Copyright for AI Model Weights

Andrew Marble
July 10, 2023

This is a follow up to my last post about licensing AI models.

There has been ongoing discussion about how AI model weights, a defining component of AI software, should be licensed. At the same time, many have argued that they may not even be copyrightable and thus not possible to license. I’m not going to look at the legal arguments (beyond briefly making my layman’s pitch that they should be copyrightable) but I’d like to share some thoughts about what the consequence could be if we can’t license them. Like software copyrights, AI model copyrights can be an important tool in maintain a vibrant open ecosystem and preventing tools from disappearing as trade secrets. We need strong and unambiguous copyright protection for AI model weights.

In the normal use case, I think intellectual property rights are usually too strong and only work to stifle innovation. Particularly when it comes to creative works like film where studios use their rights just to milk the same story over and over again (see Spider-Man, Disney). However there are differences with how software is distributed vs other works. If you publish a book, you by definition give away the “source code” in the form of human readable text. Copyright may be the only tool to prevent others from taking that text and making derivative works. With software, you can distribute an application without the source code, making derivative works less feasible. Copyrights have been used as a powerful tool in keeping open source (free) software open. GNU Public Licenses (GPL) and similar that require publishing code and similarly licensing derivative works have protected the Linux operating system and other software’s source code from being walled in by “closed” competitors. The license terms allow the code to be published so that anyone can modify, run, or distribute it, without fear that someone could take advantage of the codebase to build their own derivative work without giving back to the community.

I don’t think all software needs to be free, and I normally favor MIT style licenses that let users do whatever they want including build their own closed derivatives. But Linux is a prime example of a software that has used GPL licensing build a credible threat to closed operating systems. If Linux could not have been copyrighted in this way, I would speculate that we would see far more fragmentation, particularly proprietary variants that included some desirable features in an attempt to draw users in to their walled gardens. For example, a GPU manufacturer might build their own closed Linux with enhanced GPU support that would be required by anyone who wanted to work with their GPUs.

My two cents

AI model weights, the parameters that tell an AI model how to compute its output, may not be copyrightable some argue. I have not seen any definitive opinions on this, mostly just chatter on forums1234. If anyone has seen this position on this written by a lawyer please pass it on. The two key arguments I’ve seen are (1) the weights are a derivative work of the training data and/or (2) they are computer generated so cannot be copyrighted. I don’t buy either argument. The first has more subtlety and I could imagine cases where this was true (overfitting to a single work during training) but generally weights are influenced very slightly by each training point and training should be seen as analogous to reading a book and learning the ideas it presents rather than memorizing it. (There could be lots more to say here, I’m not a lawyer and I’m sure there is lots of work both sides can do to build arguments for and against, but common sense is on the side of it not being a derivative, and there's at least one expert that agrees5). Regarding the computer generated argument, what I’ve seen cited is the US Copyright Office providing guidance on AI generated content, saying “copyright protection depends on whether AI's contributions are ‘the result of mechanical reproduction,’ such as in response to text prompts, or if they reflect the author's ‘own mental conception.’” Weights are not AI generated content, they’re the AI itself, and the Copyright Office has not said anything about them. They are certainly not “mechanical reproduction” and for any nontrivial case are the results of the author’s mental conception, even if a computer was used as a tool in deriving them. If you’re sceptical, go look at one of the forums where people are building derivatives of Stable Diffusion (possibly NSFW, I’m not providing any links).

Regardless of arguments within the current copyright framework, weights should be copyrightable or otherwise protectable, arguably more so than many creative works, because of the protections copyright will afford to those that are openly distributed. I already wrote about how licensing can be a strategic tool6 but that’s predicated on it being possible to copyright and license weights. If we cannot, it’s going to unduly favor closed proprietary AI models – the inventive will be to keep weights closed as a trade secret, and any that are openly published can be unilaterally built upon into closed systems. The chances of superior Linux-style open versions being developed and maintaining their position will be diminished.

To my simple way of thinking, the fact that copyright could serve a valuable purpose is enough of an argument that it should apply here. Recently, Clement Delangue, CEO of HuggingFace (an open platform that hosts model weights) gave a presentation about the benefits of AI remaining an “open” discipline7. There are clear societal benefits from the AI ecosystem being open, which would be jeopardized if more models are hidden behind trade-secret protection because they can’t be copyrighted. Copyright and intellectual property rights don’t derive from some higher principle, they are compromises that should help creative society function in the post-scarcity wake of the printing press. With respect to AI weights, I see all upside and very little downside. There are already strong norms regarding model weight copyrights, and most are already released with licenses. It has rightly been argued that we need new weight-specific license terms, and copyright law may need updating to resolve current ambiguities. What’s important is that weights get the protection they need to help keep the discipline of AI open.