Skip to main content
← Back to Blog
Cool Tech8 min read

OpenAI’s New Science Claims Expose a Fight Over Who Gets to Verify Them

OpenAI released three claimed AI discoveries using three verification models, exposing a dispute over evidence access, expert selection, credit, and trust.

OpenAI announced three claimed scientific breakthroughs on October 26, 2026, then assigned each a different route toward verification. A harmonic-analysis proof was released directly with a Lean formalization. A claimed refutation of the Exponential Time Hypothesis was released with an external working group asked to examine and explain it. A quantum-field-theory result went to a selected group while the underlying AI-produced materials remained private.

The claims are consequential. If they survive scrutiny, they would resolve major questions in mathematics, theoretical computer science, and physics. They are not yet settled discoveries merely because OpenAI announced them or because two have formal artifacts. The more immediate development is procedural: after releasing hundreds of AI-generated manuscripts earlier this month, OpenAI is testing three competing ways to turn machine output into accepted scientific understanding.

That experiment has already produced a public disagreement. OpenAI says expert “shepherds” can organize verification and explanation. Its independent mathematics advisers argue that labs should not designate particular researchers for privileged roles, and they called for the withheld physics materials to be released immediately. The dispute is about more than publication etiquette. It concerns who gets early access, whose reputation validates a claim, how credit is assigned, and whether a private company should organize the scientific community around results generated by a model only it can use.

Three claims, three disclosure models

The mathematics result concerns Sogge's local smoothing conjecture for waves. OpenAI says its model produced a proof in every spatial dimension of at least two, with consequences for the Kakeya and Stein restriction conjectures. The company dates the model output to October 18 and says a Lean formalization was completed the next day. It published both the natural-language writeup and formal proof for open inspection.

In computer science, OpenAI says its model found a deterministic, subexponential-time algorithm for exactly counting satisfying assignments of fixed-width Boolean formulas. The three-variable case would refute the widely used Exponential Time Hypothesis, or ETH. The public repository includes the manuscript and a Lean artifact whose documentation states the formalized runtime and space bounds. Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams are leading an independent group asked to check, explain, and contextualize the result.

The physics claim is broader and less inspectable. OpenAI says its models produced existence proofs for four-dimensional interacting quantum field theories containing gauge fields and fermions, including quantum chromodynamics and a completion of the Standard Model. The company says the Yang–Mills proof was formalized by October 16. A group led by Nobel laureate David Gross will study the work, but OpenAI has not released the AI output or Lean materials.

Those differences create three evidence states. The first two can be inspected now, although inspection is not the same as consensus. The third asks the public to rely on a company summary and the future work of a selected group. A headline that treats all three as equally verified erases the most important fact about the release: access to the evidence is uneven.

Lean checks a proof, not a scientific institution

Formal verification is a strong form of evidence. Lean can check that a formal statement follows from its definitions, axioms, and imported libraries. The ETH documentation, for example, specifies the encoded problem, algorithmic bounds, and comparator theorem rather than merely asserting that a famous conjecture has fallen.

But a checked formal object does not answer every question a scientific community must ask. Experts still need to confirm that the formal statement matches the natural-language claim, that the assumptions capture the intended problem, that the supporting libraries are appropriate, and that the result is genuinely new. They must also determine significance, explain the ideas, locate relevant prior work, and decide how credit should be distributed. Formalization narrows the verification problem; it does not eliminate it.

OpenAI itself describes the manuscripts as raw research data. That is the right caution. The company'sOctober 7 release of hundreds of mathematical manuscriptsalready showed that generation can outpace the community's ability to read and understand the output. The new release is a material follow-up because it attempts to design an institution around that bottleneck.

The advisers reject privileged shepherding

OpenAI says its collaborations will be transparent, voluntary, unpaid, and independent. Participants may discuss intermediate work and publish their own conclusions. OpenAI says it will claim no academic credit for model-generated discoveries, while researchers should receive recognition for verification, explanation, corrections, and extensions.

The Advisory Group on Mathematics and Artificial Intelligence, or AGMAI, reaches a different conclusion about selecting the initial researchers. In an October 25 statement, the independent group said designated experts can receive an unfair head start, discourage other researchers, reflect the company's networks and priorities, and lend their reputations to claims before verification is complete. It recommends public scholarly repositories, recorded revisions, disclosure of how problems were selected and attempted, formal proofs where practical, and research support distributed through independent nonprofit institutions.

On October 26, AGMAI specifically urged OpenAI to publish the Yang–Mills, quantum chromodynamics, and Standard Model materials without delay. The group noted that OpenAI said the formalization was complete and argued that understanding should develop with equal access. OpenAI, by contrast, says the physics collaborators chose to withhold the AI artifacts while they prepare their own account.

Neither approach is free of tradeoffs. Immediate release can swamp a field with difficult material, create races to claim follow-on results, and leave errors circulating without a clear explanatory authority. Controlled access can produce a more coherent first review, but it concentrates opportunity and lets the lab choose who gets to shape the initial interpretation. OpenAI acknowledges that selecting shepherds creates information asymmetry. The unresolved question is whether that cost is necessary.

Verification needs durable provenance

AI-generated science needs more than a PDF and a verdict. Each public claim should remain connected to the exact model output, formal statement, software environment, generation date, problem-selection record, literature search, reviewer reports, corrections, and human explanations. If a theorem changes, the old version must remain citable. If experts disagree, the record should preserve the disagreement instead of compressing it into a single confidence label.

That is also a memory problem. A durable research record must distinguish what was generated, what was formally checked, what humans understood, what was later contradicted, and what evidence remains private. Retrieval without those states will repeatedly surface the strongest-sounding claim rather than the best-supported one.

OpenAI's new release makes the verification bottleneck visible, but it does not solve it. The open mathematics and computer-science artifacts can now be tested by anyone with the necessary expertise. The physics claim cannot. Until independent researchers complete that work, the honest status is three extraordinary claims at different stages of disclosure—not three finished scientific discoveries.

Sources

Your AI remembers everything. Everywhere.

Unimatrix gives you a shared, durable memory layer across Claude Desktop, Cursor, ChatGPT, and Gemini. Setup in 2 minutes. Free and paid plans available.

Looking for developer resources?

Browse our catalog of 500+ tested AI prompt profiles covering DevOps, data modeling, agent behaviors, and API wrappers. Have a prompt to share? Submit your own for review by our librarian to be featured. Completely free, no registration required. Browse prompt libraries →

Keep reading