DeVault: GitHub Copilot and open source laundering [LWN.net]

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 13:49 UTC (Fri) by edeloget (subscriber, #88392)
In reply to: DeVault: GitHub Copilot and open source laundering by kleptog
Parent article: DeVault: GitHub Copilot and open source laundering

I am inclined to follow the same direction. The EU directive is quite precise on what it implies - and Copilot as a research program would qualify for the exemption but Copilot as a commercial program does not.

The same things goes for books, photographies and so on. You can train a language model on copyrighted books for research and education. But if you want to do it for the purpose of offering a commercial service, you have to get the proper authorization (this can be costly but it's not out of reach for a company like Microsoft).

I haven't read all the comments below, so maybe I'll state a point that has already been proposed. The "model is a derivative work" question is interesting, yet I don't think this is a real issue. The problem I see (and I think it's a problem at the moment I read the "we can do it" rationale by Github) is that I don't believe them: contrary to what they say, the code written by the machine is a derivative work, no matter how hard they'll try to press on this. Not only you can easily obtain code that is a direct copy paste of existing code (with comments, if needed :)) but even if you don't, you'll end up with code that 1) has a striking similary with exiting code (after I have used it for a while, I don't envision Copilot to magically imagine new algorithms) and 2) is directly inspired by the input code (Copilot is unable to code a solution to a new problem ; for exemple, it cannot propose you to use an API it does not already know).

So, as a conclusion, I would not go by the "a model created using this code is a derivative work" clause. I would go by "any code created by a ML model trained with this code is a derivative work" clause which I find both more logical and more satisfactory. As a consequence, the tool itself can exist, but cannot be used to created anything but free software - as I see it, this would be a win-win situation (although it might be tough to market to software shops :))

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 16:36 UTC (Fri) by rgmoore (✭ supporter ✭, #75) [Link]

I absolutely agree that any code produced by Copilot that is a verbatim copy of anything from its training corpus would be a copyright violation, unless that piece fell under one of the established limits on copyright, such as purely functional material that has a restricted number of ways it can be expressed or a snippet that's too small to be considered expressive. Material that's suspiciously similar to something in the training corpus would be at least deeply suspect. But that would be true whether it comes from Copilot or from a human programmer. If your code is a copy of someone else's, it's a copyright violation regardless of how it got that way unless it isn't eligible for copyright in the first place.

The problem with trying to restrict Copilot (and similar programs) is threefold:

Existing copyright law allows authors to be exposed to and influenced by existing works. For example, if I read the source code to the Linux kernel, my brain does not become tainted by the ideas therein, making any program I write a derivative of the Linux kernel. The same thing would seem to apply to teaching a ML model on existing software.
To the extent using your code as a training corpus would be considered copying, it would likely be considered to be a sufficiently transformative use that it would fall under fair use. Their ML model is a completely different thing from your original program and serves a completely different function. That is pretty much the definition of a transformative use.
Specifically regarding Free Software under the FSF guidelines, one of the essential freedoms of free software (freedom 1) is the right to "study how the program works". Software that is licensed under a FSF license has accepted this as a fundamental principle, and it seems to me that it absolutely applies to using the program as part of a teaching corpus for a ML model.

Again, this applies only to training the model. The output of the model is a different thing and may be a copyright violation even if the model itself isn't.