OpenAI publishes hundreds of AI-generated math proofs

OpenAI has released hundreds of AI-generated mathematical proofs on
GitHub, produced by an unreleased internal model from a single prompt, sparking both excitement and unease among mathematicians.
- OpenAI says the average result used the equivalent of three hours of ChatGPT Pro thinking compute, a sharp contrast to the 88 hours and 10,000 agents used for its Navier-Stokes solution last month.
- Only 162 of the 722 papers have a Lean-formalized main result, meaning about 22% of the collection is computer-checked, and OpenAI warns that unformalized results could have issues.
- MIT mathematician Andrew Sutherland said the claim that results came from a single prompt to a single agent is unverified until the model is released, and the release omits prompts that an advisory group recommended disclosing.
- MIT mathematician Andrew Sutherland says the one-prompt, single-agent claim is unverified until the model is released, and the release omits the prompts that an advisory group at the Institute for Advanced Study recommended disclosing.
In this conversation · 1
7 OctOpenAI's math proofs face scrutiny over translation errors
OpenAI released hundreds of mathematical proofs on September 8, claiming progress on 372 open problems, but mathematicians have flagged a mistranslation in the formalization of the Navier-Stokes proof into the Lean verification language, according to
New Scientist. The company also fell short of guidelines from an advisory group of elite mathematicians, which had recommended against testing advanced problems on proprietary models and urged full formalization of proofs,
TechCrunch reports. Despite these issues, experts say some results could have earned a human author a top research award just a year ago,
Scientific American notes.
- A team of mathematicians says OpenAI made a subtle error when publishing its Navier-Stokes proofs: the natural-language version and the Lean code version do not match, though the proofs may still be correct.
New Scientist
- An advisory group of elite mathematicians says OpenAI fell short of its guidelines, particularly on the need for human understanding of results, and the group's first request was to stop testing advanced problems on proprietary models.
TechCrunch
- OpenAI released more than 700 manuscripts covering 372 open problems, but not all proofs were certified correct by Lean, some have been retracted after outside scientists found mistakes, and the company did not include full prompts or run-time details.
Scientific American
- A team of mathematicians says OpenAI made a subtle error when publishing its Navier-Stokes proofs: the natural-language version and the Lean code version do not match, though the proofs may still be correct.
Source details
- First seen here
- 5 Oct 2026 12:00 UTC
- Outlets
- 20