The science you can't see
The scientific community is in the midst of an unprecedented transparency crisis. Last Monday (October 26), OpenAI announced three heavyweight discoveries — one in mathematics, another in quantum physics, and a third in theoretical computer science — but rather than releasing the results for public scrutiny, the company introduced a system of handpicked experts to decide how and when each result would be shared. Researchers called this scheme gatekeeping and said it undermines fundamental principles of open science.
The three discoveries are substantial. The first claims to resolve the Kakeya conjecture in all dimensions, a harmonic analysis problem open since 1917. The second proposes refuting the Exponential Time Hypothesis (ETH), a nearly three-decade-old assumption stating that NP-complete problems like 3-SAT require exponential time to solve — if true, the refutation would have implications for all of computational complexity theory. The third, a rigorous existence proof for quantum field theories with gauge fields and fermions in four dimensions, including Quantum Chromodynamics (QCD) and a unified completion of the Standard Model of particle physics.
But what bothers the academic community most is not the results themselves — it is the mechanisms used to disclose them. OpenAI called these experts "shepherds." Each discovery received a different release strategy, determined by the shepherds: the math result was released as raw AI output on GitHub, leaving researchers to mount their own efforts to explain it; the computer science result will have an external group of collaborators to interpret it, while the AI output is available for anyone to examine; the physics result, however, remains restricted, with OpenAI deferring to an external group of experts on how to release the materials.
The AGMAI letter
The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), based at the Institute for Advanced Study, issued a direct warning against the shepherd model. In a statement to The Verge, the group argued that selecting certain people to receive results first is "unnecessary" and risks "unequal access, biased selection, and delays." AGMAI's recommendation is clear: release the output on scholarly repositories with formalized proofs in Lean, not privileged selections of researchers.
Constantin Kogler, a researcher at the Institute for Advanced Study himself, agreed that consulting external researchers is positive but questioned the decision to restrict access to experts chosen by OpenAI. "I think it's unnecessary," he told The Verge. Andreas Karch, a physics professor at the University of Texas at Austin, suggested that researchers should be selected through a competitive process to ensure representativeness.
Jared Duker Lichtman, a Stanford mathematician, acknowledged that "finding an approach that satisfies everyone may be impossible," but argued that releasing findings openly "could lead to better science and promote wider adoption of new tools like AI."
The legacy of preprint dumps
The frustration with the shepherd model did not emerge in a vacuum. It is the continuation of a series of incidents that have been eroding trust between OpenAI and the mathematical community. In October, OpenAI released hundreds of AI-generated preprints onto academic repositories without prior notice — an event that MIT professor Ryan Williams described as a "DDoS against mathematicians." The avalanche of unvetted articles forced entire departments to deal with dozens of unverified results that would take years for rigorous analysis.
Before that, on October 6, OpenAI had released a massive collection of AI-generated mathematical results, offering $1 million for researchers who could verify the models. AGMAI praised the release but warned that "the future of mathematical research cannot consist only of understanding results produced by AI labs."
OpenAI acknowledged the trade-offs in its official announcement. "On the one hand, collaborations empower experts to unpack results and contextualize them in the academic community. On the other hand, choosing initial shepherds necessarily shapes who gets involved first and creates asymmetry of information," the company wrote, adding: "We welcome feedback on how to broaden participation and make this process more useful."
What this means for industry
For companies that depend on frontier AI, OpenAI's transparency crisis raises serious practical questions. Compliance and procurement teams that rely on vendor-disclosed benchmarks to make risk decisions now have less basis to trust those disclosures. When the same lab that announces a capability also controls who can verify it, external validation is structurally compromised.
The EU AI Office and other regulatory bodies are assessing whether self-governed disclosure practices by frontier labs are compatible with independent audit requirements under the EU AI Act and frameworks like H.R. 8094 — the AI Foundation Model Transparency Act of 2026. The question is not just academic: companies that license these models for high-risk tasks need to know whether declared capabilities have actually been verified by independent peers.
Meanwhile, the scientific community continues navigating uncertain terrain. OpenAI itself admits that its results are "best understood as a form of raw research data" — meaning discoveries that have not yet been validated, contextualized, or incorporated into established knowledge. Until then, the tension remains: on one side, private labs producing results that transform entire fields; on the other, an academic community insisting that real science demands open scrutiny, not gated ports.
Sources: The Verge, OpenAI, TechCrunch
✓ Independent sources cross-checked and verified before publishing
