Formalization Tools Expose Reliability Gaps in AI-Generated Mathematical Proofs

The New Scientist report highlights a growing divide where AI models produce candidate proofs faster than mathematicians can manually check them, pushing formalization as the verification layer. Yet this process itself relies on human-written libraries and tactics whose completeness is never fully audited, creating an under-examined dependency chain. Recent work with Lean 4 and the Mathlib repository shows that even formally verified statements can rest on unexamined axioms imported from earlier libraries, a pattern also documented in the 2024 Nature paper on AlphaGeometry where the model’s synthetic proofs required extensive human formalization scaffolding. This reveals that formalization is not an objective oracle but a human-mediated translation whose fidelity determines downstream trust.

Philosophically, the shift reframes mathematical knowledge from intuitive insight toward machine-checkable syntax, echoing debates after the 1998 Hales Kepler conjecture proof where computer assistance sparked prolonged verification disputes. Current AI tools amplify this tension because they can generate proofs whose intermediate steps exceed human pattern recognition, forcing reliance on the formalizer’s correctness. Analyses from the 2023 arXiv survey on neural theorem proving indicate that error rates in automated formal translation remain above 15 percent for complex statements, a threshold that undermines claims of fully automated reliability.

What comes next hinges on whether the community invests in machine-checked meta-verification of formal libraries themselves. Without such second-order checks, the apparent victory of formalization may simply relocate rather than resolve questions of epistemic trust between human mathematicians and AI systems.