A scientific commons for AI: community governance as the missing layer in global science

Foundation models are being absorbed into scientific workflows faster than the governance tools available to scrutinize them. Regulation, open-weight releases, transparency audits, public procurement standards, hybrid public-academic compute, and national sovereign-compute programmes each address part of the problem. None, however, reaches the upstream question of who decides what a model is trained on, how it is optimized, and when it is retired. Here I argue that a science-governed model layer, jointly owned by an international scientific community on the CERN pattern rather than by any single state, is the upstream complement the existing toolkit requires. The case is symmetric: the layer closes the gap the other tools cannot reach, and those tools cover the residue the layer cannot, so the argument is for the full set, not the layer alone. The concern is specific to the generative foundation models now in widespread scientific use, where the same systems mediate literature review, code generation, and hypothesis articulation across fields. With the UN's Independent International Scientific Panel on AI now seated and its first assessment cycle under way, the scientific community has a live opportunity to specify what such a layer should be while the governance agenda is still forming.