Are OpenAI’s AI Math Results Verified? What We Know

• OpenAI’s released mathematical work has mixed verification status; the sources do not show that every result has been independently checked.
• A repository summary should not be treated as evidence that every result has been independently checked.
• A proof assistant checks an encoded proof against encoded definitions and assumptions. That does not by itself show that the encoding matches the manuscript.
• For a specific result, record the manuscript and formalization separately, then look for documented evidence of human review and its scope.

Are OpenAI’s released math results verified?

As of October 16, 2026, the available sources describe results at different stages of verification, not a collection in which every result has been independently checked. OpenAI’s October 6, 2026 announcement presented mathematical work produced by an unreleased internal model. ^[1]

OpenAI’s repository provides material related to the released work. ^[2] Making a manuscript available for inspection does not establish that its argument is correct, and the sources do not establish that mathematicians have checked every result. AGMAI, an independent advisory group, said its advisory role should not be read as a judgment of the results’ impact or an endorsement of OpenAI’s process. ^[3]

What does “verified” mean for AI-generated mathematics?

Verification can refer to several kinds of evidence. They answer different questions and can occur in any order:

• Manuscript available: A written argument or result is available for readers to examine. Availability alone does not establish correctness.
• Proof formalized: A proof is encoded in a language that a proof assistant such as Lean can check. OpenAI says it is publishing formalizations for many proofs as they become available. ^{[1] [2]}
• Software check passed: The proof assistant accepts the encoded proof from the encoded definitions and assumptions. This checks the formal artifact; it does not establish that the artifact captures the manuscript’s claim or that the assumptions are appropriate.
• Human review: A mathematician examines the reasoning and, where relevant, checks that the formalization corresponds to the manuscript. A documented review should make clear what was checked.
• Community uptake: The result has been assessed and incorporated into mathematical knowledge. This is not established simply by software checking or an individual review.

AGMAI described the release as “the beginning, not the completion,” of human understanding and incorporation of the work into mathematical knowledge. ^[3]

What do OpenAI’s verification numbers actually show?

The repository README is a collection-level source, but its summary alone does not provide a basis for assessing the correctness of any particular result. ^[2]

A collection-wide summary should not be treated as an independent measure of correctness. It does not establish that every formalization matches its manuscript or has passed independent expert review.

Reported process information does not establish the verification status of an individual proof. ^[2]

How can readers assess a specific result?

Use the item itself—not the collection-wide percentage—as the unit of assessment. Record what is available and what evidence supports each status. If the repository does not report a detail, label it “not reported” rather than treating it as verified.

1. Record the manuscript. Note its title or repository identifier, version or date if shown, and the exact theorem or claim you are assessing. Do not assume that one status applies to every claim in a manuscript.
2. Record the formalization separately. Note whether a formalization is linked, which theorem or claim it addresses, and any repository revision or successful check information provided. A manuscript without a linked formalization is not evidence of a software-checked proof.
3. Check the software evidence. Look for an explicit report that the formalization builds or passes its proof-assistant check, and confirm that the checked theorem is the one under review. If that information is absent, label software-check status “not reported,” not “passed.”
4. Check the human-review evidence. Look for a named reviewer or responsible author, a report or statement describing the review, and its scope. Strong evidence should say whether the person checked the mathematical argument and whether they checked that the formalization expresses the manuscript’s claim. A general statement that a paper was reviewed does not answer both questions unless its scope is clear.
5. Record community status separately. Note any evidence that other mathematicians have assessed or incorporated the result. If the only evidence is a release or a formal proof, community acceptance remains “not established.”

AGMAI recommends that authors understand mathematical arguments, verify their correctness, and take responsibility for a paper’s content. ^[4] Those recommendations describe responsible practice; they are not evidence that a particular OpenAI item has met each criterion.

What does the announced partnership establish?

The announcement does not establish the status of any particular result.

Frequently asked questions

Does an unformalized result mean the result is wrong?

No. “Unformalized” means that a software-checked proof is not provided as a formalization. It does not settle whether the manuscript’s argument is correct; that requires separate evidence about the argument and any human review. ^[2]

Sources

• ^[1] OpenAI, announcement dated October 6, 2026; passages on the mathematical work, planned formalizations, and Simons Foundation initiative. The exact announcement title is not specified in the draft source information.
• OpenAI repository, “README.” The README date or repository snapshot is not specified in the draft source information.
• ^[3] Advisory Group on Mathematics and Artificial Intelligence (AGMAI), statement on the release; passages on the advisory role and the ongoing process of human understanding and incorporation. The statement date and exact title are not specified in the draft source information.
• ^[4] AGMAI guidance on author understanding, verification of mathematical arguments, and responsibility for paper contents. The document date and exact title are not specified in the draft source information.
• Simons Foundation announcement.