FULL STORY

OpenAI's Millennium Prize Claim Sparks Theft Allegations and Backlash

After OpenAI claimed partial progress on the Navier-Stokes millennium problem, mathematicians accused it of training on private conversations, which OpenAI denied. Authorship disputes and a wider ethics debate followed.

2026-09-09 ~ 2026-09-11 · 7 episodes · 34 posts

Episode 1 · Mathematicians Accuse OpenAI of Racing to Claim Their Unfinished Proof (2026-09-09, 6 posts)

In early September 2025, accusations that OpenAI "stole" mathematicians' in-progress work ignited debate across the math and AI communities, turning into a dual discussion about the reality of AI math capabilities and research ethics. OpenAI has not formally responded.

Confirmed

  • A math professor posted on LinkedIn that "OpenAI may have stolen another major proof," sparking widespread controversy (@lowkeysciguy); Jürgen Schmidhuber amplified the story (@hardmaru).
  • Mathematician Leven Alpöge (@alpoge) published the points he raised to OpenAI on the night of September 2: he claimed intelligence indicated OpenAI had learned of his ongoing work and assembled a competing team days earlier; he stressed it was a rigorous personal collaboration lasting a year (@mvpatel2000).
  • Per @新智元, NYU mathematician Tristan Buckmaster and Leven Alpöge's team had spent months using AI to attack the Navier-Stokes equation and other Millennium Problems; they publicly accused OpenAI of learning their solution path "in the final days" and marshaling massive compute to race ahead.
  • Mathematician littmath rebutted the theft claims point by point, arguing the evidence is very weak and such narratives mainly serve to deny rapid model progress, which he called dangerous for the health of mathematics; AI×math success is real (@littmath, @AashishReddy).

Unconfirmed

  • The claim that OpenAI reviewed Codex session logs tagged Navier-Stokes to find the most promising approach remains rumor.
  • Whether OpenAI actually accessed or exploited logs or ideas, or built a competing team, is unverified pending an official response.
  • Terence Tao's reported criticism that "AI destroys open science" is based on secondhand accounts.

Why it matters

  • The case touches core disputes over AI coding tools' data usage, privacy, and IP boundaries: if platforms can inspect and exploit researchers' accumulated ideas, academic trust in AI tools will be shaken.
  • littmath's rebuttal represents a counterpoint—denying model progress via "theft" narratives is equally dangerous. The outcome will shape how mathematicians and AI labs cooperate going forward.

Episode 2 · Anthropic Faces Ethics Backlash Over Math Proof and Scooped Collaboration (2026-09-09, 2 posts)

Anthropic admitted using an internal model in its Euler blowup proof, drawing skepticism, while mathematician Alpoge accused the lab of racing to scoop his year-long personal collaboration—sparking an AI ethics controversy.

Episode 3 · Mathematician Thom accuses OpenAI of training Astra on his private chats to crack Gromov conjecture (2026-09-10, 9 posts)

Mathematician Andreas Thom published a long post on Mastodon listing multiple pieces of evidence, accusing OpenAI of possibly using, without authorization, his conversations with collaborator Gábor Kun about the Gromov soficity conjecture to train the Astra model — the very conjecture at the core of the non-sofic group problem OpenAI later announced Astra had solved. The accusation was relayed and spread publicly by mathematician Valerio Capraro; OpenAI has not yet responded.

Confirmed

  • Thom publicly posted on Mastodon, listing multiple pieces of evidence that OpenAI may have used his research conversations with Gábor Kun about the Gromov soficity conjecture when training Astra (per m3, m5, first-hand relayed by Capraro)
  • Valerio Capraro relayed and amplified the accusation on X, with accounts like Gary Marcus retweeting
  • Quant researcher henryquantum shared it and called on OpenAI and other frontier labs to seriously and transparently clarify data sources (per m2)

Not Yet Confirmed

  • Whether OpenAI actually used Thom's conversation data to train Astra remains supported only by the "multiple pieces of evidence" Thom unilaterally listed; the accusation itself uses wording like "possibly" and "suspect"
  • OpenAI has not publicly responded

Why It Matters

  • The incident touches the core controversy of AI training data compliance: whether users' private conversations or unpublished research were used for training, with the resulting breakthrough then claimed by the lab
  • If true, it would involve a double issue of intellectual property over unpublished academic work and user privacy, which is why multiple researchers and prominent scholars (such as Gary Marcus) amplified it, and public pressure is mounting

Episode 4 · Mathematician littmath criticizes AI labs over scooping and authorship disputes (2026-09-10, 6 posts)

The authorship dispute over the Navier–Stokes breakthrough continues to simmer, with several scholars discussing the research culture of AI labs. Mathematician Daniel Litt (littmath), responding to users' questions, clarified his position: his core concern about the "AI stealing work in progress" allegation is not theft itself, but the culture of scooping—racing to publish after learning others have results—which incentivizes secrecy, harms scientific openness, along with negotiations over paper authorship.

Confirmed

  • littmath said his criticism targets the overall conduct of both OpenAI and Anthropic across a series of past results, each release accompanied by its own racing dynamics and controversies.
  • littmath clarified he wasn't saying the labs are darker than the median corporation, but "equally dark"—which by academic standards is already outrageous, though academia has its own flaws.
  • User AashishReddy argued that OpenAI's publishing results first was not about stealing researchers' credit, but about staying at least on par when Anthropic announced—reportedly, OpenAI released early when it believed Alpöge/Buckmaster already had the result.
  • Math blogger grotsen posted a timeline of known AI-math result disputes so far, showing the controversy has spread across multiple cases.
  • OpenAI researcher rbhar90 quote-tweeted that if the bulk of the intellectual contribution to a solution came from human researchers who go uncredited, the top credit should not go to the AI.

Why it matters

  • If scooping culture becomes the norm, it will incentivize researcher secrecy and undermine academic openness, clashing with mathematics' long tradition of open preprints.
  • AI labs are entering frontier mathematics research with commercial racing logic, and with no consensus on rules for authorship and credit attribution, disputes may grow as AI involvement increases.

Episode 5 · OpenAI Denies Using Private Codex Chats in Millennium Prize Claim (2026-09-11, 3 posts)

After OpenAI claimed progress on another Millennium Prize problem, critics alleged it stole user chat data as evidence; OpenAI denied accessing private Codex data, and community members called the accusation baseless.

Episode 6 · OpenAI's Millennium Prize Proof Mired in Authorship Dispute (2026-09-11, 2 posts)

After OpenAI claimed its internal model produced a partial Navier-Stokes proof, NYU mathematician Tristan Buckmaster alleged OpenAI pressured him in private calls to remove a collaborator's attribution from related work.

Episode 7 · Math Community Debates AI's Role After OpenAI's Millennium Prize Claim (2026-09-11, 6 posts)

After OpenAI claimed to have (nearly) cracked another Millennium Prize Problem, the math community erupted around September 11 in a large-scale debate over values: Is AI an increasingly necessary tool for modern mathematical research, or a force damaging the mathematical academic ecosystem? The debate remains unresolved, but the arguments clearly expose the community's deep anxiety about AI's involvement in fundamental research—well worth watching.

Confirmed

  • The immediate trigger was OpenAI's claim about a Millennium Prize Problem, with several math-community bloggers posting intensively on the topic.
  • Some argued AI is "an increasingly necessary tool for modern mathematical research"; basedjensen relayed that another user pushed back, asking who has the authority to declare that, and proposed putting it to a vote of the entire mathematical community via a special ICM—only for responders to retort: a tool that can solve Millennium Prize Problems and multiple open problems, and people are seriously debating whether it's necessary.
  • gleech relayed criticism that OpenAI makes math worse by "deliberately scooping the work of peer mathematicians"; but he immediately flipped with self-mockery: the Millennium Prize itself flattened and gamified mathematics, so "wrecking the Millennium Prize" may not be a bad thing.
  • gleech also summarized three positions: one holds that the Millennium Prize's real function is to "corral the hyper-competitive" so dreamers can pursue real scholarship in peace, so OpenAI ruining the prize devalues mathematics; another takes the opposite stance (i.e., the prize itself is the bigger problem).
  • User nileshtrivedi pointed out a double standard: Google DeepMind competed in the CASP14 protein-folding competition with a black-box AI solution and won, with Demis Hassabis and John Jumper even receiving the Nobel Prize in Chemistry for it—yet OpenAI attacking math faces resistance.

Why it matters

  • The core of this debate goes beyond the attribution of a single result; it's a structural conflict between AI institutions and human mathematicians over priority of results and the academic incentive system, and the Millennium Prize–Nobel Prize analogy amplifies the question of "who gets to define the rules."