Blogger Claims Lean4 Lacks Public Consistency Proof, Challenging Trust in AI Formal Verification
Milo Moses argues Lean4 has no public consistency proof relative to ZFC and cites kernel bugs, warning against blind trust in AI-generated formal certificates.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Lean4 currently lacks a publicly available proof of consistency with respect to standard set theory axioms (ZFC).
Background: As AI is increasingly used to generate mathematical proofs and code verification certificates, Lean4 is relied upon as a core tool. However, its design based on dependent type theory prioritizes compile-time speed, resulting in strong recursion capabilities that invalidate traditional consistency proof strategies.
Conclusion: Author Milo Moses states on LessWrong that, as of October 2026, there is no public evidence that Lean4 is consistent with ZFC. He cites Lean creator Leo de Moura acknowledging new soundness bugs in the kernel (such as #14576 and seven subsequent ones), noting that AIs are good at exploiting these vulnerabilities.
Boundary: This is a personal blog opinion, not an official audit. The risk that "True=False" could be accepted is the author's self-assessed probability, and the summary does not specify whether tests were simulated or on real hardware.