Machine-generated proofs are poised to reach largescale, human-unreadable artifacts. They foreshadow what we call the Fourth Mathematical Crisis. This crisis crystallizes around three fundamental tensions: trusting proofs that no human can inspect, understanding results that no one can fully read, and verifying systems that themselves resist verification. As a minimal yet principled response, we propose the Human Understandability (HU) meta-axiom, which requires that every proof admits at least one projection that is resource-bounded, divergence-measured, and acceptable to a verifier. Confronting these questions opens a timely research agenda and points toward new directions in scalable reasoning, interpretable inference, and epistemic trust for the era of machine-scale mathematics.