If I tell you something I learned from a copyrighted book, I am not doing something illegal.
If I produce a copy of someone else’s output and state it’s my own, I am. That’s plagiarism. And I think this is the best way to view AIs that create artworks “in the style of” etc — as plagiarism for output, not copyright violation for input.
I, and I am the author of several things including proprietary software (published, sold, even acquired!) and a novel I hope to publish soon, increasingly think copyright is an evil. I like getting income from what I make but I don’t believe in stopping people accessing what I create. Using it to restrict learning absolutely is an evil. And an AI uses it to learn.
Modern copyright is abusive and evil - we are already at the spot where large corporations entirely control the content, and pretending that copyright is somehow helping indie publishers is a pure fantasy.
Copyright can exist in a way that I would not find so egregious (and historically - it was much closer to that form when originally proposed). But the modern IP rules are explicitly designed so that large corporations reap all the profit while they pay artists as little as possible.
If AI cracks that apart.... eh - I don't see it as much of a loss.
At least for now - plenty of AI models are open weight and publicly available for general use. As long as that trend continues, I don't see the problem.
what kind of breakthrough are you expecting here?
Like humans, the model has the capacity for both high-level abstractions that could be described as merely "influence" but also it has the very-specific minute details that when reproduced in art are described as "infringement."
While I agree that some of what the model learns could/should be described as Fair Use, it's not courts strong suits to create tests for NN graphs to determine just how well abstracted things are. So, to be safe, they're likely to lean towards "just license the content."
With billions of dollars invested in AI, I wonder if lobbying could influence copyright legislation to tip the copyright duration a little back towards sanity.
I have a great deal of sympathy for artists who feel like their work is stolen. But I think there's an inevitable future where these tools are powerful and commonplace. So IMO we should focus on compensating artists for their work, especially so if they're used to train ML models.
It's very possible for neither party to be right and I wish people would acknowledge such situations more often.
Humans have constraints that are - at the moment - very hard to alter. An average human can/will read at best a few hundred books in its lifetime. You have to have some innate or acquired skill or taste to come up with something that is unique, to select what works you gonna spend your time on, that would make you better at what you do and elevate your work above others or offer something truly unique.
AI is not human, AI is billions of humans. Let this sink in, alrite?
No, copyright is not evil. Patents are not evil. Ideas are yours and you are an idiot if you give them away for free. These are your best friends if you know how to wield them.
Just look at what happened to the internet after Google put its overweight arse on it. And Cloudflare is looming in too with their totally walled garden ecosystem.
Copyright is evil... suuuure. Then what AI is that expropriates skills of humans, who have put their life to become best in whatever they do. And AI redistributes this to those who have money and power, to the elite.
It's like saying a lone fisherman and a 50m industrial trawler both catch fish, so they should be bound by the same rules and regulations. It's absurdly reductive and pathetic not to notice this error
I see AIs today who have done that. I am sympathetic to them because I wish I could have been them and achieved what they have. Don’t demean human learning just because we have systems that can learn like us.
You'd have to devour the library of congress, and every public scientific paper, and most of the text on the internet. Likely much more than you could read in a lifetime.
Then you'd have to be able to regenerate a realistic imitation of any book, on demand, in a minute, for a few cents.
Intellectual property (in the context of art, movies, books, games, etc) is so valuable because it stands on the input of how humans behave, which in turn shapes their output.
In other words, a movie script that people like and generates mimicry behavior is some sort of machine built using human parts. It has value because that mimicry or fan-related reaction is behavior that can be shaped (by a reboot series, a sequel, another work using similar tropes, a critic, a meme and so on).
That doesn't mean we are all models. Fuck that. It means a lot of us are trapped by those systems (even if I don't mimic culture, peer pressure puts me in a position where it seems I do).
The rights to a popular "intellectual property" is power. Silicon AI (in this context) stands on the web as an additional way to shape that collective behavior.
We're in for a wild ride in the following decades.
We need to understand those mechanisms, be wary of them (soundtrack use, camera perspective influence on perception of meaning, script writing and editing, sequels, rewrites, movie critics, duality play, all of it). IT folk in particular need to be very aware of such things.
With the AI model, it can regurgitate something from the internet word for word and it's on you to check it.
There's a lot of work being done to make AI more transparent but it seems like that has a way to go.
Or you realise that the models are all trained on pirated data anyways and no one cared so if you are a big enough company you can just do whatever you want.
Predatory contracts that leave only the choice of starving versus being a mind slave are commonplace in industry. Idea production is claimed by masters, instead of workers, very often.
The comparison is just a trick to make people talk about one thing while it means another.
