OpenAI Sends AI-Generated Maths and Physics Proofs to External Experts

What happened

OpenAI says its internal AI models have resolved three major open questions, one each in mathematics, computer science and physics. One of the results relates to a Millennium Prize Problem, and each of those problems carries a $1 million prize. The company has asked outside experts, including a Nobel laureate, to check whether the proofs hold up.

The computer science result carries the highest stakes. If that proof stands on its own, hundreds of research papers that depend on the open question would lose their foundation. The claims are still under external review, so they should be treated as unverified until the experts report back.

Why it matters

The significance lies less in the claimed breakthroughs than in the process being used to test them. AI systems are now generating outputs that sit at the frontier of formal research, and that shifts the bottleneck from producing answers to validating them. For organisations adopting AI in research, engineering or advisory work, the question is no longer only whether a model can produce a result, but how quickly and reliably a human team can confirm it.

The episode also shows how dependent a field can become on a single result. When one proof underpins hundreds of papers, an error in that proof has consequences far beyond the original claim. Any organisation whose decisions rest on AI-generated analysis faces a similar exposure, even if the stakes are lower.

By the numbers

• 3 open questions, one each in mathematics, computer science and physics, reported as solved by OpenAI's internal models
• $1 million prize attached to each Millennium Prize Problem, one of which is linked to the claimed results
• Hundreds of research papers that could lose their foundation if the computer science proof does not hold

The Renascence take

Most coverage will focus on whether AI has cracked famous problems. The more useful question for leaders is what happens after the answer appears, and who is accountable for checking it.

The real product here is verification, not the proof. Trust in an AI output is built through visible, independent checkpoints, and any operator that relies on machine-generated work without designing those checkpoints in advance is borrowing confidence it has not earned. Treat every AI output as a hypothesis until a named reviewer has tested it, and map which decisions depend on it before anyone celebrates the result. The customers who will trust your AI are the ones who can see how it was checked.