright-violating content, not on the company that made the AI. People use Photoshop to infringe copyright all the time, should Adobe be held liable for that?

It’s also worth noting that for AI companies to be held directly liable for infringing outputs, AI outputs would have to be copyrightable. As it stands now policy states that accidental infringement is impossible. There has to be some kind of authorial intent for there to be infringement, and A.I.s aren’t legally considered authors.

Edited by archonspeaks on Mar 20th 2024 at 12:48:08 PM

Apparently the movie Late Night with the Devil has a few AI-generated images and people are making a stink about that.

Now the governor of Tennessee has signed a bill regulating AI and deepfakes. Source is deadline:

https://deadline.com/2024/03/ai-deepfakes-tennessee-1235865303/

Edited by BigBadShadow25 on Mar 21st 2024 at 7:52:43 AM

Bumping: Bloomberg says the CEO of OpenAI is trying to court Hollywood artists through filmmakers and agencies.

https://www.bloomberg.com/news/articles/2024-03-22/openai-courts-hollywood-in-meetings-with-film-studios-directors

That doesn't sound like particularly good news

Agreed. They learned NOTHING from the strikes.

Ok can't read much of it so judging from what little I did get to read and recent events it seems like jobs could get on the chopping block

Honestly kinda seems like regardless of whether AI training dataset is comparable to human memory or not, the public opinion seems for people who have opinion on it(I think majority opinion is "sounds neat, I don't know much about it") seems to be that its plagiarization to not pay for use of training data? Like no matter what, that its unfair someone can just take someone else's work and use it to train ai without permission

(I kinda have question related to that: has there been similar case of it being considered distasteful to use non public domain work to teach people to write or draw art or such?)

Edited by SpookyMask on Mar 22nd 2024 at 10:26:42 AM

Never mind. The comment I linked to was deleted.

Edited by BigBadShadow25 on Mar 26th 2024 at 4:42:52 AM

its been deleted so..I wouldn't believe everything you hear on reddit

Yeah, I’m starting to figure that… but it’s probably good to keep an eye on things.

Bumping. Apparently the BBC used AI to promote Doctor Who… and they’re stopping because they got complaints.

Source is Deadline: https://deadline.com/2024/03/bbc-doctor-who-ai-complaints-1235867333/

It's proof complaining works

Came across all wrong.

Edited by BigBadShadow25 on Mar 27th 2024 at 3:10:49 PM

Complaining to a megacorporation that using AI to avoid paying people is bullying now?

…I’m sorry, I was joking.

It’s good that they stopped using AI.

Edited by BigBadShadow25 on Mar 27th 2024 at 3:10:33 PM

I hope this sets a trend of speaking up against AI and companies listening

I'm a bit curious about what the complaints were like. Was it that people immediately clocked that the writing was AI and complained before it was confirmed?

Because that's a thing I've noticed. Companies immediately back off and start dialing down plans when they realize that...oh, a lot of people actually can tell when something is AI created and it's not a difficult task.

Amusingly there's an AI program browser extension that detects if something is likely written by an AI.