OpenAI’s Math Paper Raises a Basic Question: Can Anyone Read the Proof?
Partition Principle
A long-running mathematics problem discussed in OpenAI’s paper.
Axiom of Choice
A rule about choosing one item from each collection.
Lean
A language that helps a computer check mathematical proofs.
What happened
OpenAI, the company behind ChatGPT, announced a result about a long-running problem in set theory. It says the Partition Principle does not imply the Axiom of Choice, a rule about choosing one item from each collection. OpenAI published a preprint and Lean code. Lean is a system for checking mathematical proofs with a computer.
Asaf Karagila, a mathematician who has worked in this area, discussed the release in an October 8, 2026 blog post. His post is an expert’s critique, not a verdict from the whole field. He says he read the preprint briefly. He did not examine the large Lean code in detail.
What did he object to?
Karagila says the paper is unclear and poorly organized. He says some terms are used oddly. He also points to the way several theorems are stated. The references concern him, too. The paper cites multiple unpublished, unrefereed lecture notes. One is Karagila’s own work, even though he believes published papers would have been better references.
That criticism is about research communication. It does not by itself prove that the mathematical claim is false. A proof can be correct and still difficult to read. But a difficult paper makes checking, teaching, and building on the result harder.
Why this matters
OpenAI’s announcement describes progress. News coverage can make it sound like ChatGPT solved a famous problem. That gap matters. When a company releases hundreds of solutions, researchers must decide which results deserve their time. If the papers are hard to understand, the release can create a large review burden. Karagila compares this pressure to a denial-of-service problem for researchers: attention is pulled away from other work.
He also worries about public expectations. Policymakers, funders, and non-specialists may see machine-produced results and conclude that mathematicians can soon be replaced. His point is not that AI has no use. He says AI can help organize information, make graphics, or proofread email. He argues that the field needs rules for AI-assisted mathematics before it rushes ahead. Those rules might include how to report AI use and whether reviewers should see chat records.
What is confirmed, and what is not?
The confirmed facts here are limited. OpenAI released the preprint and Lean material. Karagila publicly criticized the writing, structure, and citations. His post does not provide a full independent check of the proof. We still do not know whether the central result will survive broader expert review, how much of the work is genuinely new, or whether the code makes the mathematical ideas understandable.
The story drew attention on Hacker News, where the submission had 196 points and 296 comments in the supplied snapshot. Those figures measure community attention. They do not establish that the paper is correct.
What to watch next
The next useful test is not another headline. It is careful reading by specialists, clearer explanations from the authors, and independent checking of the formal code. It will also matter whether OpenAI invites experts to improve the exposition and references.
AI may increase the number of mathematical results produced. This episode asks a prior question: can the results be shared in a form that other people can understand, verify, and use? In mathematics, a correct answer is only the start of a conversation.
OpenAI shared a math result. A mathematician says the explanation matters.
📰 Full story: OpenAI’s Math Paper Raises a Basic Question: Can Anyone Read the Proof?
A computer can find a result quickly. People still need to read and check it.
preprint
A paper shared before formal journal review.
Lean
A computer language for checking proof steps.
Hacker News
A website where people discuss technology news.
💡 The gist
- OpenAI shared a result about a famous math question.
- A mathematician found the paper hard to follow.
- Hacker News discussed it widely, but attention is not proof.
What did OpenAI share?
OpenAI, the company behind ChatGPT, shared a preprint about the Partition Principle. A preprint is a paper published before formal journal review. The question involves the Axiom of Choice. This is a rule about choosing one thing from each group. OpenAI said the Partition Principle does not force that rule to be true.
OpenAI also shared Lean code. Lean is a computer language for checking proof steps. A computer check can help. It does not automatically explain the main idea to people.
What did the mathematician say?
Asaf Karagila studies this area of mathematics. He wrote about the paper on October 8, 2026. He said the writing was unclear. He also found the structure strange. Some words seemed poorly chosen. Some references pointed to unpublished lecture notes. He said better published references existed.
Karagila did not say his short reading proved the result false. He also did not study the large Lean code closely. His complaint focused on whether other mathematicians can understand the work. Read the original post
That matters because math needs more than an answer. Other people must follow the steps. They must find mistakes. They must learn from the work. They also need to build new work from it.
Why does this matter?
AI can produce many results quickly. That speed can create a new problem. Experts may spend their time sorting and checking unreadable papers. They may have less time for their own research or students.
Karagila is not against every use of AI. He says AI can organize information, make pictures, and fix emails. He wants clearer rules for AI in mathematics. For example, researchers might report how they used AI. Reviewers might need access to important chat records.
What happens next?
The central math claim still needs wider checking. We also need to know what ideas are truly new. The community must see whether the proof is clear enough to use.
The story received 196 points and 296 comments on Hacker News. Those numbers show interest. They do not show that the result is correct. Next, specialists will read the work. The most useful outcome would be a clear explanation that others can test and use.
💬 Is an AI math answer enough?
The source post discusses OpenAI’s announcement that the Partition Principle does not imply the Axiom of Choice [source post](https://karagila.org/2026/openai-pp/). HN commenters disagree about whether this is useful progress or work that lacks enough explanation and checking. Numbers and bugs are commenters’ own reports.
- Critics say that a proof passing Lean is not enough if people cannot understand what it proves. They worry that mathematicians must do the checking while the public announcement spreads first.
- Supporters say that if the statement matches the problem and Lean checks the proof, it can be stronger evidence than a normal human paper. People who want to read it can continue the work, and nobody is formally forced to do so.
- Others say mathematics needs more than an answer: it needs earlier research, clear explanations, and a way for others to reproduce the result. One commenter self-reported that the model and prompts were not released, so independent repetition was impossible.
- People also disagree about AGMAI. One commenter said mathematicians supported releasing the work; another said they did not support testing hard problems this way and wanted existing results researched, credited, and rewritten clearly.
- In a commenter’s self-report, a sub-n log n multiplication record was updated 115 times, moving from 1 - 1.63e-55 to 1 - 9.87e-5. Supporters see this as new research growing from the release.
- On the evaluation question, a commenter self-reported 372 results out of 4,000 problems and said the test was not yet saturated. Other self-reported examples involved an AI-assisted Collatz disproof broken by a kernel bug and a sign error; another commenter said those examples were not OpenAI’s Lean-backed results.
growing digest at 300 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
An AI made a math paper. People still need to read it.
📰 Full story: OpenAI’s Math Paper Raises a Basic Question: Can Anyone Read the Proof?
A quick answer is not enough. Friends must check the answer together.
Partition Principle
The name of a very old math question.
Lean
Something that helps a computer check math steps.
Hacker News
A website where people talk about technology news.
What happened?
OpenAI, the company behind ChatGPT, shared a math paper. The paper talks about the Partition Principle. It is a very old math question. OpenAI also shared Lean code. Lean helps a computer check math steps.
Asaf Karagila is a mathematician. He read some of the paper. He said the writing was hard to follow. He said some source names were not good choices. He did not check all the computer code. So he did not prove the answer wrong.
Why does reading matter?
A math answer needs a path. The path shows how the answer was found. People use the path to find mistakes. People use it to learn more. AI can make answers very fast. People still need time to understand them.
Many people noticed
Hacker News is a website about technology news. People there talked about this paper. Many readers noticed the story. That shows interest. It does not show that the math is right. Other mathematicians must check the paper next. They will see if the answer works. They will also see if it helps others.
💬 Who checks the AI’s math?
The source post is about OpenAI saying that the Partition Principle does not imply the Axiom of Choice [source post](https://karagila.org/2026/openai-pp/). The comments include both excitement and worry. Numbers and mistakes are reported by commenters themselves.
- The AI was said to find a very hard math result. The explanation is difficult to read, so people may need a long time to check it.
- Supporters say Lean is a careful checker and a checked proof can help. Critics say math also needs clear explanations and a way for other people to repeat the work.
- People disagree about AGMAI. One side says mathematicians supported releasing the results; another says they did not support testing hard problems this way and wanted the work explained and credited first.
- In commenters’ self-reports, a community record was updated 115 times, and an evaluation had 372 results out of 4,000 problems. Other self-reports mention bugs and errors in AI-assisted math, so people still need to check it.
growing digest at 300 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 Who should understand and verify AI-generated mathematics?
The source post discusses OpenAI’s announcement that the Partition Principle does not imply the Axiom of Choice [source post](https://karagila.org/2026/openai-pp/). HN commenters split between seeing Lean-backed formal proofs as useful progress and seeing a mass release of hard-to-read work that shifts explanation, attribution, and verification onto mathematicians. All numerical, performance, and bug claims below are attributed to commenters’ own reports.
growing digest at 300 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.