A short argument that model outputs and training-data priors can improve rough approximations of data attribution.

A short argument that "augment, do not replace" is about data control.

A short argument that AI systems pool diffuse human work into privately governed, public-good-like model weights.