A critical look at Lean4: why no one can prove the proof assistant is consistent
thomasahle · x · 2026-10-12
A LessWrong essay argues that Lean4's consistency is unproven: we don't know why it should be consistent, and there are real obstructions to every known proof strategy.
- By Gödel's second incompleteness theorem, any sufficiently strong system (Peano arithmetic, ZFC) can't prove its own consistency, so a consistency proof for Lean must rest on stronger meta-theoretic assumptions
- This challenges the "formally verified = absolutely reliable" narrative
- It matters more as AI generates large volumes of Lean proofs: the reliability of the checker itself becomes the bottleneck
A foundational reflection on the Lean-centered theorem-proving ecosystem, relevant to formal methods and AI for math practitioners.
Related event: LessLong Post Questions Lean4's Unproven Consistency(2 posts)→
More from Research
- Paper: LLMs repeatedly hop between shallow heuristics and generalization during pre-training — thomasahle · 2026-10-12
- tACS tuning brain alpha rhythms reduces illusory speech percepts in McGurk effect — AnnaCiaunica · 2026-10-12
- PlaySuite benchmarks 14 open models across 5,734 video games — cgmsnoek · 2026-10-12
- ENCORE: Coding Agents Write Robot Policies From 5 Demos, Hitting 96.3% on LIBERO-PRO — 2077AIDataFoundation · 2026-10-12
- NVIDIA: Offline Student Rollouts Often Beat On-Policy Distillation in 14 of 17 Pairs — nvidia · 2026-10-12
- S-Space: Researchers Uncover a Linear Spatial Workspace Inside Multimodal LLMs — MirroS-Lab · 2026-10-12