OpenAI's New Math Advisors Are Its Best PR Move Yet

The Audit That Looks Like a Concession

OpenAI announced on September 21 that it has formed an independent advisory group of nine mathematicians, headquartered at the Institute for Advanced Study in Princeton. Among them are Fields Medalists Timothy Gowers of Cambridge, Edward Witten and Martin Hairer of IAS. The group’s official name—the Advisory Group on Mathematics and Artificial Intelligence—sounds like an academic body. It is not.

OpenAI also revealed that an internal model, whose training began on August 28, has solved more than 100 unsolved problems across nearly every branch of mathematics. Among them is the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems with a $1 million reward attached. The company had previously disclosed in August that a model codenamed “Astra” had solved ten such problems. The new figure is ten times larger.

This is where the move is clever. OpenAI is handing the math community exactly what it demanded—external scrutiny—while retaining full control over when, how, and whether those results ever see peer review.

The Backlash That Forced the Hand

The advisory group did not emerge from a vacuum. In a public letter titled “A Severe Misalignment of AI in Mathematics,” twenty-five Fields Medalists warned that the field was being pulled into a trap. The concern was not that AI would solve hard problems. It was that the act of solving them—without proof, without context, without engagement with existing literature—would become the metric by which AI progress is measured.

Mathematics is not a sport. A correct answer to an open problem is worthless without a verifiable argument. When AI systems begin producing solutions that cannot be checked, the entire knowledge-transfer mechanism of the discipline breaks. That is what the mathematicians were protesting.

OpenAI heard the objection. It simply reframed the response as collaboration rather than compliance. The distinction matters because it shifts the narrative from defense to openness—and OpenAI has spent years building that narrative into its brand.

What the Group Can and Cannot Do

Here is the fine print that matters. The advisory group will advise on how to evaluate and publish OpenAI’s mathematical outputs. It will weigh in on research norms and whether tools should support learning or research. But OpenAI explicitly stated that the group bears no responsibility for advising on the pace of the company’s internal math research.

The members receive no compensation from OpenAI. They can offer unsolicited advice. They can publicly critique the company’s impact on mathematics. These are real freedoms. But none of them affect the core question: who decides whether a claimed solution is actually a solution?

The answer remains OpenAI. The advisory group can say whether the company should release its findings, and how. It cannot say whether the findings are true. That is the distinction that turns this from accountability into a legitimacy machine.

Consider the structural asymmetry. OpenAI controls the data, the models, and the timeline. The advisory group operates at arm’s length—it can observe and comment, but it cannot independently verify the outputs through traditional channels, since OpenAI has not agreed to share raw model artifacts or training pipelines. This is not an oversight; it is a deliberate boundary drawn by the company itself.

Who Wins, Who Loses

OpenAI wins the most. It gets an endorsement architecture without losing steering control. The moment the group—especially its high-profile members—begins validating specific results, every claim carries institutional weight that did not exist before. Even cautious language from the group will function as de facto signal boosting. A sentence like “we find these results worthy of further investigation” from Timothy Gowers carries disproportionate gravity in a field where his opinion alone can shift funding and hiring patterns.

The mathematicians on the panel gain a platform, but also a risk. Their names now sit adjacent to claims that the broader community has already flagged as potentially hollow. If the group endorses results that later fall apart under scrutiny, the reputational damage will not be confined to OpenAI. It will ripple across the very people who trusted their judgment. There is a growing awareness of this dynamic among some potential participants—several senior figures were reportedly approached but declined, citing the conflict between academic independence and corporate proximity.

The wider math community gets a channel it asked for, but one that runs in only one direction. OpenAI has not agreed to submit its outputs to traditional peer review. The advisory group is an alternative pathway—one that is faster, more controlled, and fundamentally asymmetric. The company produces. The group advises on packaging. The community receives.

There are also second-order effects to consider. University departments that have struggled to attract graduate students in pure mathematics may now face pressure to incorporate AI-generated results into their curricula—not because the results are verified, but because the market demands it. Funding bodies may quietly adjust their priorities. Postdocs who spend months wrestling with a single theorem may find their career trajectories undervalued against a system that produced a hundred solutions in two months. The advisory group does not create these pressures, but it will be seen as part of the ecosystem that normalizes them.

What Happens Next

The immediate test is publication. The group has set up a form to collect input from mathematicians on how to handle what OpenAI describes as a large volume of important results. That is the operative word: important. OpenAI gets to classify its own outputs. The advisory group advises on the timing and framing. The rest of the field is invited to comment through a web form.

If OpenAI’s claims survive even a modest round of independent verification, the precedent is enormous. Solving a Millennium Prize Problem would be the single most visible AI milestone in mathematics since DeepMind’s work on protein folding. It would reshuffle the hierarchy of what counts as intellectual achievement and force a reckoning about the relationship between human expertise and machine generation. The $1 million prize from the Clay Mathematics Institute—still awarded to humans—would become a symbolic relic if a machine claimed the solution first.

If the claims do not survive scrutiny, the advisory group becomes a case study in how institutional credibility can be borrowed without earning it. We have seen this pattern before in other domains. Scientific advisory boards captured by industry have lent false precision to contested findings, from tobacco to pharmaceuticals. The difference here is that the stakes are epistemic rather than commercial—the truth of mathematics is not negotiable, and the field will eventually sort signal from noise, regardless of how polished the packaging.

The model behind these results has been training since late August. That is roughly two months. Claims of over 100 solved open problems in that window are not just ambitious—they are outside anything the field has seen from any source, human or machine. The onus now shifts entirely to the advisory group to produce a public account of what has actually been verified, what remains unverified, and what the company is choosing not to share.

The line between breakthrough and press release has always been thin in AI. OpenAI just built a mathematician’s committee to make it thinner.