Mathematicians Are Using AI to Work Faster — Even as They Accuse It of Stealing Their Work

Leading mathematicians are increasingly relying on AI tools from OpenAI and Anthropic even as they accuse the companies of using their research without permission to accelerate major breakthroughs. The tension, sharpened by OpenAI’s disputed claims around solving the Navier-Stokes existence and smoothness problem, shows how deeply AI has already embedded itself in mathematics and how hard it may be to pull back.

What began as a fight over attribution has grown into a broader reckoning over credit, privacy, research ethics, and whether human mathematicians can still meaningfully compete in a field where AI systems can search through proof strategies at industrial scale.

Why this dispute matters beyond one proof

The clash over OpenAI’s math announcements is important because it goes to the heart of how scientific discovery is supposed to work. Mathematics has long been built on transparency: researchers publish ideas, others check the work, and credit is assigned through citations and peer review. AI complicates that model by making it harder to see where an idea originated and who contributed what.

For mathematicians, the issue is not just whether an AI system can solve a famous problem. It is whether the field’s core norms can survive when large labs can deploy thousands of agents, harvest patterns from vast amounts of text, and then present the result as a machine-driven triumph.

Tristan Buckmaster, a New York University mathematician at the center of the latest dispute, argues that even if researchers distrust the companies behind these systems, they are still often stuck using them because the tools are so effective and the market is so concentrated.

That frustration captures a broader dynamic in academia: the same systems people believe may have absorbed their work are now being used to edit papers, generate ideas, and speed up routine tasks.

What happened in the OpenAI Navier-Stokes controversy?

OpenAI said it had used AI agents to help solve the Navier-Stokes existence and smoothness problem, one of the most famous unsolved questions in mathematics. Buckmaster says the company learned the equation was close to being solved only after it had already drawn on his approach, which he says helped it reach the result faster.

After Buckmaster made those accusations public, OpenAI reviewed the matter and updated its announcement. The company said Buckmaster’s Codex prompts in the two months before the September 8, 2026 paper and announcement could not have influenced the system in any way, including through training.

OpenAI pointed to that revised statement in its response to WIRED. Buckmaster, meanwhile, says his own recent work has been consumed by the fallout, leaving him with less time to do mathematics and more time to examine how the company may have built on his ideas.

He has still been using OpenAI’s Codex coding agent to clean up papers and, by his account, to help reconstruct the logical path AI tools may have taken from earlier work to the final proof.

How are mathematicians using AI even while criticizing it?

They are using it because it is fast, accessible, and increasingly hard to avoid. Buckmaster says he has kept using Codex despite his objections because it helps with paper cleanup and other tedious work. Other mathematicians describe a similar pattern: skepticism about the companies, but practical dependence on the tools.

That contradiction is one reason the backlash has not translated into a broad exit from AI. In many labs, the technology is already woven into writing, coding, literature review, and proof exploration. Refusing to use it can feel like opting out of a competitive advantage that peers will not ignore.

• AI tools can draft text and clean up manuscripts.
• They can suggest proof directions and help test logic.
• They can accelerate literature searches and coding.
• They can also create uncertainty about where ideas came from.

That last point is especially troubling in mathematics, where attribution and priority are central. If an idea shows up in a model’s output after being absorbed from countless documents, who deserves the credit: the original researcher, the prompt writer, the lab, or the model itself?

How did OpenAI respond to the criticism?

OpenAI’s response has been to deny that the disputed prompts influenced the result and to revise public statements that overreached. In the Navier-Stokes case, the company amended language in its announcement after Buckmaster pushed back, specifically addressing whether his prompts could have shaped the system’s behavior.

That change did not end the controversy. For some researchers, the central concern is not simply whether a specific prompt affected a specific model run. It is whether their work may have been absorbed into training data or into surrounding workflows in ways they cannot see, control, or verify.

And even where a company insists a particular result was independent, skepticism remains. The broader worry is that the combination of private models, opaque training data, and large-scale agent systems makes it nearly impossible to determine how a result was produced.

What is the deeper mathematical worry?

The deeper worry is that AI is changing the nature of mathematical contribution itself. Andreas Thom, a German mathematician who studies geometric group theory, says he was stunned when OpenAI claimed its Astra model had used techniques from that field to prove a problem he had been working on.

Thom says he contacted OpenAI researchers after noticing that the company’s statement that there had been “no progress” in a decade ignored his 2019 paper and related work by other mathematicians. OpenAI later amended its release. Thom and a colleague had also been using ChatGPT to help with their own work on the problem, but he says he was told that their interactions had not been used in training.

Thom says he has little confidence that he will ever know for certain whether his research influenced the system, and he believes AI is weakening the old chain of attribution that academic science depends on.

That loss of traceability is not a side issue. It strikes at the foundation of academic life, where citations, authorship, and priority help define careers and keep the research ecosystem honest.

What are mathematicians saying about credit and ownership?

Many mathematicians say the problem is not only whether AI can discover something new, but whether companies are respecting the human work that made those discoveries possible. Buckmaster has argued that presenting machine-generated proof breakthroughs without properly acknowledging the underlying human research is irresponsible, especially when the companies involved are preparing for major financial events such as IPOs.

He also warns that even relatively small uses of AI can have unexpected consequences. A researcher may use a model to improve grammar or polish a draft, only to discover later that their work has entered a commercial system or been folded into outputs that benefit someone else.

That concern resonates with a wider coalition in the field. Twenty-five Fields Medalists signed an open letter saying AI companies and mathematicians are badly misaligned. More than 4,000 people have also endorsed the Leiden Declaration, which offers recommendations for how researchers, funders, and governments can reduce the risk that AI swallows mathematical work without fair protections.

What do critics want?

They want stronger rules on data use, clearer disclosure when AI is part of a proof process, better citation practices, and more care around how results are announced. Some also want institutions to slow down events that normalize corporate control over mathematics.

Their message is not necessarily anti-AI. It is more precise than that: if AI is going to become part of mathematics, then the field needs guardrails before the technology makes old norms impossible to restore.

Why are universities and students paying close attention?

Universities are paying attention because AI is already affecting how students and faculty work. Matthias researchers at major schools now ask not just what models can do, but how to get access to stronger versions and how to use them without compromising privacy or authorship.

Cornell mathematician Alex Townsend says many colleagues have started asking what they need to learn about AI and how to subscribe to more powerful systems. He describes the mood as a mix of excitement and anxiety: the tools can make impossible tasks tractable, but they also raise uncomfortable questions about the role of human expertise.

Townsend says the field is hearing a new, more existential question from researchers: if AI can do so much, what exactly is the human mathematician’s purpose?

That question is already reaching students. Townsend says younger mathematicians are asking what their careers will look like if AI systems become capable of tackling some of the field’s most prestigious problems.

How does AI change the culture of mathematical research?

It changes the culture by speeding up work while making authorship less visible. In traditional mathematics, a paper usually reflects a chain of reasoning that can be followed, debated, and checked. AI can compress that process into a sequence of machine-generated outputs that are difficult to interpret after the fact.

For some researchers, that is a productivity boost. For others, it is a warning sign. If AI can help draft a proof, clean up notation, search prior work, and even suggest the next step in an argument, the line between assistance and authorship starts to blur.

That blurring is especially pronounced when a company deploys tens of thousands of agents, as OpenAI reportedly did in its Navier-Stokes effort. A result produced at that scale may be scientifically impressive, but it also raises a practical question: how much of the discovery was actually understood by the humans presenting it?

What is the Navier-Stokes problem, and why do people care?

The Navier-Stokes existence and smoothness problem is one of the Clay Mathematics Institute’s famed Millennium Prize Problems, with a $1 million reward attached. It asks, in simplified terms, whether the equations used to model fluid motion always produce smooth, well-behaved solutions in all cases.

The question matters far beyond pure mathematics. Navier-Stokes equations underpin aerodynamics, weather prediction, ocean modeling, and many engineering systems. A proof or disproof would deepen understanding of how fluids behave and could influence how scientists think about complex physical systems.

That prestige is one reason the OpenAI dispute got so much attention. If AI systems can meaningfully contribute to a Millennium Prize Problem, then they are no longer just productivity tools. They are becoming participants in the most elite tier of mathematical research.

Why are some mathematicians calling for a pause?

Some researchers believe the field needs to slow down before AI becomes the default research infrastructure. That is why more than 2,000 people with ties to Caltech asked organizers to suspend an AI-focused math hackathon sponsored by Anthropic and, at one point, also OpenAI. OpenAI later withdrew from the event.

These critics are not claiming that AI should disappear from mathematics. Their argument is that the current rollout is happening too fast, with too little oversight and too little discussion of what happens to research culture, younger scholars, and intellectual ownership.

• They want clearer standards for disclosure.
• They want stronger privacy protections for research inputs.
• They want credit policies that reflect human contributions.
• They want institutions to think carefully about student dependence on AI.

Still, not everyone thinks a brake can be applied for long. Thom believes the efficiency gains are likely too large to resist, especially for early-career researchers trying to keep up. In his view, refusing to use the tools may leave people isolated in a field that is rapidly reorganizing around them.

What comes next for mathematicians and AI labs?

The next phase may be less about whether mathematicians use AI and more about how the field sets rules for doing so. Buckmaster is calling for a détente: a period of practical restraint in which researchers and AI companies agree on standards before the damage to trust deepens further.

He says he plans to clean up several papers he posted before they were fully finished, including one he now describes as AI slop. He also says he feels a responsibility to explain what his group did and to fix the record for other mathematicians.

That desire to repair the record may be one of the clearest signs of the moment the field is in. Researchers are no longer debating AI in the abstract. They are trying to preserve the basic mechanics of scholarly trust while using the very tools they fear may be changing it.

Could mathematicians and AI companies work together?

Possibly, but only under stricter conditions than exist now. Buckmaster says he is open to speaking with OpenAI, though he does not want the issue to become a permanent fight. Even so, he remains cautious about whether collaboration between mathematicians and AI labs can happen without new safeguards around credit, data use, and disclosure.

For now, the field is left with an uneasy balance: mathematicians are frustrated by AI companies’ power and opacity, yet they keep using the tools because they are too useful to ignore. That contradiction may define the next chapter of mathematical research.

As Thom puts it, the old idea that every contribution can be traced cleanly back to a person may already be over. If that is true, the challenge for mathematics is not just to solve harder problems with AI. It is to decide what counts as a contribution at all.