'...and the proof is checked in Lean': how to judge AI-generated math proofs
Spotted 2026-11-16 23:07 (UTC+8) · original event 2026-11-12 · sources: Hacker News · spreading on 1 platformHeat score
Rank score: 8540% — where it sits on the charts
Discussion score: 7040% — how much people are talking about it
Cross-platform score: 3020% — how many platforms it spread to
Composite heat score:68.0rank × 0.4 + discussion × 0.4 + cross-platform × 0.2
Why it matters
'Checked in Lean' sounds like a guarantee, but an AI-generated proof can game the statement, the imports, or the checker itself - and most of the OpenAI theorem dump has not yet passed these four tests.The angle
AI proved 722 theorems - but did it? The four checks that decide whether 'verified in Lean' actually means anythingSuggested format: Short video · for Math and AI audience, formal-verification crowdEvidence chain
• HN #9, 125 pts/50 cmt (2026-11-16)• Kevin Buzzard (Xena Project) blog 2026-11-12: four questions before trusting 'checked in Lean' - (1) does the Lean statement match the informal theorem (38 of ~290 OpenAI Lean entries do not even state the main theorem• e.g. paper 194 on Lech's multiplicity conjecture, open since 1960)• (2) does the code compile (Anthropic's 13M-line Fermat proof needed 96 cores / 500GB RAM to compile)• (3) does the proof prove the statement as stated, with no redefinition tricks (lake comparator checks statement identity and axioms used)• (4) does it exploit a Lean kernel bug (a soundness bug was fixed 4h after discovery• multiple independent kernels + comparator --paranoid as defense)• OpenAI theorem dump: 722 papers, ~45% with Lean code, 3 non-Lean papers retracted over a sign error• single source, no independent coverage found, cross stays 30