I think we need to go one level up in the law, and recognize that existing models of copyright and ownership simply don’t apply here, and come up with something new.
There is an undeniable core mathematical principle that should be recognized in law somehow: the emergent capabilities of LLMs are not possible without, effectively, the entire recorded intellectual output of human civilization since the dawn of recorded history as training data.
Trying to determine who “owns” that data, or each individual piece of it, leads us to insane distortions like LLM labs scanning and destroying books. Nobody, including the labs, thinks this is the best solution, but they’re doing it to comply with today’s interpretation of the law.
The reality is that if we are considering the entire corpus of training data, we all own it. We’re not really talking about who owns one book. We’re talking about who can be said to own all the knowledge in all the books. Who can be said to own the existence of books, or the conditions of civilization that cause us to value and preserve books in general.
The preservation and archiving of knowledge that enabled the emergence of LLMs is the product of a system of social relations that belong to the entire civilization. It is a commons. And when we talk about the future impact of these systems, we’re generally talking about taking the value of that commons, and concentrating the value of it in the hands of about a dozen people.
To me, there is no other way to think about a situation like that, where a tiny handful of people realize all the surplus value latent in the entire intellectual inheritance of human civilization, as anything other than theft.
One-off settlements with individual rights holders don't fix that. Neither does scrambling the output of LLMs so that they don’t reproduce their training data verbatim. I don’t think making LLM output non-copyrightable, or all LLM output open source, fixes it either.
I think we need to establish an entirely new framework for thinking about this stuff, and then go forward from there.
I completely agree, I was never able to put this into words but you did. All existing laws don't apply, they were written to settle disputes between humans. But now it's humans vs robots. We are in uncharted territory. No current existing law was written today's world in mind. We might be in luck and stumble upon laws that are so elegant they even apply in the AI era. But if they are not, then we need to change the law.
> we all own it
That's the only acceptable answer. LLMs at their current and future capacity would not exist without the global-scale theft of intellectual property. They are made of "public domain" without consent, therefore everything they produce belongs to the public, period.
The legal weaseling about training data, derivative work, etc., is a distraction that obscures the fundamental fact.
>There is an undeniable core mathematical principle that should be recognized in law somehow: the emergent capabilities of LLMs are not possible without, effectively, the entire recorded intellectual output of human civilization since the dawn of recorded history as training data.
Problem is, that's simply untrue.
> The preservation and archiving of knowledge that enabled the emergence of LLMs is the product of a system of social relations that belong to the entire civilization. It is a commons.
This is the right framing. LLMs are a marvel built on the commons. Unfortunately, these wondrous information synthesis engines are currently captured and controlled by a tiny insane elite.
They deserve to be rewarded for their hard work, the risks they’ve taken, and the capital that was invested. However, the output should belong to civilisation. What should be open source, is the models themselves.