<p>This paper examines the fundamental limitations of artificial intelligence in mathematics through the lens of Norbert Wiener’s warning that automation’s danger lies not in its failures but in its successes. Through analysis of recent developments—automated theorem proving, AI-generated proofs, and machine learning in mathematical discovery—I demonstrate that even in mathematics, the domain most amenable to formalization, essential aspects of mathematical practice resist mechanization. Drawing on philosophy of mathematics, particularly recent work on value judgments in mathematical practice, I argue that mathematical understanding requires three capacities that AI systems cannot achieve: (1) grasping why theorems are true rather than merely verifying that they are true, (2) recognizing mathematical significance through insight rather than pattern recognition, and (3) exercising value judgments about which problems, methods, and frameworks merit attention. These limitations are not temporary technical constraints but structural features of what mathematical understanding requires. The paper concludes that optimizing mathematical practice for AI capabilities risks transforming mathematics itself, privileging aspects amenable to automation while marginalizing those requiring human judgment about meaning and significance—precisely the transformation Wiener warned against.</p>