Open Source and AI Licensing

The conversation started, as these things often do, not with a crisis but with an accumulation of discomfort. On March 18th, the Fair Source Software movement re-entered the room with a pointed question: do the licensing frameworks that served open source through two decades of web infrastructure actually hold up when the thing consuming your code is not a developer but a training run? The answer, increasingly, is no. A day later, on March 19th, a more structured attempt emerged — a principles document aimed at giving the open source community a shared vocabulary for navigating generative AI. Two documents in two days is not a movement, but it is a signal that the same frustration has reached critical mass in enough places simultaneously that people are starting to write things down. The core problem has not changed since the previous reporting: copyleft was designed around the idea of derivative works, and a foundation model that trains on your code does not produce a derivative work in any legally recognisable sense. It produces something else entirely, something that carries the value of your contribution without triggering any of the reciprocal obligations that made open source sustainable in the first place.

If confirmed as a genuine inflection point, here is what this means. The economics of open source maintenance — already strained, already dependent on a small number of people absorbing costs that the broader industry externalises — face a structural break. Organisations building on open source foundations are currently receiving a free option on AI capability that the people who built those foundations never priced in and cannot reclaim. Fair Source licensing, which restricts certain commercial uses while preserving openness for everyone else, becomes less a philosophical curiosity and more a practical instrument. The second-order effect is harder but more important: if major open source projects begin migrating toward Fair Source or similar hybrid models in response to AI extraction, the composition of the training data available to future models changes — quietly, incrementally, but irreversibly. What gets built next depends on what remains openly available to train on. The community working through this licensing question is, without quite intending to, making decisions about the shape of AI development years from now.

Watch for whether any significant open source project formally adopts Fair Source terms in direct response to AI training concerns — that would move this from principled discussion to precedent. Watch also for whether any AI lab responds with a concrete licensing reciprocity proposal, which would signal that the pressure is landing somewhere.