Lean checks do not alone prove AI-generated theorems correct
Original title: “…and the proof is checked in Lean”
Of OpenAI’s 722 dumped papers, about 45% include Lean, and roughly 38 of 290 do not state the main theorem.
Kevin Buzzard’s Xena Project post explains what “checked in Lean” does and does not establish for AI-generated mathematics. OpenAI’s dump of 722 papers included Lean code for about 45%; three papers without Lean were later retracted over a sign error. He says about 38 of 290 Lean-accompanied items do not even state the paper’s main result, and that statement matching, compilation, and related checks have not been done systematically.
Unconfirmed: Buzzard says the extra checks have not been run systematically and he has no reason to expect they will fail.
- OpenAI never claimed the Lean code proved each paper’s headline theorem.
- Paper 194’s Lean headline matches Corollary 5.4, not the full conjecture.
- Axiom’s finite simple groups statement sits on over 250,000 lines in Tau Ceti.
- Anthropic’s Fermat formalization was about 13 million lines of generated Lean.
- A Hodge conjecture statement PR relies on over 10,000 lines above Mathlib.
From the original article
8 min read at xenaproject.wordpress.com
This post is an attempt to explain in simple terms how to interpret claims of the form “this theorem is correct because it was checked in Lean”. The summary is: there is an important signal that can be extracted from Lean code, but if this code is AI-generated then you need to do some tests first.
By matt_d · xenaproject.wordpress.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.