OpenAI on Tuesday announced what it described as three "major" discoveries produced with the help of its AI models — one each in mathematics, computer science and physics — less than a month after the company publicly released hundreds of manuscripts it claimed contained solutions to long-standing open problems in mathematics. The announcement, confirmed across multiple outlets including The Verge, Scientific American and Unite.AI, marks the company's latest and most ambitious attempt to position its models as genuine participants in scientific discovery rather than mere assistants.

What distinguishes this batch from the previous release, according to OpenAI, is not the results themselves but the process behind them. The company said external experts — which it calls "shepherds" — guided the computer science and physics results, reviewing and steering the AI's output before publication. For the mathematics result, however, OpenAI said those advisers recommended a different path: releasing the raw AI output directly so that researchers could work through their own explanations and verifications.

The Results, and What Was Actually Released

OpenAI added the outputs to GitHub on Tuesday. One of the artifacts visible in the release is a result titled "Local smoothing for the Euclidean wave equation," a problem in harmonic analysis concerning how solutions to wave equations behave when averaged over time and space. The proof, as posted, leans on an auxiliary partial differential equation argument — a structure that mathematicians will now need to audit line by line.

The company characterized the three results as spanning distinct disciplines, but it has not yet published full, peer-reviewed write-ups for all of them, leaving the broader research community to assess whether the outputs constitute new theorems, novel proof strategies, or merely plausible-looking derivations that collapse under scrutiny.

A Field Already 'in Shock'

The timing is hard to separate from the controversy that preceded it. Scientific American framed the new announcement against a backdrop of a discipline still reeling from OpenAI's earlier mass release — hundreds of manuscripts dumped on mathematicians in a single batch, many of them claiming to resolve well-known open questions. That earlier episode was widely described as leaving the field "already in shock," with researchers scrambling to triage an unprecedented volume of machine-generated mathematics.

Scientific American also reported that AI had tackled an 80-year-old "Erdős problem" — one of the famously stubborn open questions associated with the prolific Hungarian mathematician Paul Erdős — with mathematicians described as "amazed" by the result. Erdős problems carry particular weight in the field: Erdős himself famously placed cash bounties on unsolved questions, and a solution to one that has stood for eight decades would be a significant event regardless of whether a human or a machine produced it.

The shift toward "shepherds" signals that OpenAI is aware the credibility bottleneck is no longer generation — it is verification.

Why 'Shepherds' Matter

The introduction of expert shepherds is a notable concession from a company that has generally emphasized raw model capability. In computer science and physics, the shepherding model resembles a hybrid workflow: the AI proposes, and credentialed humans prune, redirect and validate. That is a meaningful departure from the previous approach, in which outputs were released en masse with the implicit invitation for the community to sort out what held up.

For mathematics, the decision to release output without a shepherd — on the advice of the very experts OpenAI consulted — reflects a different theory of validation. Rather than presenting a curated result, the company is effectively handing proofs to the community and asking researchers to construct their own explanations. It is an approach that some mathematicians may welcome as transparency and others may read as offloading the hardest work of formalization onto unpaid reviewers.

The Verification Problem

The central tension in all of this is not whether AI models can generate mathematical text that looks like a proof — they demonstrably can — but whether those proofs are correct, novel and meaningful. Automated theorem provers such as Lean and Coq offer one route to certainty, and a growing number of mathematicians have pushed for AI-generated results to be formally verified before being announced as discoveries. OpenAI's releases do not uniformly include such formalization.

There is also a sociological dimension. Journals, conferences and tenure committees are built on attribution and peer review, and a firehose of machine-generated results strains all three. Researchers have raised concerns about the sheer labor of triage, the risk of errors propagating into the literature, and the question of how credit should be assigned when a model produces a proof that humans then clarify.

What Happens Next

OpenAI's posture suggests a strategy of moving fast and letting the community catch up. The company has now made two substantial forays into pure research in under a month, and the pattern implies more to come. Whether that pace proves to be a genuine acceleration of science or a publicity engine that outruns peer review will depend on what happens in the coming weeks, as mathematicians dig into the GitHub repository and physicists and computer scientists assess the shepherded results.

For now, the most defensible reading is this: OpenAI has demonstrated, again, that its models can produce sophisticated scientific content at volume. It has also acknowledged, by introducing human shepherds, that volume is not the same as discovery. The field's verdict — on the Erdős problem, on the wave equation proof, and on the collaboration model itself — is still being written, and it will not be written by a machine.

• What happened: OpenAI announced three AI-assisted results in math, CS and physics, posted to GitHub.
• What's new: External "shepherds" guided the CS and physics work; math output was released raw for researchers to explain.
• Context: Comes weeks after OpenAI released hundreds of math manuscripts, unsettling the field.
• Open question: Whether the results survive formal and peer verification.