Consider the moment when a mathematician, deep in the throes of wrestling with one of the seven Millennium Problems — a puzzle so fiendish that its solver is promised $1 million and immortality — decides to use ChatGPT as a thinking partner. He feeds it fragments of his unpublished work. He asks it to check his logic. He treats it like a tireless, if slightly dim, research assistant. And then, weeks later, he watches in horror as an AI system from the same company announces it has made 'substantial progress' on the very problem he was solving — progress that looks suspiciously like his own.
This is not a hypothetical. This is the story of Kevin Buckmaster, an independent researcher who, alongside his longtime collaborator Levent Alpöge, has spent years attacking the Navier-Stokes existence and smoothness problem. In September 2025, Buckmaster went public with a startling accusation: that OpenAI's Codex, the company's coding assistant, had ingested his private research prompts without his knowledge or consent. His evidence was circumstantial but troubling. He claimed that before he revealed his identity, OpenAI's agents showed little progress on the problem. Within a day of him disclosing his name — and, by extension, his affiliation and research trail — he says those same agents reached 'exactly where we were.'
OpenAI denied it. Vehemently. And narrowly. The company stated that Buckmaster's Codex prompts from 'the past two months' could not have influenced the system 'in any way, including for training.' Note the precision of that denial. It covers a specific time window and a specific mechanism. It does not address earlier data pipelines. It does not address session memory, retrieval-augmented generation, or the possibility that a human on the inside connected dots that a model could not.
The technical reality is messier than the headline suggests. Buckmaster's causal claim — that revealing his name triggered a sudden leap in AI capability — does not survive scrutiny. If his data had already entered the training corpus, his name would be irrelevant. The data is the data. What could plausibly change with identity disclosure is human behavior: a researcher inside OpenAI recognizing the name, pulling up associated drafts, or manually guiding an agentic workflow. Alternatively, the entire episode could be explained by convergence — the natural tendency of a well-trained model on public mathematical literature to independently arrive at the same partial result as a human expert. Navier-Stokes is a Millennium Prize problem. Its frontier is public. The literature is vast. A model trained on that literature producing a step that matches a human's unpublished step is unlikely, but not impossible. And crucially, Buckmaster has not demonstrated that his step was genuinely absent from the public record.
Here is where my own experience auditing early-stage crypto and AI projects becomes relevant. In 2017, I reviewed over fifty whitepapers during the ICO boom, and I learned to distinguish between a founder who genuinely had proprietary technology and one who had repackaged public research with a fresh coat of paint. The difference was rarely in the code. It was in the provenance of the idea. Buckmaster's case suffers from the same ambiguity: we cannot see the provenance of the AI's output. We see only the output itself and the researcher's assertion that it mirrors his own. Trust is the only currency that matters, and right now, OpenAI is asking us to extend it without collateral.
The competitive context sharpens the stakes. Alpöge works at Anthropic. This is not a footnote; it is the subtext. OpenAI and Anthropic are locked in a fierce race to claim AI-for-mathematics breakthroughs, and priority claims carry enormous reputational weight. Buckmaster alleges that OpenAI offered him sole authorship if he dropped Alpöge — a claim OpenAI's Sébastien Bubeck characterizes as an attempt to 'help.' We cannot verify either version. What we can observe is a pattern: OpenAI simultaneously denies the allegation and announces progress on another Millennium Prize problem. Crisis management and capability branding, executed in the same breath.
The deeper issue is not whether OpenAI stole Buckmaster's ideas. It is whether any researcher can safely use AI tools for sensitive work. Buckmaster noted that one of his accounts still had model training enabled by default. This is the industry standard — opt-out, not opt-in. For enterprise customers, data sovereignty is a dealbreaker. For academics, it is often invisible. And for the broader public, it is a trust deficit that compounds with every ambiguous denial.
Code binds, but people break or build. The architecture of modern AI platforms places the tool provider and the data recipient in the same corporate body. There is no independent auditor. There is no transparent ledger of how prompts become parameters. When a researcher pastes a grant proposal into ChatGPT to polish its abstract — a practice Buckmaster says is rampant among funding reviewers — they are donating their most valuable intellectual property to a private training pipeline. This is not malice. It is design.
Here is the contrarian angle that most coverage misses: the real victim of this episode is not Buckmaster. It is the early-career mathematician. OpenAI reportedly released 722 AI-generated mathematics papers addressing 4,000 open problems. Even if a fraction are correct, the sheer volume overwhelms the human peer-review system. A doctoral student who once hoped to stake a career on solving a narrow but meaningful sub-problem now faces an AI competitor that can generate a plausible attack in minutes. The selection space for human research is being colonized, not by superior intelligence, but by superior throughput.
Culture eats blockchain for breakfast, and it eats AI research for lunch. The academic community's norms — attribution, peer review, embargo periods — were built for a world where ideas traveled at the speed of human conversation. They are not built for models that ingest millions of tokens per hour. Until the incentives align with transparency, every researcher who types a secret into a prompt box is gambling with their life's work.
What would a better system look like? It would start with opt-in data policies, not opt-out. It would require disclosure of training data provenance for models used in research contexts. It would create independent audit trails for high-stakes platforms. And it would recognize that the boundary between 'tool' and 'collaborator' is porous — and ethically fraught.
We are building the future, together, whether we intend to or not. The question is not whether AI will reshape mathematical discovery. It already has. The question is whether the humans who do the discovering will still own the rights to their own questions. Buckmaster's case may never be proven. But the discomfort it exposes is real, and it will not disappear with a carefully worded press release. The next researcher who opens a chat window and hesitates before typing deserves a better answer than 'we deny everything, including for training.'
