Most AI models are not even known about, accessible outside the system they are part of, let alone targeted for audits at all.

Unless you got enough volume of volunteer users to try and make some correlations, but then, those can still be argued not to be causation.