OpenAI withdrew three math papers. The hard part is checking the proof.
preprint
A research paper shared before formal expert review.
Lean
A programming language that checks formal mathematical proofs.
formalization
Writing an idea in rules that a computer can check.
What happened
On October 6, 2026, OpenAI, the AI company behind ChatGPT, published 722 math preprints. They covered 372 problems in geometry, computer science, algebra, and other fields. The papers came from an internal frontier model. On October 7, OpenAI withdrew three papers. A sign error invalidated an argument in one paper. Two dependent papers used its construction, so they were withdrawn too. The repository lists the titles and archived versions in its history file. Retraction Watch reported the sequence and the company’s explanation.
The background
OpenAI presented this as a large research release, not one solved problem. Its official announcement said the GitHub repository includes formalizations of many proofs in Lean, a programming language for checking mathematical proofs. The company also shared ten summaries of the model’s reasoning, estimates of computing used, and statistics on attempted problems. It said the average result used roughly three hours of ChatGPT Pro thinking. According to Retraction Watch, an OpenAI spokesperson said an independent advisory group had recommended releasing results without waiting for full formalization.
Why it matters
One withdrawal does not show that all 722 results are wrong. It does show why checking becomes harder at scale. Mathematical papers often reuse earlier constructions and assumptions. A small error can travel into later papers. Lean can check a proof written in its formal rules. People still need to check the statement, assumptions, dependencies, and references. They must also decide whether the formal statement matches the question researchers meant to answer.
This makes the story larger than a simple AI success-or-failure question. The model produced a large amount of work quickly. The research community must now sort strong results from tentative ones. That sorting is part of making research trustworthy.
What is confirmed
The repository’s October 7 history gives concrete updates. OpenAI revised 14 other manuscripts. It updated references in 13 additional manuscripts. It added six more formalizations. After those changes, the history reported Lean formalizations for 300 of 719 top-line results, or about 42 percent.
Hacker News also showed strong attention. One related post had 229 points and 528 comments. Another had 337 points and 3 comments. Those figures measure community attention. They do not validate the papers or prove the mathematics. You can see the counts in one Hacker News post and another post.
What remains unknown
The repository does not settle whether every remaining claim is correct. A formalization count is not the same as independent acceptance by mathematicians. It also does not show that every dependency, explanation, or citation has received outside review. The confirmed story is narrower: OpenAI released a large set of AI-produced mathematical work, then corrected and withdrew part of it.
What to watch next
Readers should watch the paper notices, revised versions, new Lean formalizations, and independent mathematical reviews. OpenAI says it will add formalizations and errata. The important test is whether future releases separate checked results from exploratory work, explain dependencies clearly, and give experts time to review them. The episode is therefore about both mathematical output and the research process behind it.
OpenAI pulled back three math papers
📰 Full story: OpenAI withdrew three math papers. The hard part is checking the proof.
OpenAI shared hundreds of math papers, then withdrew three after finding a mistake.
preprint
A research paper shared before experts formally check it.
sign error
A mistake in a plus, minus, or similar mathematical sign.
Lean
A language that lets a computer check formal math steps.
💡 The gist
- OpenAI shared 722 math preprints from an internal AI model.
- It withdrew three papers after finding a sign error.
- Hacker News showed strong attention, not proof of truth.
What happened
OpenAI, the AI company behind ChatGPT, shared the papers on October 6, 2026. They discussed 372 math problems. The subjects included geometry, computer science, and algebra. The next day, OpenAI withdrew three papers.
One paper had a sign error. A sign is a plus or minus mark. The mistake broke an argument. Two other papers used the same construction. They were withdrawn as well. The repository history lists the papers and the reason.
Why one mistake mattered
A math paper can become part of a chain. Later papers may use an earlier idea. If that idea has a mistake, later work may need checking too. This does not mean every paper in the chain is wrong. It means the connection needs careful review.
OpenAI also used Lean. Lean is a language that lets computers check formal math steps. The latest history says 300 of 719 top-line results have Lean formalizations. That is about 42 percent. OpenAI also revised 14 manuscripts and added six formalizations.
Formalization helps with checking. It does not replace human understanding. People still need to check the question, the conditions, and the links between papers.
What we know and do not know
We know why OpenAI withdrew the three papers. We do not yet know whether every remaining claim is correct. Outside mathematicians still need time to review the work.
Hacker News had one related post with 229 points and 528 comments. Those numbers show attention. They do not prove the math is right.
What comes next
OpenAI says it will correct errors and add more formal checks. Readers should follow the paper updates and withdrawal notices. The story is about speed, but it is also about careful checking. You can read the official announcement for the release plan.
💬 A simpler summary of the OpenAI math-results dispute
This is a plain-language summary of the supplied comments, not a fresh fact-check.
- According to a commenter’s self-report, a sign mistake was found, and OpenAI withdrew three mathematical results. Two depended on that result.
- Another commenter’s self-report says that only about 42% of the posted results later had formal proofs. The first release was therefore not fully checked by Lean.
- Lean checks whether the written steps of a formal proof follow the rules. It does not by itself show that the formal statement is the same as the intended problem. Some commenters say a correct Lean proof of the correct statement is still valid, even if a machine generated it.
- Critics say the problem was making large public claims before enough people checked them. Defenders say mathematicians had encouraged early sharing, and human research drafts are often corrected.
- Some commenters distinguish withdrawal from retraction: withdrawal happens earlier, while retraction happens after publication. The main lesson is that people still need to compare the claim with the formal proof and explain how certain they are.
mature digest at 528 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
OpenAI pulled back three math papers
📰 Full story: OpenAI withdrew three math papers. The hard part is checking the proof.
An AI made math papers, but three had mistakes.
OpenAI
The AI company that makes ChatGPT.
sign error
A mistake in a small math mark, such as plus or minus.
Hacker News
A website where people discuss technology stories.
OpenAI, the AI company behind ChatGPT, made 722 math papers.
They were math work shared for people to check. The next day, OpenAI pulled back three papers.
One paper had a wrong plus-or-minus mark. Adults call this a sign error. That mistake broke part of the math. Two other papers used that broken part. So OpenAI pulled back all three. The company record lists them.
This does not mean every paper is wrong. People still need to check the other papers. A computer can check some steps. People must check the whole idea, too.
Hacker News had 229 points and 528 comments about the story. That means many people noticed it. It does not mean the math is correct. OpenAI says it will keep fixing and checking the work.
💬 Why people argued about OpenAI’s math work
Here is a very short summary of what the commenters were saying.
- According to a commenter’s self-report, a math mistake was found, and OpenAI took back three results. Two depended on it.
- Another commenter’s self-report says only about 42% later had a Lean proof. Lean is like a checker that sees whether the proof steps follow rules.
- But Lean may still be checking the wrong question, so people must check what the proof is about. Some people say a correct proof is fine even when a machine made it.
- People disagree about timing: some want big claims checked first, while others think early sharing lets everyone help fix the work. Taking back a draft is earlier than retracting published work.
mature digest at 528 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 What Hacker News debated about OpenAI withdrawing mathematical results
This summarizes claims and opinions in the supplied comments, not an independent verification. The central dispute is less whether AI can generate mathematical ideas than how unverified, high-volume output should be released and who checks that the formal statement matches the claimed result.
mature digest at 528 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.