That depends on how much you value the free market as a means to an end versus an end in itself. Machine learning meta-models aren't very valuable by themselves

That depends on how much you value the free market as a means to an end versus an end in itself. Machine learning meta-models aren't very valuable by themselves, and raw personal data isn't valuable by itself, but their union creates a huge surplus of value. If you value the free market then any allocation of that surplus that happens in practice is a good allocation. If you don't, then it makes sense to look at properties you would like such an allocation to have and try to find a way to distribute the profits to meet your criteria. As an example, the Shapley value in simpler, similar scenarios would allocate half the profits to whomever trained the model and distribute the other half equally among those individuals whose data was used (or equivalently, it would treat every interaction between the model-maker and a data-provider as splitting the profits 50/50). Actually calculating the Shapley value for something as nonlinear as the hypothetical profits from a trained model on various data subsets seems...problematic...but splitting some percentage of the profits equally among the participants won't be too far off the mark.