AI and Copyright: Medium Sized Thoughts

Ok, still working on BIG thing about AI and writing and images and originality, but let me pull off a piece of that I want to address more quickly (but in more depth than I could in a Bluesky post) right now. Namely, some musings about AI, machine reproducibility, and intellectual property rights.

I wanted to get this out quickly in part because Scott Weingart dropped an absolute gem of bluesky post last night that really clearly summed up something about AI and the economy of information that I've been thinking about for awhile. His post succinctly identifies that the problem that AI proponents promise to solve for information: making more of it more cheaply, is the wrong problem to solve. As I've pointed out before, the problem with information is we have too much of it, this has been true at least since print, and Scott points out an important reason for this. Too much information negates itself, becomes noise. The best I could do to try to add to his observation was to make a Hegel joke (an incredibly successful joke for an audience of exactly three people worldwide).

Here I want to address something related to Scott's point. See, one of the ways we traditionally managed the economics of too-abundant information has been copyright.

Copyright has been invoked rather a lot in the debate about AI. For opponents of LLMs and other generative AI systems, the fact that these systems were trained on copyright-protected works without the author's permission is theft. The creators and proponents of AI, on the other hand, argue that the training of systems is an allowed form of transformative use, no different - they argue - from what a person does when they learn from and build on previously published work.

There are, in fact, important differences in what LLMs do from what people do when they learn. That's what the longer piece is about. There are also important differences between an LLM's output and a traditional "copy." But right now, what I want to focus on is this: the moral language of theft tends to distract us from one of the more pragmatic purposes of copyright. A pragmatic purpose we need to think about clearly if we're going to navigate the post-AI information economy, and one, I will argue, we can't achieve simply by extending property rights further.

One of the key things copyright is designed to do is to give creators an incentive to make work public. Before mechanical reproduction, authors had little reason to keep work secret, or worry about property rights in text, because the sheer effort that it took to reproduce a written document (to say nothing of images) was so profound that anyone willing to copy a document for you was basically doing you a favor. The economy of information was not based on authors selling text, there wasn't any way to do that at scale. Instead, authors built patronage relationships with the powerful and the wealthy. The public circulation of text was an economic after thought.

Print changes the game. Now if an author makes work public, anyone who owns the means of reproducing text (a printing press) can take that work, copy it, and profit off of it and (in the absence of copyright) do so without extending any benefit to the author. Under these conditions, creators have a strong incentive to keep work a secret so they can maintain their economic control over it! That's what copyright is supposed to address (in part). Giving an author an economic incentive to make work public by making sure those who own the means of reproduction have to share their profits with them.

This was, as anyone who has studied copyright and the relationship between various creators and various publishers knows, a deeply problematic and highly unequal exchange in which publishers often had the upper hand. Still, it got us a couple of centuries of published work so it wasn't entirely unsuccessful.

Digital technology upsets these traditional copyright relationships around the turn of the 21st century in a variety of ways. Now there are amateur producers building off of copyrighted work with a potential mass audience (remix), now there are aggregators (search engines) that don't quite copy work but do profit from it.

AI, of course, is making that aggregation problem much worse. Google served ads around your content, but at least linked back to you so you could get traffic and maybe revenue. This was another very unequal status quo but one that was at least somewhat workable for some players. Now AI chatbots summarize content so that users don't even have to click the link to it. That's a real problem.

But wait, there's more! Generative AI, as Scott points out, can also generate new content based on your content in ways that were not mechanically possible before. Just to name one example, take the well know use of "in the style of (Artist's Name)" prompts in image generation models. It was always possible, and generally protected, to borrow from another person's style. It was not mechanically possible. The mechanical possibility to copy style shifts a power relationship just as the mechanical possibility to copy text shifted a power relationship. Now those who own the means of reproduction (AI models) can profit off someone else's stylistic work at scale.

This means there is now an incentive to keep your work secret. I honestly wonder if that's one reason some social media spaces are seeming to dry up. Why share what will be used by someone else with no benefit to you?

It can be tempting, then, to think that we can solve this by extending copyright. Let's give people property rights in their style. In their ideas. If their work is used to train a generative model, pay them.

I don't think this works though! For one thing, "style" is a much fuzzier object than a specific text ever was. If I authored a text, I authored that text. But my "style?" I think I have a fairly recognizable written voice. I try to. Still, even I understand that there are days that voice is stronger and days it is weaker. Times when I turn it on and turn it off. Where does it begin and where does it end? More importantly, my voice is absolutely a mash up of the styles of a bunch of other writers I admire. Is it really mine? Can I own that?

Just as important, as Scott points out, once a model is trained its output is theoretically infinite. Even if we somehow magically solved the identification problems I mentioned above, we only solve it for authors right now. Imagine your favorite author gets paid handsomely to license their style to an LLM. Now it writes novels in their voice. Forever. What happens to future authors? The triumph of dead labor over living could hardly be more literal outside of very literate zombie movie.

So no, I don't think intellectual property rights can work here. They become incoherent. Instead, we have to understand that if human beings are to retain agency in their society, if we are not going to literally turn it all over to the machines (and I don't think we should), we are going to have to continue having human beings do intellectual labor. And that labor will have to be supported.

More about that in the longer thing I swear is coming. Really.