It's fairly clear to me that trained models are a transformative rather than derivative works. However, without distributing forms which are not transformed, open/distributed/permissionless work in this space is either severely hampered, or basically impossible. I think a lot of people instinctively feel that removing barriers to their work is particularly important, since if we don't it could remain impossible for anyone outside a few megacorps -- who have the resource to obtain whatever they need in multiple ways without the additional liability of distributing non-transformed copies externally -- to build highly scaled up machine learning models.