OpenAI publishes hundreds of AI-generated mathematical results
The manuscripts have been made public, but many proofs still need to be checked by mathematicians.
_-1.jpg){kind=link}
OpenAI has published a large collection of mathematical results generated by an internal AI model. The company is sharing manuscripts, formal proofs and information about the working process, but emphasises that not every result has been fully verified.
According to OpenAI, the publication contains a broad range of new results from mathematics and theoretical computer science. The company is placing the manuscripts in a public GitHub collection and is also publishing descriptions of the method used, unsuccessful attempts and estimates of the computing power required.
Some of the proofs have been formalised in Lean. This is a system in which a computer can check whether individual steps in a proof follow logically from one another. Such a check says something important about a proof’s formal validity, but does not automatically make a result understandable, relevant or scientifically important.
OpenAI says that many of the results came from internal tests in which the model was asked to investigate open problems. According to reporting about the collection, it consists of hundreds of manuscripts divided into different groups of related results. The frontier models used are not available as a public product.
The publication is therefore not a collection of peer-reviewed articles. Other mathematicians still need to read the results, compare them with existing literature and check whether the formal translation accurately reflects all assumptions and definitions. Even when a proof is formally correct, the question remains what new insights it provides.
That distinction is important because AI systems have previously produced mathematically convincing-sounding but erroneous texts. OpenAI therefore says it intends to update the collection when problems are found. The openness chosen makes checking possible, but at the same time presents the research community with a large amount of material.
The development mainly shows that AI in mathematics is shifting from solving exercises towards suggesting possible research results. Whether this will actually produce new mathematics cannot yet be established. Independent verification, explanation and follow-up research are needed for that.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The publication, the public repository and the formal Lean proofs have been confirmed by OpenAI and an independent technical source. The article explicitly does not present the results as peer-reviewed breakthroughs.
- confirmed OpenAI published a broad collection of mathematical results generated by an internal AI model. — Description of the publication on OpenAI’s website. source
- confirmed Some of the proofs have been formalised in Lean. — Mentioned by OpenAI and visible in the public repository. source
- confirmed The collection comprises hundreds of manuscripts. — Independent reporting mentions 722 manuscripts. source
Editor's note
The publication and the use of Lean have been confirmed. The scientific significance and accuracy of all the manuscripts have not yet been independently established; caution with breakthrough language is therefore needed.Sources
- Sharing AI progress in mathematics — OpenAI
- OpenAI数学の新成果722編を公開 — ITmedia NEWS
- OpenAI math repository — GitHub
More on this in Dutch media
- NRC — „openai wiskunde”
- NU.nl — „openai wiskunde”
- De Telegraaf — „openai wiskunde”