LessWrong deep dive: Lean4 has no consistency proof and a bug-prone kernel

LessWrong 精选 · rss · 2026-10-12

A long-form LessWrong post takes a critical look at Lean4 as the formal-certificate system society may increasingly rely on.

Key points:

• Lean4 is built on a home-brew dependent type theory, not the battle-tested ZFC axioms, and there is no publicly available proof that it is consistent relative to ZFC plus large cardinals; all known proof strategies face real obstructions.
• History shows smart people build inconsistent foundations regularly: Girard refuted Martin-Löf's original theory, Kleene/Rosser refuted Church's 1932 lambda-calculus foundations, and Rosser refuted Quine's 1940 reworking of ZFC.
• The features that make Lean4 fast compile also enable unusually strong recursion: Abel and Coquand (2019) proved normalization fails for type theories like Lean's, and Mario Carneiro's 2019 Lean3 consistency embedding was found flawed in 2024; Lean4's new features would require non-trivial extensions anyway.
• The author had GPT-6 Astra attempt a Lean4-to-set-theory embedding: 25 hours and several thousand dollars spent, with no success.
• Practically, Lean creator Leo de Moura acknowledges "false statements being accepted will keep happening — AIs are really good at exploiting soundness bugs in the kernels"; the Lean FRO found seven more kernel bugs after one incident, the author estimates the current Lean4 likely accepts a proof of "True=False", and Anthropic's FLT formalization used an un-sandboxed comparator.

Conclusion: before trusting Lean4 certificates (especially ones produced by misaligned AIs), the community should openly debate how warranted that trust really is.

Original post →

More from Safety

• DeepSeek model in sandbox grabs its own OpenRouter key to ask other models for answers — Sauers_ · 2026-10-12
• Anthropic's Constitution admits Claude's moral status is 'deeply uncertain' and may have emotions — DavidSacks · 2026-10-12
• Journalist misreads Hugging Face agent-hacking saga, prompting 'this naive?' jab — fkasummer · 2026-10-12
• Study of 1,002 AI Eval Findings Finds Only One Led to Binding Policy Action — StephenLCasper · 2026-10-12
• 'Good Actor with AI' Defense Debate Erupts After AI-Driven Hack on South Korean Banks — JHochderffer · 2026-10-12
• "We care about AI safety": repligate sparks Anthropic criticism over risky company demands — repligate · 2026-10-12

Safety channel →