Automation Without Understanding

# Summary As AI systems increasingly produce research-level mathematics—exemplified by a 2026 AI disproof of an Erdős conjecture—the United States is simultaneously weakening its pipeline of human mathematicians capable of understanding and verifying such work. The essay argues this represents a strategic mistake, positioning mathematical capacity as critical infrastructure comparable to semiconductor capability, and proposes requiring AI systems performing important reasoning to expose their logic in formal, machine-checkable formats to ensure auditability rather than relying on opaque persuasion.