Much of my work today is advising executives on AI strategy and training software teams to write production-quality software with AI. The question about copyright comes up repeatedly. From developers. I can’t recall a single executive asking about copyright from the standpoint of “did AI companies steal someone’s work?” Instead, the C-suite is typically asking about their own liability (hint: don’t use shadow AI). They don’t care if Anthropic or OpenAI violated anyone’s copyright. To be brutally honest, business today has little ability to simultaneously police all of their suppliers and earn a profit. Make of that what you will.
So far, the courts have often held that the use of copyrighted works in LLMs is fair use . Criticism, review, summarization, teaching, research, are all allowed under fair use and that’s the vast majority of what LLMs are being used for. Today, you can generally ask an LLM a summary of your favorite book, but not the first chapter.
If I do these things, or if I write software that can do these things, it’s legally protected behavior. So the core question as viewed by the courts is less “how is this being used” and more “how was this obtained?”
And that’s where massive theft occurred. Anthropic torrented massive amounts of copyrighted material and, when facing a class action suit, settled for a whopping $1.5 billion rather than risk losing if the case went to trial . However, that was for torrenting the material; the court had already agreed that using the material was “fair use.”
But I do know that anyone trying to build an LLM is cursing the 1886 Berne Convention because it created a massive headache. Under that, all creative work is copyrighted automatically. I sneeze into a napkin and outline the result in a blue marker et voila, there’s a good chance that’s protected work.
There is no opt-in; it’s automatic. So the vast majority of the world’s material is copyrighted, but with zero possibility of determining who owns the copyright or how to contact them. It’s off limits to everyone, not just AI companies.
This has been a huge problem long before LLMs, but I’ll skip that. Easy to look up. In short, there are tons of works which were authored, but we don’t have copyright information on, so this work is effectively “memory-holed” and no one wins. We can’t digitize and share it. We can’t compensate an author. We can’t gift it to the public. Should we just throw it away?
In fact, a report from The National Library of Wales (pdf) has this interesting claim: “In 2011, the British Library estimated that as many as 40 per cent of all in-copyright works are orphan works.” . Orphan works are works still under copyright but whose owner can’t be identified or located. No one can lawfully get permission to use them. No one can be paid for them either. These works are “memory-holed” too.
So we have several categories (not exhaustive). Works out of copyright are fair game, so long as we can unambiguously verify that they’re out of copyright. Otherwise, we have copyrighted works ...
- Without a declared copyright but for whom the author can’t be known
- Without a declared copyright but for whom the author can’t be found
- With a declared copyright but for whom the author can’t be found
- With a declared copyright the author failed to renew. Fair game in some jurisdictions, but not others
- Works with declared, active copyright, where the author can be found
That, frankly, is a damned mess. We have a hodge-podge of laws crafted over centuries, none of which thought about how to deal with this situation because this situation didn’t exist. For many items above, we have no clear way of resolving this under current law.
So we have this powerful technology which, at present, provides great value, but we’re not allowed to use unless the world’s governments get together and clean up the legal mess? You and I both know that’s not going to happen any time soon.
Or we can say, “you can only train on material that is unambiguously out of copyright (often hard to know), or for which you have compensated the copyright holder (often hard to know).” Some argue that this would destroy the value of LLMs because much of the value for LLMs undoubtedly comes from recent works which are, in effect, off limits to those without a fat wallet. Others claim that’s nonsense. As usual, the truth is probably between those two extremes.
But let’s say we require compensation to copyright holders (which seems fair), what do we do about all of the open-weight models out there? How can we enforce that? As technology improves and the training becomes easier, there will be more and more people who will simply claim that all information should be free and train these models anyway. Is a law that is largely unenforceable something that should be a law?
And there’s also a curious copyright angle. If I paint in the style of Keith Haring, but I don’t represent the result as his, I’ve done nothing wrong. And that includes if I sell that painting, so long as I’m only copying Haring’s style. But what happens if I use AI to do this? Still not a problem. What happens if I use AI to create millions of “Haring-esque” images and flood the world market, potentially crashing the value of Haring’s work? AI isn’t just exposing old, unsettled questions. It’s forcing us to ask new ones.
I genuinely don’t know the answers to these questions, but I do know that a simple yes/no is too tidy of an answer.