Computer scientist Scott Aaronson, writing on his blog on September 15, 2026, dropped a claim that's more consequential if true than almost anything else in this week's AI news cycle — and he was careful to frame it exactly as what it is: a rumor. "According to rumors that I've heard," Aaronson wrote, AI companies — having been "burned by the hostile response to the Navier-Stokes proof" — are now sitting on solutions to some very major open problems in theoretical computer science, "until they figure out a better way to handle things."
Wharton professor Ethan Mollick amplified the claim the same day, calling it "from someone with inside knowledge" and "plausible," while framing the real question correctly: if the norms of sharing become strained, will labs start hoarding knowledge to avoid PR issues? Neither Aaronson nor Mollick named a lab, a specific problem, or a primary source. This is worth covering not because it's confirmed — it isn't — but because the underlying dynamic it describes is real, checkable against recent events, and matters regardless of whether this specific rumor holds up.
TL;DR
| Question | Answer |
|---|---|
| What's the claim? | AI labs may be withholding solutions to major open problems to avoid a repeat of the Navier-Stokes backlash |
| Who said it? | Scott Aaronson, relaying a secondhand rumor; amplified by Ethan Mollick as plausible |
| Is it confirmed? | No — no lab, problem, or primary source has been named or has confirmed this |
| What triggered the theory? | OpenAI's Navier-Stokes announcement became a public credit dispute, followed by 25 Fields Medalists' "Severe Misalignment" declaration |
| What's the counterargument? | Mathematician Shuohao Liao: coordinated release (claim + proof artifacts + independent verification together) beats secrecy as a norm |
| Why cover an unconfirmed rumor? | Because the incentive structure it describes is real and checkable, independent of whether this exact claim is true |
The chain of events that makes this rumor plausible
To evaluate whether this claim is worth taking seriously, it helps to lay out exactly what actually happened before it, all of which is confirmed:
- OpenAI announced a claimed Navier-Stokes solution on September 8, 2026, produced by roughly 10,000 coordinating AI agents over 88 hours — covered here in detail.
- NYU mathematician Tristan Buckmaster publicly disputed the announcement's framing within a day, alleging OpenAI's effort was triggered by rumors of his own private research with Anthropic researcher Levent Alpöge, and that OpenAI misrepresented how independent its result actually was.
- 25 Fields Medalists, including Terence Tao, published "A Severe Misalignment of AI in Mathematics" on September 11 — a declaration arguing that AI labs treating famous open problems as PR benchmarks, without proper attribution or human understanding of the resulting proofs, damages the mathematical community's core process.
- Aaronson's rumor, published September 15, claims labs are now responding to that exact backlash by going quiet on further results rather than fixing the underlying process issues the Fields Medalists raised.
Read in that sequence, the rumor isn't a wild leap — it's a plausible next chapter in a story that's been escalating in public for over a week. Whether it's true is a separate question from whether it's coherent with the incentives already on display.
Why this would be a genuinely bad outcome, even for AI labs
The most direct response to Aaronson's rumor, quoted in the same discussion thread, came from mathematician Shuohao Liao: "Withholding results because the reaction may be messy is a bad equilibrium. A better norm is coordinated release: claim, proof artifacts and independent checks at the same time." That's a sharp diagnosis of exactly what went wrong with the Navier-Stokes announcement — OpenAI published a headline claim before the verification, attribution, and process questions were settled, which is what produced the public dispute in the first place. Liao's proposed fix isn't secrecy, it's the opposite: releasing the claim together with the artifacts and independent checks that would head off exactly the credibility fight that followed.
If labs actually did start hoarding solved problems rather than fixing the release process, it would plausibly make the underlying trust problem worse, not better — the mathematical community would have even less visibility into what AI systems can actually do, at exactly the moment public trust in AI-generated math claims is most fragile. Several replies to Mollick's post picked up on this directly, with one commenter noting there would be little practical benefit to sitting on results for long, since "it's only a matter of time until an individual is able to make similar advances" independently — and another (more cynically) observing that "the frontier is gated by PR risk now."
The honest epistemic status of this claim
It's worth being unusually explicit about what we actually know here, because the whole story rests on secondhand information:
- Aaronson is a credible source — a working theoretical computer scientist with direct professional ties to complexity theory and quantum computing, not a random commentator. His own framing ("according to rumors that I've heard") is appropriately hedged.
- No lab has confirmed this. OpenAI, Anthropic, Google DeepMind, and other frontier labs have not publicly stated they're withholding any specific solved problem as of this writing.
- No specific problem has been named, beyond Aaronson's parenthetical that it's "not P≠NP or other complexity class separations, but think about some of our other biggest problems" — vague enough to cover a wide range of possibilities, or nothing concrete at all.
- This could plausibly be true, exaggerated, or entirely mistaken — rumors relayed secondhand about internal lab decision-making have a mixed track record, and "sitting on a result to manage PR" is also a claim that's nearly impossible to falsify from the outside (there's no way to prove a lab isn't sitting on something).
Treat this as a serious question worth watching, not a settled fact to repeat as news. The healthiest response is the same one Aaronson himself seems to be modeling: take the claim seriously enough to think about its implications, while being explicit that it hasn't been verified.
What this means for AI research norms going forward
Regardless of whether this specific rumor holds up, the underlying tension it names is real and will keep recurring: AI labs increasingly have both the capability to produce headline-grabbing research results and a strong commercial incentive to control how those results are announced and attributed. The Navier-Stokes dispute showed what happens when that incentive collides with the academic norm of careful, collaborative credit — and the Fields Medalists' declaration was explicitly a call for AI labs to adopt something closer to the coordinated-release norm Liao describes, rather than unilateral announcement followed by damage control.
Whether labs respond to that pressure by improving their release process (the outcome mathematicians are asking for) or by simply going quiet (the outcome Aaronson's rumor describes) is a genuinely open question — and one with real consequences for how much the outside world can trust or verify AI-generated research claims in any field, not just mathematics.
The broader context: Aaronson's essay wasn't only about this rumor
It's worth noting the rumor about hoarded results was a small part of a much longer essay, titled "The Age of Wonders and Terrors," in which Aaronson argued more broadly that AI has crossed a threshold in mathematical research that he — a self-described skeptic for most of his career — no longer feels he can dismiss. He pointed to a cluster of results beyond Navier-Stokes: a disproof of the Jacobian conjecture, improved bounds on Grothendieck's constant, a Lean-verified proof of Fermat's Last Theorem, and progress on several open problems in quantum complexity theory that Aaronson himself has worked on. His framing was that the volume and difficulty of these results, taken together, represents a genuine shift in what AI-assisted mathematics can do — independent of any single result's disputed provenance.
The hoarding rumor sits inside that larger argument as a specific, practical worry: if AI capability really has reached this level, and the reputational fallout from announcing results poorly is real, the rational response for a lab optimizing for its own PR risk might indeed be silence rather than a better process — which is precisely the "bad equilibrium" mathematician Shuohao Liao warned against. Whether labs choose transparency or silence from here is likely to shape how much independent verification and trust the field can maintain as AI-assisted results keep arriving.
FAQ
What did Scott Aaronson actually claim? That AI companies, per rumors he's heard, are sitting on solutions to major open problems to avoid a repeat of the Navier-Stokes backlash — with no lab, problem, or source named.
Is this confirmed by any AI lab? No — no lab has confirmed withholding any specific solved problem as of this writing.
Why would a lab withhold a solved math problem? Stated logic is reputational risk management, following the public credit dispute and Fields Medalists' criticism after the Navier-Stokes announcement.
What would be lost if labs actually did this? Reduced trust and visibility into AI capability, and a missed opportunity to fix the actual process problem (rushed, uncoordinated announcements) that caused the backlash in the first place.
How does this connect to the Navier-Stokes dispute? Directly — Aaronson cites it as the specific triggering event for the rumored shift in disclosure behavior.
Should this rumor be treated as reliable? No — treat it as a credible, worth-tracking claim, not established fact, given its secondhand and unconfirmed nature.
Related reading
- OpenAI's Navier-Stokes proof is now a credit and data dispute
- 25 Fields Medalists accuse AI labs of "severe misalignment" in math
- OpenAI used 10,000 AI agents to propose a solution to a 90-year math problem
- Did Claude solve Navier-Stokes? The Millennium Prize rumor, fact-checked
- Millennium Prize Problems and AI: what's actually been solved, fact-checked
- Will AI replace mathematicians?
- Official: Scott Aaronson's Shtetl-Optimized: "The Age of Wonders and Terrors"
This piece covers an explicitly unverified, secondhand rumor as reported by Scott Aaronson on September 15, 2026, and amplified by Ethan Mollick. No AI lab has confirmed the underlying claim; treat it as a credible question worth tracking, not as established fact.