Federated learning is often promoted as a privacy‑preserving approach: personal data stay on the device and only model updates are shared. It borrows terminology from the federated social web but flips the logic, distributing computation while the final model remains with the entity that convened the training. We argue that federation alone does not solve extractive AI, because outcomes depend on who controls the data and the model and who has agency over the practices that shape them.\ \ A creative community can exercise governance at three layers: storage, circulation, and learning. Artist‑run trusts, cooperatives, and consent infrastructures have already established creator governance for storage and circulation, but they stop short at learning: contributors may consent to training yet have little influence over the resulting model or its federation.\ \ From this observation we map a research space that pairs technical open problems with the underlying human questions, and we propose four design principles for a creative data commons that governs models and their federation, not just datasets: 1) govern the model, not only the corpus; 2) make the terms legible at the moment of contribution; 3) design refusal as a first‑class state; 4) decide stewardship openly and be accountable for it.\ \ Review