OpenAI's AI-Generated Math Proofs: What Is Claimed, What Is Verified, and Why Caution Matters
OpenAI says its AI disproved the Exponential Time Hypothesis and proved two other results. What is claimed, what is under review and what to believe.
OpenAI's claim that its AI disproved the Exponential Time Hypothesis is one of the boldest announcements in AI research this year, and it is still unverified. OpenAI says internal models produced three significant results in mathematics and physics, and it has handed them to outside experts to check. For readers who follow AI, the story is a useful lesson in how to treat dramatic claims: wait for independent review. The details below come mainly from one newsletter summary and several secondary posts, so treat them as reported, not confirmed.
The three claims
First, the Exponential Time Hypothesis (ETH). Roughly, it is a belief that some famously hard computing problems, such as 3-SAT, cannot be solved much faster than trying an exponentially large number of possibilities. Think of looking for a key among a pile that doubles in size with every extra lock. According to the summary, OpenAI says an internal model found a way to solve 3-SAT in time that is faster than exponential but still not polynomial, and that this was formalised in Lean on 17 October after the disproof on 16 October. The same summary notes this would leave the famous P versus NP question untouched.
Second, a proof of a harmonic-analysis result called the local smoothing conjecture, in all dimensions. The summary says this would also imply the Kakeya conjecture, so that claim depends on the first proof holding up.
Third, a proof that pure Yang-Mills theory exists in four dimensions, which relates to one of the Clay Millennium Prize problems.
What "verified in Lean" means
Lean is proof-checking software. If a proof is fully written in Lean and the software accepts it, then every logical step has been mechanically checked, a bit like a spellchecker for reasoning rather than words. The Lean file and a model write-up are reportedly public on GitHub. However, the summary says it is not clear whether the Lean file checks every step independently or leans on an outline the model filled in, and that is an important open question. Lean also only checks that the statement formalised is the statement claimed, so experts still need to read the definitions.
Who is reviewing
For the ETH result, researchers Russell Impagliazzo, Ramamohan Paturi and Ryan Williams are reportedly reviewing it, and they say their involvement should not be taken as certification. For the Yang-Mills work, a physics group led by Nobel laureate David Gross is reportedly withholding the AI-produced material and plans to write up the work itself. Another news item says the physics release was delayed. OpenAI reportedly seeks no academic credit and gave $5 million to the Simons Foundation for related workshops. An arXiv paper cited in the summary reports no confirmed substantive error in a principal result in the assessments it examined, but notes review depth varies.
How to read this news
Be careful in both directions. It is possible that these are real advances; machine-checked proofs are a strong form of evidence. It is also possible that subtle gaps will surface during months of expert review. OpenAI itself reportedly expects extended peer review. For now, the fair description is "claimed, under review."
Practical takeaways
- When AI companies announce research results, look for public artefacts such as code, proof files and named independent reviewers.
- Separate the claim from the certification: a reviewer saying they are "looking at it" is not an endorsement.
- If you use AI for technical work, build in independent checks rather than trusting outputs on confidence alone.
What to watch
Expect updates from the ETH reviewers first, since the Lean file is public. The physics results will likely take longer.
Key takeaways
- OpenAI says internal models disproved the ETH, proved the local smoothing conjecture (implying Kakeya) and proved existence of pure Yang-Mills theory; none is independently confirmed yet.
- The ETH result has a public Lean file, but it is unclear how completely it checks every step.
- Treat the claims as "under review" until named experts report back.
Sources
- [AI Weekly: OpenAI hands ETH, Kakeya, Yang-Mills proofs to reviewers](https://aiweekly.co/alerts/openai-hands-eth-kakeya-yang-mills-proofs-to-reviewers)
- [AI Weekly: AI news today, 27 Oct 2026](https://aiweekly.co/ai-news-today)
- [Fox News live blog, 27 Oct 2026](https://www.foxnews.com/live-news/ai-news-anthropic-claude-openai-chatgpt-10-27-26)
Tags: OpenAI AI research mathematics Exponential Time Hypothesis AI verification