GitHub recently announced a tool called Copilot, a tool which uses machine learning to provide code suggestions, inciting no small degree of controversy. One particular facet of the ensuing discussion piques my curiosity: what happens if the model was trained using software licensed with the GNU General Public License?
Disclaimer: I am the founder of a company which competes with GitHub.
The GPL is among a family of licenses considered “copyleft”, which are characterized by their “viral” nature. In particular, the trait common to copyleft works is the requirement that “derivative works” are required to publish their new work under the same terms as the original copyleft license. Some weak copyleft licenses, like the Mozilla Public License, only apply to any changes to specific files from the original code. Stronger licenses like the GPL family affect the broader work that any GPL’d code has been incorporated into.
A recent tweet by @mitsuhiko notes that Copilot can be caused to produce, verbatim, the famous fast inverse square root function from Quake III Arena: a codebase distributed under the GNU GPL 2.0 license. This raises an interesting legal question: is the work produced by a machine learning system, or even the machine learning system itself, a derivative work of the inputs to the model? Another tweet suggests that, if the answer is “no”, GitHub Copilot can be used as a means of washing the GPL off of code you want to use without obeying its license. But, what if the answer is “yes”?
I won’t take a position on this question1, but I will point out something interesting: if the answer is “yes, machine learning models create derivative works of their inputs”, then GitHub may itself now be considered a derivative work of copyleft software. Consider this statement from GitHub’s blog post on the subject:
During GitHub Copilot’s early development, nearly 300 employees used it in their daily work as part of an internal trial.
If 300 GitHub employees used Copilot as part of their daily workflow, they are likely to have incorporated the output of Copilot into nearly every software property of GitHub, which provides network services to users. If the model was trained on software using the GNU Affero General Public License (AGPL), and the use of this model created a derivative work, this may entitle all GitHub users to receive a copy of GitHub’s source code under the terms of the AGPL, effectively forcing GitHub to become an open source project. I’m normally against GPL enforcement by means of pulling the rug out from underneath someone who made an honest mistake2, but in this case it would certainly be a fascinating case of comeuppance.
Following the Copilot announcement, many of the ensuing discussions hinted to me at a broader divide in the technology community with respect to machine learning. I’ve seen many discussions having to wrestle with philosophical differences between participants, who give different answers to more fundamental questions regarding the ethics of machine learning: what rights should be, and are, afforded to the owners of the content which is incorporated into training data for machine learning? If I want to publish a work which I don’t want to be incorporated into a model, or which, if used for a model, would entitle the public to access to that model, could I? Ought I be allowed to? What if the work being used is my personal information, collected without my knowledge or consent? What if the information is used against me, for example in making lending decisions? What if it’s used against society’s interests at large?
The differences of opinion I’ve seen in the discussions born from this announcement seem to suggest a substantial divide over machine learning, which the tech community may have yet to address, or even understand the depth of. I predict that GitHub Copilot will mark one of several inciting events which start to rub some of the glamour off of machine learning technology and gets us thinking about the ethical questions it presents.3
Though I definitely have one 😉 ↩︎
I support GPL enforcement, but I think we would be wise to equip users with a clear understanding of what our license entails, so that those mistakes are less likely to happen in the first place. ↩︎
I also predict that capitalism will do that thing it normally does and sweep all of the ethics under the rug in any scenario in which addressing the problem would call their line of business into doubt, ultimately leaving the dilemma uncomfortably unresolved as most of us realize it’s a dodgy ethical situation while simultaneously being paid to not think about it too hard. ↩︎