Yeah good point, TBH I feel like the most practical way forward right now isn’t to wait for Fair Use to clarify AI, it’s probably going to take training data that’s better curated and partitioned into groups that explicitly allow certain kinds of ML use, and models that aren’t leaky across partitions (which sounds like what the problem is with Copilot in this case). This might mean new kinds of code licenses that specify whether it can be used for ML training or not, and companies are going to need to start respecting those licenses rather than assuming they can train on everything reachable from the internet.

Otherwise, if we are going to let companies ignore copyright and licenses because of their pinky promise that their models won’t be recognizable copies of any single input, then we really do need a legal criteria for how much of a single input is fair game, right?