> Yet, I know no consequences that would be measurable in physics

One can build a physical device modeled off of a Turing machine that enumerates all proofs within ZFC. The machine halts if an inconsistency is discovered, and runs forever if not. Now a prediction can be made about a process in the physical universe whose outcome depends on the axiom of choice.

I’m not trying to sound facetious actually. Highly abstract mathematics plays a critical role in inductive inference (in the sense of speeding up universal search by mapping a search over program space to a search over proofs in formal systems). This appears to be the direction some recent ML research is heading, so it wouldn’t surprise me if a lot of “unphysical” axioms end influencing our ability to efficiently approximate Solomonoff induction.