AI's Quiet Coup in Mathematics: Not Just Faster Proofs, But a New Episteme

The New Scientist piece captures mathematicians' unease at AI's rapid gains on advanced problems, yet frames this as a job-security worry rather than the epistemic rupture it represents. Beyond the reported astonishment lies a deeper pattern: AI systems are no longer merely verifying conjectures but generating novel proof strategies that humans struggle to anticipate, echoing the 1976 Four Color Theorem computer-assisted proof but scaling it generatively. Synthesizing DeepMind's AlphaGeometry work (announced 2024, not peer-reviewed, relying on synthetic training data from 100 million geometry diagrams without formal human oversight) with emerging arXiv preprints on LLM-guided formalization in Lean, the shift suggests mathematics is moving from human intuition as sole generator to human curation of machine-proposed structures. What mainstream coverage misses is the cultural implication—proofs may soon be accepted whose internal logic exceeds individual comprehension, redefining rigor itself. Limitations abound: current models still falter on open-ended conjectures without massive curated datasets, and no large-scale peer-reviewed studies yet quantify error rates across subfields. This is less replacement than redefinition, with implications for physics modeling and algorithmic design where mathematical certainty underpins trust.