Copyright vs AI
Updating IP Regulations for AI Distillation
Andrew Marble<br>marble.onl<br>andrew@willows.ai<br>July 18, 2026
Copyright law isn’t “natural” in the sense that it doesn’t stem from<br>some higher principle. It’s easy to see that laws against murder or<br>theft derive from morals and natural rights. Copyright is more like a<br>government subsidy, where we’ve decided that someone who creates<br>something has the exclusive right to sell it, and the state will spend<br>its own money to intervene to protect that exclusivity. There’s<br>certainly a fair version of the world where you can copy anything you<br>encounter freely and don’t owe anybody anything. The ostensible argument<br>for copyright, as I understand it, is that, along with other<br>“intellectual property” laws like patents, it encourages innovation.<br>People wouldn’t create anything if nation states didn’t put their might<br>behind making sure nobody copied it.
Anyway, copyright is unnatural – it’s something we collectively<br>decided on to try and balance personal freedom with promoting<br>creativity. One can argue about how effective it is, if it’s the right<br>balance, etc. but this is where we are. What I think is clear, is that<br>it’s completely irrelevant to AI model training. I want to focus<br>specifically on “distillation” where someone effectively copies some<br>part of an AI model’s characteristics by generating lots of text (or<br>other outputs) with it and using that to train a new model. This is what<br>some of the Chinese labs have been accused of doing to create their open<br>models. Companies like Anthropic claim that the Chinese AI providers run<br>millions of Claude queries to get the conversation traces the need for<br>training, essentially copying the capabilities of Claude or GPT or<br>others, instead of training models from scratch themselves. The<br>implication is that this is unfair, it takes the hard work of others and<br>profits from it, without compensating them.
So if for example, Anthropic has to pay $1 Billion to get data<br>labelers to provide datasets, to invent training techniques, etc.,<br>someone like DeepSeek could come along and spend a small fraction of<br>that to exfiltrate conversations from Claude that encompass that<br>knowledge and use them to train their model. Since they spend<br>comparatively little, they can charge way less, undercutting Anthropic<br>(who would have liked to be a monopolist or oligopolist). In the long<br>run, companies like Anthropic won’t want to invest anymore because they<br>know copycats will simply undercut them, there will be no more<br>innovation, etc. etc.
The interesting question for me, is what, if any, state intervention<br>here makes sense. Copyright does not generally appear to be relevant to<br>AI model training, for a number of reasons, not the least of which is<br>the model is not a copy, and that the capability to violate copyright is<br>not the same as violating copyright. Successful copyright cases have<br>generally involved things like training on “pirated” works instead of<br>paying for them; once a work is purchased, training on it is fair game,<br>just like it’s ok to read a book and remember the plot after.
So my question is, do we need another kind of bargain, another<br>unnatural law created as a compromise to spur innovation in relation to<br>AI. It seems that there are ongoing attempts at this including some that<br>are roundabout or subversive, like trying to frame things in terms of<br>national security. So it might be valuable to have a more open<br>discussion about what kind of tradeoff would be appropriate.
Some positions that come to mind are:
No distillation : outlaw (in the same sense as<br>copyright) training on model output unless specifically licensed. This<br>would essentially be putting the weight of the state behind current<br>terms of use that models have. The challenge is that it would be<br>extremely hard to enforce. If I start sharing songs online, it’s easy to<br>check if they are copyrighted. It’s not really possible to tell<br>definitively that a model has been distilled or that traces have been<br>used in training, so cases would be less black and white and hinge on<br>expert analyses, etc. instead of just common sense.
Permissive / do nothing : Given the unnaturalness<br>of it, there is no moral or higher imperative to do anything about<br>distillation. We can just let it happen. This would keep the current<br>competitive environment (to the extent that distillation is even<br>relevant) which is currently producing better models every week from a<br>variety of sources. The counterargument is that over time innovation<br>will slow because there are not the incentives to keep investing if it<br>doesn’t provide an edge. And unfortunately, doing nothing creates a<br>vacuum in which people will want to “do something!” and the something<br>might be worse if we don’t tackle it head on. (See the next<br>point)
Oblique rules targeting distillation : this I’ll<br>come out and say is the worst option but may be the most likely. Instead<br>of actually tackling intellectual property concerns, we instead fight a<br>proxy...