Did OpenAI Really Solve ETH and Yang-Mills, or Will AI Answers Just Say It Did?

Did OpenAI Really Solve ETH and Yang-Mills, or Will AI Answers Just Say It Did?

OpenAI's October 26 announcement says its models refuted the Exponential Time Hypothesis, proved the local smoothing conjecture behind Kakeya, and showed that pure Yang-Mills theory exists. Every one of these is a claim awaiting independent review. Coverage is split between headlines that attribute the claims and phrasing that states them as fact. AI answers will likely inherit whichever framing gets cited most.

What is the gap between the source text and the headlines?

The gap is in the verbs. OpenAI's own announcement uses a flat declarative section heading, "The Exponential Time Hypothesis is false." The page's body is more careful. It describes the manuscripts as a form of raw research data and describes outside groups that will verify and explain the results.

Secondary coverage made different choices:

• The Next Web headlined that OpenAI says its model disproved ETH. That framing attributes the claim and keeps it as a claim.
• The Neuron's daily digest described them as unusually ambitious claims for independent review to settle.
• Some aggregator headlines repeated OpenAI's declarative phrasing more or less word for word, with "claims" sometimes the only hedge.

For citation analysis, that one-word difference matters. A language model summarizing "OpenAI says X" and one summarizing "X is false" can produce very different answers to the question "Is ETH true?"

Which caveats are most likely to get lost?

The caveats most at risk are the ones that take more than a clause to explain. Three stand out:

1. Review is not certification. According to The Next Web, ETH co-proposers Russell Impagliazzo and Ramamohan Paturi, joined by Ryan Williams, wrote that leading the examination does not certify the claim.
2. "Part of" a Millennium problem is not the problem. The Clay Mathematics Institute describes the challenge as proving Yang-Mills existence and a mass gap. OpenAI describes its existence proof as one part of that problem.
3. The physics can't be examined yet. A group led by Nobel laureate David Gross chose not to release the AI-produced material and will write up the work itself.

Lean formalization is a fourth point that summaries can easily distort. OpenAI says the ETH disproof was found October 16 and formalized in Lean October 17. Machine checking is strong evidence that a formal argument is internally valid. On its own, it doesn't show that the formal statement captures the conjecture as experts understand it.

Who ends up credited when an AI answer retells the story?

When an AI answer retells the story, the company's name is the one most likely to survive. OpenAI's stated principles reject academic credit for model-produced discoveries and say the human groups who write the explanations deserve recognition. Compressed retellings tend to keep the brand and drop the reviewers.

The ETH reviewers flagged this dynamic directly. According to The Next Web, they warned that credit may go to whoever publishes first or holds the most powerful tools, while verification and explanation get less recognition. The Leiden Declaration is a June 2026 statement endorsed by the International Mathematical Union. It names weak attribution and overhyped claims among the five threats AI poses to mathematics. It also notes that AI output often fails to cite the prior human work it relies on. In this case, that prior work includes the 2019 two-dimensional local smoothing proof by Guth, Wang and Zhang and the 2025 three-dimensional Kakeya proof by Wang and Zahl, both of which OpenAI's post credits.

How should communications teams read this episode?

Treat it as a case study in claim governance, not just science news. The pattern applies to any organization whose announcements get repeated by AI systems:

• Your headings get quoted more than your body text. A declarative subhead can outrun the caveats in the paragraph below it.
• Attributed framing protects accuracy. "X says" keeps a claim open to revision, while a bare assertion closes it early.
• Credit follows clarity. Named people and institutions survive summarization only when sources state their roles plainly.
• Corrections lag. If review overturns a result, AI answers built on early coverage may keep repeating the original claim until newer sources outweigh it.

Brands can track how this plays out for their own claims. MarkGrid's Model Share tracks how AI models mention and recommend brands. Its GEO guide covers how to structure pages so qualifiers and attributions survive citation, and teams that want to test messaging before launch can look at Neuro-Marketing. For the full breakdown of what is proven and what is claimed, see our MarkGrid analysis.