AI math needs an audit trail
OpenAI withdrew three mathematics manuscripts and revised others. The useful question is what has been checked—and who can explain it.
OpenAI's new mathematics collection now comes with a visible correction record. The company released the work on October 6; its October 7 repository history records three withdrawn manuscripts after a sign error undermined an argument and two dependent papers. It also records revisions to 14 other manuscripts. These are dated corrections to this week's release, not a new withdrawal announced this morning.
The current repository lists 719 manuscripts across 372 families of related findings. Those counts do not mean 719 independently verified breakthroughs. OpenAI explicitly describes a collection at different stages of verification, with some results lacking formal proofs. Retraction Watch's October 8 reporting documents the withdrawals and the company's response.
What does “verified” actually mean?
Lean lets a computer check a formally expressed mathematical argument. That is valuable, but another question remains: does the formal statement faithfully represent the claim people read?
A new preprint by Alexander Bastounis, Fabian Circelli and Anders C. Hansen examines this translation problem, including OpenAI's earlier Navier–Stokes work. The authors identify mismatches between natural-language arguments and their formal versions. We have not independently checked their mathematical analysis. Their paper is a reason to inspect the relationship between the two artifacts, rather than treating a successful formal check as a universal guarantee.
OpenAI's collection reports about 42% of its top-line results formalized and promises further additions. That is its stated status, not our verification of the proofs. The company says it will preserve earlier versions and record corrections. Keeping that history accessible matters: a reader should be able to tell which claim was checked and which version changed.
Understanding needs time and support
The Advisory Group on Mathematics and Artificial Intelligence's recommendations put human understanding at the center. They ask labs to provide clear exposition, proper attribution, formalization status and supporting artifacts, and to support community-led work that explains the results. The group also advises against using mathematical releases chiefly as marketing.
These are scholarly recommendations, not a certification of OpenAI's collection. They offer a useful standard for asking whether a release helps people understand the work or simply hands them an enormous checking task.
The wider push is accelerating too. On October 8, the White House announced an expansion of the Genesis Mission and substantial industry commitments for AI-assisted science. Announced resources are not completed scientific results or proof of safe deployment.
Our view: welcome useful discoveries and demand an audit trail. Before repeating an AI breakthrough claim, find the original result, its current version, its checking status and who can explain it. Corrections should remain visible. Understanding and independent scrutiny deserve resources alongside generation.