OpenAI's AI Proof Claims Need Peer Review Before We Cheer

OpenAI announced on Monday, October 26, that its models produced results in computer science, mathematics, and physics. If they hold up, they're enormous. The key phrase is if they hold up. These AI proof claims come from the company itself, and independent checks are still pending, according to reporting from The Next Web.

We make AI creative tools content, not theorems, so we're not going to pretend we can referee a conjecture. What we can do is read the announcement the way we read any big AI claim: slowly, with the sources open, and with a clear line between "reported" and "verified."

What OpenAI Says It Found

Per TNW's report on the announcement, OpenAI claims:

• Computer science: an internal model found a disproof of the Exponential Time Hypothesis (ETH), a conjecture about how hard Boolean satisfiability is. OpenAI says the model found it on October 16 and that it was machine-checked in Lean the next day, with the write-up and Lean proof on GitHub. OpenAI itself called ETH "widely believed to be true."
• Mathematics: its models proved Christopher Sogge's local smoothing conjecture, proposed in 1991, in every dimension. OpenAI says this would also settle the Kakeya conjecture and Stein's restriction conjecture in all dimensions.
• Physics: a rigorous proof that certain quantum field theories exist in four dimensions, including quantum chromodynamics, with a proof that pure Yang-Mills theory exists. That's one component of a Clay Millennium Prize problem.

Those are company statements. We have not independently verified any of them, and neither has TNW.

Who Is Checking It

The more interesting part of this story is the review structure.

For the ETH result, OpenAI asked Ryan Williams (MIT), Russell Impagliazzo (UC San Diego), and Ramamohan Paturi (UC San Diego) to organize an open, independent examination. In an open letter, the three said their involvement "should not be taken as certification of the claim." Williams said on X that the group signed no NDAs and won't be paid.

For the physics results, a group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov, will lead the review. They chose not to release the AI-produced material for now and plan to write up the work themselves.

Pro tip: When a headline says "AI proved X," look for the sentence that says who is checking it and whether they're being paid or bound by an NDA. That sentence is often more informative than the headline.

The Credit Question

A thread running through the coverage is credit. OpenAI says it does not seek academic credit, that the researchers who write the explanations should be recognized, and that it gave $5 million to the Simons Foundation for workshops on AI-produced results. It also agreed to release raw ETH data and fund a conference.

The ETH reviewers, meanwhile, warned that credit could go to whoever publishes first or has the most powerful tools, rather than to the people who check and explain the work. TNW notes that mathematicians raised similar concerns in the Leiden declaration in June. We find that worry credible and worth taking seriously, whatever the results turn out to be.

What This Means If You Make Things With AI

Most of us aren't proving conjectures. But the pattern here shows up in our corner of the industry constantly, so here's how we'd carry it over.

1. Separate the demo from the claim

A gorgeous generated video proves a model can make that clip once. It doesn't prove it makes that quality reliably. Same energy as a proof announcement: one striking output isn't a track record.

2. Prefer checkable outputs

The best detail in this story is the Lean check. Machine-verifiable output is a different class of evidence from "trust us." In creative work the equivalent is something you can reproduce: a prompt, a seed, a model name such as nano_banana_2 or veo3_1, and the settings you used. If someone shares a result without those, treat it as inspiration, not evidence.

3. Wait for the second set of eyes

We'd rather be a week late than confidently wrong. For our own posts, that means no invented versions, quotes, or stats, and a clear label when something is a company claim.

4. Watch the incentives

A lab announcing breakthroughs is also a lab with funding rounds, products, and competitors. That doesn't make the work wrong. It means independent review matters more, which is exactly what several of these researchers are insisting on.

Our Take

There are two ways to get this wrong. One is to wave it all away because it's from a big lab. The other is to treat a press announcement as settled science. The careful reviewers named above seem to be modeling the sensible path: examine the work in the open, don't confer certification early, and be loud about credit and process.

If even part of this survives review, it's a notable moment for what AI systems can contribute to research. If parts don't, we'll have learned something about how to evaluate claims like these. Either way, the interesting news won't be Monday's announcement. It'll be what the independent reviews say, and we'll be watching for it.

Until then, our advice is simple: be excited, be patient, and keep a clear label on every claim that says who said it and who checked it.