One sign error pulled three of OpenAI's AI-produced math papers. Find the claims that hold up the rest

On October 6, OpenAI posted 722 math manuscripts produced by an internal OpenAI model. The vast majority of the results used an unreleased internal model. A day later it withdrew three of them: a sign error in one paper invalidated an argument and a construction used by two other papers. The lesson for anyone who acts on AI work: know which claims were checked, and what rests on each one. That is why a decision map keeps evidence, claims and their links separate.

This reads public documents through the Of One method. It adds no facts and judges neither the mathematics nor the people involved.

What happened

From OpenAI’s public repository and its change history (all figures company-reported):

• The work. The manuscripts were produced by “an internal OpenAI model.” The vast majority of results used an unreleased internal model, and on average each of those results used three hours of ChatGPT Pro thinking compute. Over the evaluation, the model was posed about 4,000 problems.
• The first catalogue. 722 manuscripts in 372 families. The README said many, but not all, of the manuscripts have been formalized in Lean.
• The warning. From the first version, the README said some unformalized results “could have issues.”
• The October 7 update. A sign error in “Algebraicity of Weil classes on split abelian eightfolds” invalidated an argument and a construction used by two dependent papers. All three were withdrawn. Fourteen other manuscripts were revised, including proof repairs and corrected statements, and 13 more were updated to cite the revised editions.
• Where it stands. 719 manuscripts, with 300 top-line results formalized, about 42 percent.

None of the three withdrawn titles appears in the repository’s Lean formalization catalogue, either in the first release or now. The history does not say who found the error.

The same day, the Association for Human Mathematics published a statement, reposted as a guest post on Terence Tao’s blog. It said mathematicians did not ask for this work and urged them to stop working with OpenAI. Its sharpest line: “Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power.”

Three levels of checked

The repository already separates formalized manuscripts, unformalized results, and reasoning summaries, and anyone can copy that:

1. Machine-checked. A main result with a Lean formalization.
2. Written, not formalized. The README says some unformalized results “could have issues.”
3. Narrated. Ten families come with abridged summaries of the model’s reasoning. A summary shows how the model got there.

Of One’s skill asks every map to show what each claim stands on. Confidence is low, medium or high, with a named basis such as provenance strength and source independence. Every strong claim lists its support, its gaps and how it could fail.

What broke was a dependency

Three papers came down because two leaned on the first. Of One’s skill makes that link an object: each claim records what it depends on, and new evidence “patches only the affected dependency closure,” the exact chain of claims, edges, options and renderings that depend on it. Its validator also asks, “Does any option depend on a disputed claim?”

OpenAI’s history reads like that patch, done after the fact. The more useful question comes before the error: which unchecked claims hold up the most others?

A test to run this week: the load-bearing claim test

Pick one AI-assisted analysis your team will act on soon.

1. List the five to ten claims the decision rests on.
2. Tag each one checked (a tool or an independent source confirmed it), reviewed (a named person read the source and signed) or stated (only the model says so).
3. Draw the depends-on lines between claims and moves.
4. Count what sits downstream of each stated claim.
5. Decision rule: a stated claim that holds up two or more other claims, or any move that is hard to undo, gets checked before you move. If you choose not to check it, a named person accepts the risk in writing, with an end date and the fact that would reopen it.

The fallback in step 5 mirrors the skill’s rule for accepting risk on an open unknown: a named person with authority, a bounded scope, an expiry and reopening conditions.

A small worked example (illustrative, not a client). An AI memo backs a second location. Claim A: foot traffic near the site rose last year (stated: the model’s reading of a report). Claim B: demand supports a second site (depends on A). Move C: sign a five-year lease (depends on B). A is stated, it holds up B and C, and C is hard to undo. So A gets checked against the original report first. If A fails, you already know B and C fall with it.

Where the human gate goes

Of One’s skill requires human review for low-provenance evidence in high-consequence claims, and for moves that touch money, legal, safety or reputation. In your own work, that gate sits where a stated claim meets a move you cannot take back. The repository does not say what review preceded the release, so this post makes no call on it.

What to do this week

• Run the load-bearing claim test on one AI-assisted decision.
• Add a check-status column (checked, reviewed, stated) to AI reports your team shares.
• Record what each claim depends on, so one withdrawal shows you everything that falls with it.
• Keep a change history. OpenAI says it will log corrections as new versions and keep old ones accessible. Do the same.
• Name a person at every step that rests on a stated claim and cannot be undone.

The free tool at of1.ai/ask walks one decision through the eight starting questions, ranks its open gaps and asks who signs off. A consult is time with Utlyze running Ask, Map, Move on a decision of your own, live: $400 an hour, in 5-hour blocks.