But in that case, physical objects you own are being taken and used, presumably you notice that your own resources are missing/dwindling. Larry didn't have the right to give what he borrowed from you to others. But in the real world, Laion isn't giving anything to anyone. If someone else misuses their giant database, it's not their problem. It's not even a problem if you find out it wasn't "someone else" and they were secretly funded by some company who benefits, because the act that would get them in trouble is the misuse of the dataset, not the entirely legal creation of it. A benign act doesn't become illegal because of a criminal act that later happens derived from it.
I don't think this is why you can't sue Larry. I mean, in this case, you probably CAN sue him, because he's actively giving your stuff to someone else without permission.
But in real world terms, I don't think the reason you can't sue Laion is because they're a nonprofit doing research. The reason you can't sue them and expect to win is because they're not doing anything illegal. Like trying to sue a rival company for sending someone to walk into your store and look around and take notes. Sorry, they're simply allowed to do that. If they actually perform some illegal action like directly copying your trade dress, then you have a case.
Analogies don't really work because there was never anything quite like it, so I apologize is my poor larry/barry story had an unintentional © maximalist-shaped hole in it. I just thought it sounded close enough to put one in the shoes of someone in a similar predicament to those of writers whose IP got swept up in a data woodchipper and used to compete with them.
Again, you're completely right when you say that Laion isn't doing anything wrong in itself. But when you take a step back and see how the data it collected was being funneled into a for-profit venture, I know you can put 2 and 2 together as well as anyone.
To elaborate further, have you tried any AI model based on open source stuff? They are awful, barely functional. It turns out that most of the value of these LLMs is derived from an Internet's worth of copyrighted pictures/text they fed into their models.
They need it. They need your copyrighted furporn very badly. But they won't pay you for it. In fact, their entire business model hinges on making certain that they do not have to pay royalties to anyone, especially not the people with the skillsets they're trying to displace, or it would be even less profitable than it already is. So when they don't argue that training on copyrighted stuff is fair use, they are doing their best to obscure where they sourced the data.
Laion is the one case where the AI startup actually showed where they got everything from. They were savvy enough to use the nonprofit status and managed to stay out of trouble that way. Other big AI players took notice of it and got so fucking scared of this potential Achille's heel that they are finding every excuse in the book not to divulge where they sourced any of it from.
They know what they're doing, and so is everyone aware of such things. Again, this is where the indignation comes from, and in my opinion, these IP holders aren't always old egomaniacs screeching at clouds.
This is what pro-IP cannot fathom, they make these allegories comparing limitless supply and limited supply and equates intellectual theft with physical theft.
View attachment 7209920
For the record I'm mostly pro-piracy, on one hand because IP laws are antiquated and riddled with bullshit none should have to deal with, but also because I'm convinced that small-time piracy it is harmless and even beneficial for IP holders the majority of the time.
However in the case of AI, the copy isn't just some p2p-transferred file existing in some NEET's hard drive, it's actively being weaponized against the original creators in an unprecedented industry-wide assault.
I'm not legal expert, neither I'm american so my understanding of american law is very narrow to say the least. I was wrong about reach in the US law of copyright laws. But from perspective of a random guy it seems that AI generally is fine in most criteria (Substantiality of the individual art pieces is very low because it uses shit tone of training data, In case of purpouse it's heavily dependent one the context of use, Nature of coyrighted work as far as I know primarily is applied to fiction, when LLM's create some kind of story they don't copy preexistent work but generate new ones based on training data) the only case when it seems questionable is the effects on the potential market.
Although case of AI and copyright is rather complex one, I'm sure of one thing: It isn't art theft. AI creates completely new things from training data. The only problem is the occasional problems as you mentioned, but they are Errors, these things shouldn't happen and due to that we can hope that as tech develops this issue becomes obsolete.
It is definitely not wholesale theft, we wouldn't even be having this discussion if it were. But I do strongly believe that they are abusing the current system and exploiting the work of artists, and that in the long run it could have dire consequences on creative industries as a whole. Your logic is good and I wish people even thought that much about it, as well as the caveats I brought up.
btw I didn't mean to bombard you with a fucking redditpost worth "erm ashktually", but I thought your initial post was a good excuse to offer this thread some contrasting views based on what legal knowledge I have.
I'm sorry. I think I polluted this thread enough for a bit now, but its the only place I get to read interesting opinions. Please show me some cows soon.