Haha, AI can't do $THING

I have heard this argument before, and I've thought very carefully about it.[0] Regardless of your feelings on the morality, I think the law as it is currently structured is unprepared to recognize this as "theft" (or infringement, or whatever you want to call it).

To understand why, we need to go back to the 17th century. Copyright law is a brand-new idea, and we need to codify how it works. You *could* write something like "don't rip off other people's books etc. using a movable type printing press." But then you'd have to update the law every time a new printing technology is invented, and that'd leave a lot of authors (and other artists) unprotected in the meantime.

Instead, they wrote something more like "don't rip off other people's books etc. using any technology that exists or will ever exist." In modern terms, the law is technology neutral.[1] For the last ~300 years, this has worked out pretty well in practice. Technologies come and go, and the law adapts. But we don't get this flexibility for free. Judges and lawyers have to figure it out on a case-by-case basis, and they need a coherent theory of how that works. That theory is now the problem.

Broadly and informally speaking, most analysis of infringement uses a black-box model. That is, instead of focusing on the specifics of how the technology works, it focuses on the inputs and outputs to the process. For example, when file sharing gets invented, and pirates argue in court that e.g. "BitTorrent only shares small pieces of the work at any one time," judges can make short shrift of that argument. The overall effect is to make and distribute copies of the work, so it's infringing. We don't need to care about the details of the BitTorrent protocol (nor any of the other file sharing apps). We just look at inputs and outputs, and in that view, the file sharing app looks functionally identical to a copying machine.

But AI opens a can of worms, because its inputs and outputs do not look much like classic infringement, at least in some cases. Sure, there are situations where the AI produces an output that looks like Mickey Mouse or another copyrighted character. Those are easily litigated and somebody probably loses (the main open question is whether that "somebody" should be the AI company or the end user who generated the image). But there are also situations where the AI's output is not substantially similar[2] to any individual training image. The law has seen that fact pattern before. It looks very much like the inputs and outputs you get from a student learning to draw or write.

Obviously, the learning process of an AI is quite different to the learning process of a human. But if you're black-boxing that process away, then you don't care. You look at the inputs and outputs, conclude that the AI's training is broadly similar to something that has been legal for three centuries, and dismiss the case.

So that leaves the question of what to do about it. I think there are broadly three paths that the law could take:

0. Do nothing, and accept that AI is (probably) legal in some applications.
1. Modify the law to carve out the specific kind of mimicry that AI performs, while still being technology-neutral in principle.
2. Modify the law to treat AI differently from other technologies.

(1) is probably impossible. The distinctions between the inputs and outputs created by an AI on the one hand, and a human art student on the other, are too fine for judges to reliably analyze in the courtroom. More importantly, there are many other edge cases that we would need to exclude, such as making sure that the Tolkien estate does not accidentally gain ownership over the entire genre of high fantasy. And of course, AI is rapidly evolving technology and any "AI fingerprint" we identify now will probably be worthless in six months.

(0) is obviously workable (regardless of its merits), so needs no further analysis. That leaves us with (2) to consider. The problem with (2) is that it turns the law into a game of Whac-a-Mole. Legislators will write up some description of how AI works, and then a few months later, some startup will find a way to make an AI-like-thing that doesn't quite fit the legal definition. This repeats until lawmakers get tired of it, and then we're back to (0).

Alternatively, we end up with a definition so broad that it covers some non-AI technologies such as procedural generation. Would Minecraft infringe just because some of its terrain generation loosely resembles some famous landscape paintings? Most of those paintings are in the public domain, so I suppose not, but that feels like a scary question to even have to ask. Tech neutrality is the seal that holds back these bizarre and awkward questions, and breaching that seal leaves us in a world of confusion.

--

[0]: Disclaimer: I work for big tech and have both monetary and job-related biases in favor of AI. All opinions are my own.
[1]: The law is technology neutral on the question of infringement. It's not neutral on other questions, such as the threshold of originality, but we're not talking about that (yet).
[2]: See https://en.wikipedia.org/wiki/Substantial_similarity and references linked from there.