Scientific research in the AI era: the validation dilemma

OpenAI announced, on October 26, 2026, three results produced by its internal models in mathematics, computer science, and physics. This time, the approach was different: instead of dumping hundreds of manuscripts on the academic community as it did on October 6, the company hired external experts — which it called "shepherds" — to guide the dissemination of each discovery. The response was not the expected relief. Instead, it revealed a deeper conflict over who has the right to access, interpret, and validate scientific knowledge produced by machines.

What was discovered

The three results are substantial. In harmonic analysis, OpenAI claims to have proved Sogge's local smoothing conjecture (1991) in all spatial dimensions — a result that, if valid, would imply versions of the Kakeya and Stein restriction conjectures, extending earlier findings limited to more restrictive contexts. The company points to a milestone in two dimensions (2019) and another in the Kakeya conjecture in three dimensions (2025) as prior milestones.

In computational complexity, the model found a disproof for the Exponential Time Hypothesis (ETH) — a widely used assumption in computer science theory that states that satisfiability problems with fixed Boolean formulas cannot be solved in subexponential time. The result was formalized in Lean on October 17.

In theoretical physics, OpenAI claims to have produced an existence proof for four-dimensional quantum field theories, including quantum chromodynamics (QCD) and pure Yang–Mills theory. Unlike the first two, the physics result has not been released.

The new "shepherd" model

OpenAI adopted three distinct dissemination strategies. For the mathematics result, it published the AI-generated output (including natural-language manuscripts and Lean code) and left it to the research community to organize its own interpretation efforts. For computer science, the company convened an external group to explain the formalized result. For physics, the dissemination was postponed: OpenAI deferred to the external expert group on how and when to make the materials accessible.

This hybrid model was presented by the company as an evolution in response to criticism from the academic community. At the October 6 briefing, when it released 722 AI-generated math manuscripts — some claiming solutions to open problems — the reaction was immediate and negative. Academics pointed out that unreviewed results, released in bulk, created an overwhelming burden for already exhausted communities. The Simons Foundation itself responded with a $15 million joint effort (with participation from Anthropic and OpenAI) to create analytical capacity.

This time, OpenAI says it is experimenting. But experts are not convinced.

Researchers push back

Constantin Kogler of the Institute for Advanced Study in Princeton praised the volume reduction — three results instead of 722 — but questioned the logic of manually selecting which scientists would get privileged access to results before the broader community. "I am fine with them releasing the problems as they did today," Kogler said, but added that public announcement should, in principle, mean open access for any researcher to examine the work.

Andreas Karch of the University of Texas at Austin went further: allowing large labs to choose their own reviewers creates an unhealthy concentration of power. For Karch, the proofs or underlying outputs should be made public immediately, regardless of whether an external advisory group is also formed.

The Advisory Group on Mathematics and Artificial Intelligence — a group OpenAI had pledged to consult — issued a warning the day before the announcement. The group stated that selective early access is unnecessary, risks biased selection and delays, borrows the reputations of chosen experts, and should be avoided in favor of open academic repositories, formal proofs, clear problem selection details, and support for broader verification. The day after the announcement, the group urged OpenAI to publish without delay the natural-language and Lean outputs behind the Yang–Mills and QCD results.

The control paradox

There is something peculiar about this debate. OpenAI is producing results that the scientific community classifies as potential breakthroughs — advances that could rewrite mathematics textbooks, refute fundamental computational hypotheses, and confirm century-old physics theories. And at the same time, the company is restricting access to these results through a validation process controlled by itself.

This paradox is not new. Since AI models began producing plausible scientific results — the AlphaFold case for protein folding in 2021 being the most successful example — the question of how to validate machine-generated discoveries has become central. What distinguishes the current situation is scale: OpenAI is not producing a single notable result. It is continuously producing results that, at least formally, appear correct.

The irony: the very tool that enables OpenAI to produce these results — models capable of formal reasoning in mathematics and physics — is the same one that makes traditional validation more necessary than ever. If a model can write proofs that seem correct to limited experts, why would we trust a validation process that is even more limited?

The academic community has already demonstrated that it can handle massive volumes of preprints — arXiv is the proof. The problem is not quantity, but quality and legitimacy of the process. And a process in which a private company selects who can validate its own results sounds less like science and more like marketing.

Sources: The Verge, HyperAI, MLQ News

✓ Independent sources cross-checked and verified before publishing