News

Your AI Research Assistant Works for Someone Else

• By John K. Waters
• 09/10/2026

Scientists are being encouraged to trust artificial intelligence with some of their hardest problems. They give AI systems unpublished ideas, experimental approaches, half-finished proofs, code, and questions they may have spent years learning how to ask.

An awkward fact is buried in that relationship. The AI belongs to somebody else.

That problem came into unusually sharp focus this week after OpenAI announced that an internal AI system had produced what the company says is a solution to the Navier-Stokes existence and smoothness problem, one of mathematics' seven Millennium Prize Problems.

As we reported separately, the result is potentially significant evidence that frontier AI systems are beginning to move beyond accelerating research and toward participating in discovery itself.

But the circumstances surrounding the result have exposed another question that may prove nearly as consequential: What happens when the company providing your AI research assistant can also become your research competitor?

A Rumor and 10,000 AI Agents
The controversy centers on Tristan Buckmaster, a mathematics professor at New York University, and Levent Alpöge, a mathematician at Anthropic.

The pair had been working on closely related problems involving singularities in fluid dynamics. They used several AI systems during that research, including Anthropic's Claude and OpenAI's Codex, according to reporting by WIRED.

Their results were not the full Navier-Stokes solution OpenAI would later claim. They concerned related equations, including the Euler equations, which describe fluid motion without viscosity.

While Buckmaster and Alpöge were finishing their work, however, word of their progress began circulating. OpenAI acknowledges that this sparked its own effort.

“On Tuesday, September 1, we heard rumors that two Millennium Prize problems had been resolved,” the company said in its post about the project. OpenAI later determined that the rumor was connected to Buckmaster and Alpöge.

OpenAI had something the two mathematicians did not: extraordinary computational resources and a new internal AI model the company describes as significantly more capable than GPT-6 Astra. OpenAI turned those resources loose on the problem.

The company initially assigned agents to several major mathematical problems. After its system produced a result involving the Euler equations, OpenAI decided Navier-Stokes looked especially promising and shifted additional resources toward it.

Eventually, roughly 10,000 agents were involved in the group that produced the Navier-Stokes result. OpenAI says the agents reached their proposed solution after about 88 hours, followed by another 17 hours of formalization and verification using Astra. The Navier-Stokes effort generated approximately 130 billion output tokens and 2.7 million messages. OpenAI executives put the computational cost in the millions of dollars.

The ability to marshal resources on that scale creates a new kind of competitive asymmetry in scientific research. A mathematician may spend months pursuing an unusual approach to a difficult problem. A frontier AI company that learns the approach is promising can potentially direct thousands of artificial researchers toward it almost immediately.

None of this suggests that OpenAI did anything improper, of course. But it does suggest that the playing field has changed.

The Codex Question
The more uncomfortable issue is that Buckmaster and Alpöge had been using OpenAI's technology themselves. Buckmaster questioned whether information from their Codex use could have played some role in OpenAI's work. WIRED reported that he specifically asked OpenAI whether the company had accessed their Codex logs. OpenAI denies doing so.

“We (the researchers and the agents) did not see any of their work through any means until they released it publicly,” the company said.

OpenAI says it did not access any specific user data to solve the problem and that its proof differs significantly from the mathematicians' work. It has also recognized Buckmaster and Alpöge's priority on their related Euler result. But in its statement, OpenAI included an important qualification.

“While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models,” the company said.

No evidence shows that Buckmaster and Alpöge's private research produced OpenAI's Navier-Stokes result. OpenAI explicitly denies that its researchers or agents saw the work. But that distinction doesn't make the broader problem disappear. It may actually define it.

Researchers increasingly use AI systems not merely to polish finished work, but during the messy intellectual process that precedes it. They test hypotheses. They expose failed approaches. They ask questions whose significance may not be apparent to anyone outside a highly specialized field.

In other words, the conversation with the AI can itself reveal where the researcher believes something valuable might be hiding.

Then Came the Fight Over Credit
The dispute became more personal after OpenAI and the mathematicians began talking about how their respective results might be announced.

OpenAI says it contacted Buckmaster and Alpöge after completing its project and Lean verification on September 6. (Lean is a software tool that translates mathematical reasoning into formal logic and checks, step by step, whether a proof actually holds.) Believing the pair had also solved Navier-Stokes, OpenAI says it offered a concurrent release that would recognize their priority. Only then, the company says, did it learn that their result concerned forced Euler rather than Navier-Stokes. Buckmaster describes the subsequent conversations differently.

According to WIRED and Axios, Buckmaster alleges that OpenAI researcher Sébastien Bubeck discussed a proposal under which Buckmaster could help announce OpenAI's Navier-Stokes result without Alpöge's name attached. Buckmaster interpreted the proposed exclusion as related to Alpöge's employment by Anthropic, one of OpenAI's principal competitors.

Bubeck has disputed Buckmaster's characterization that OpenAI sought to remove Alpöge from credit. OpenAI says it recognizes the priority of both researchers' work on forced Euler.

Buckmaster also alleges that the discussions became heated. He says that after indicating he might make the disagreement public, Bubeck asked him, “Why would you ruin your career?” Axios reported the allegation, while other accounts make clear that OpenAI disputes Buckmaster's characterization of the exchange.

Those allegations concern private conversations and have not been independently established.

But the disagreement over what happened in those conversations points toward a much larger problem than a fight among mathematicians.

Who Is the Research Assistant Working For?
Generative AI has created an unusual relationship between researchers and technology companies. The researcher may regard the model as a tool. The company operating the model may regard interactions with that tool, subject to its policies and user settings, as data that can potentially contribute to improving future systems. And the resulting system may eventually become capable of doing the same research.

That possibility extends well beyond mathematics.

A biologist could use an AI system to explore an unusual experimental result. A software engineer could discuss a new algorithm with a coding agent. A startup founder could use a model to work through an invention before filing a patent. The value isn't necessarily contained in the final document. Sometimes the valuable thing is knowing which problem is worth attacking.

The OpenAI episode shows how powerful that signal can become. OpenAI says it began its Millennium Prize effort after hearing a rumor that major problems had been solved. After its agents made progress on Euler, it redirected enormous resources toward Navier-Stokes. Within days, the company says, its system had produced a solution to a problem mathematicians had been unable to resolve for decades.

Again, none of that demonstrates that OpenAI appropriated Buckmaster and Alpöge's work. It demonstrates why researchers may increasingly care about the difference.

The New Research Confidentiality
Scientists have always had reasons to protect unfinished work. Researchers decide when to circulate drafts, whom to tell about promising results, and when to publish. Academic priority can determine reputations and careers. AI inserts another participant into that system. And unlike a traditional research assistant, the AI can potentially become vastly more capable after the researcher has used it.

And that creates a problem even if every company involved behaves exactly according to its published policies. Today's model may be incapable of finishing a problem. Tomorrow's model may not be. The company operating it may have vastly more compute than the researcher. It may employ scientists working in the same field. And, as OpenAI demonstrated last week, it may be able to mobilize thousands of AI agents against a promising problem once it decides that problem is worth pursuing.

Business Insider noted that the controversy raises precisely this broader issue: whether researchers can safely entrust frontier AI systems with valuable ideas when the companies operating those systems may eventually be capable of pursuing the same ideas themselves.

That is not simply a privacy question. It is a conflict-of-interest question.

Discovery Changes the Rules
For the first generation of generative AI, the stakes were relatively familiar. Don't paste your company's confidential financial data into a chatbot. Don't expose customer records. Understand what happens to your prompts. Those remain sensible precautions. But research-grade AI introduces something different. The sensitive information may not be a document or a dataset.

It may be an idea.

And as AI systems become capable of generating genuinely novel scientific results, determining who discovered something, who deserves credit, and what role the company behind the model played will become considerably harder.

The OpenAI-Buckmaster dispute may ultimately have an uncomplicated explanation. OpenAI says its agents independently produced a different solution after hearing only a rumor. Buckmaster believes the circumstances warrant much more scrutiny. Both positions are now part of the public record. The larger question doesn't depend on resolving their disagreement.

Scientists are rapidly acquiring extraordinarily powerful research assistants. Those assistants can read papers, write code, test ideas, and increasingly participate in the intellectual work of discovery. But they aren't graduate students sitting down the hall. They are services operated by some of the most powerful technology companies in the world.

Before telling one your best idea, it may be worth remembering who owns the lab.