I’ve been guilty, I think, of waving aside questions about how AI will test copyright law (or vice versa) as being, well, not Real or Serious Questions. Not because the development of AI hasn’t been real and in the pipe for quite a long time, and not because copyright isn’t implicated, but rather because it struck me that copyright law was doctrinally well-suited to handle the tension and, moreover, some of the smartest and most knowledgeable copyright experts had been pondering these questions decades before they matured into today’s AIs, and have already provided much of the groundwork for reaching sensible answers.
I was wrong here. Which isn’t to say my opinions about the right (copyright) answers have changed! For instance, I still believe:
- Non-human authors shouldn’t get copyrights, but we should be open to copyrighting machine outputs — with shorter terms — once we have a better understanding of the economics of non-human authorship at scale. I agree with Pam Samuelson (or at least the Pam Samuelson of 1986? (PDF link)) that if there is going to be a copyright it should generally belong to the user.
- Use of copyrighted material as training data should generally be fair use, and jurisdictions without fair use should provide for it by law. We’re talking fair use here, so, obviously there will be edge cases and exceptions, and other legal regimes (most notably data protection/privacy) will continue to pose much, much thornier questions.
- Outputs, obviously, can and will infringe. Liability should rest first with the user — that is, my view is that providers will generally, but not always, lack the necessary volitional step.
Where I was very, very wrong, first and foremost, was in underestimating the public’s concern with and interest in the whole topic. Which, I admit with the benefit of hindsight, should have been obvious! This tech has the potential to radically reshape the creative economy and copyright is at the core of how that economy works. Fundamentally, this ain’t a series of concerns for specialists, experts, and policy wonks: the public is directly implicated, and they’re very rightly going to be directly engaged. Technocrats don’t get to hammer this one out in a backroom and that should have been bleeding obvious from the get-go.
In particular, the training data question is so much more of a live wire than I had imagined. At a basic level, it sure looks like machines are using our work product to train our replacements. For my part, I don’t think the doom level is actually that high — I see lots of room for increased productivity, and no reason to believe that human creativity is on the chopping block — but I have a much better understanding in 2023 of why it’s important we take the concern seriously.
Another error: I failed to imagine the sheer volume of crap machines can make (again: duh!). For all the handwringing about information overload in the Web 2.0 years, it’s just nothing on what we’ve got coming. There’s a copyright angle to this in that a world of sharply decreased content creation costs might well demand substantial tweaks to our incentives model . . . but more generally, wow, I shudder to imagine the internet where many dozens of these things are spewing out text on demand. Could be a good incentive to unplug and spend more time with books, though.
I doubt that this little bit of introspection is going to make me a better augur going forward, but hey. Can’t hurt.