Terence Tao, the UCLA professor widely regarded as one of the world’s top pure mathematicians, has warned that AI is accelerating through mathematics in a way that could drain the field of its most valuable open problems.
In a post published yesterday on Mathstodon, the Mastodon instance used by the math community, Tao said the issue is not simply that models can produce proofs or papers faster. His concern is that AI may be consuming the supply of good open problems—the unsolved questions that actually push mathematics forward.
He argued that anyone can generate endlessly many mathematical questions. The googol-th digit of pi, for example, is technically an open problem if nobody has calculated it. But almost none of those questions matter, because most do not teach researchers much about the broader structure of the field. The real skill, Tao wrote, is knowing what is worth the effort.
“In short, the indiscriminate use of powerful solution-extraction tools can achieve the immediate short-term goal of solving problems at hand, but at the cost of sustaining the ecosystem for the next wave of progress, or in understanding the progress already obtained,” Tao wrote.
Reasoning models are turning math into a contested arena
With the release of reasoning models and the latest generation of frontier AI systems, labs have started directing enormous compute resources toward mathematical and scientific problems. Anthropic and OpenAI have both tested their models on questions that resisted human mathematicians for years, and in some cases for decades.
The reported results span quantum physics, applied mathematics, and medicine. Tao’s argument is that mathematics sits in a special position. Truly difficult problems are relatively scarce, and researchers often spend months or years deciding which ones deserve sustained attention.
That process used to depend on what Tao described as a field’s “difficulty landscape”: a rough sense of which problems are easy, which demand major effort, and which remain out of reach with current tools. New methods have always flattened part of that terrain, but they also tended to expose fresh frontiers beyond their own limits. Tao said AI breaks that pattern because nobody can say with precision where a model’s capabilities stop.
From rumor to race
Tao’s warning was tied to recent events, not a distant hypothetical.
In May, an OpenAI model disproved the Erdős unit-distance conjecture, an 80-year-old problem about how many pairs of points in a plane can lie exactly one unit apart. The result was checked by outside mathematicians, including Fields Medal winner Tim Gowers.
During the same week, Anthropic researcher Levent Alpöge ran the same problem through Claude Mythos, the company’s unreleased top-tier model, in an offline setting so it could not copy OpenAI’s published solution. Anthropic engineer Sholto Douglas described the output as a “cute, simple proof,” saying it was shorter than OpenAI’s version. Mathematician Daniel Litt said Mythos produced something “a bit worse” than OpenAI’s proof, though the model also found OpenAI’s own solution.
Douglas wrote at the time: “Huge credit to the OAI team for solving the unit distance problem with 5.5 - it is now my go to example that models can in fact pull together disparate ideas into new discoveries.” He added: “As with all 4 minute miles, we had to try and cross it too! Turns out mythos solves it with a cute,…”
This week, Anthropic also formalized a centuries-old proof of Fermat’s last theorem. A few days later, OpenAI solved a 90-year-old problem just hours after a researcher published his own proof, then coauthored a paper with an Anthropic researcher.
That dynamic is what Tao said worries him most. “We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential,” he wrote.
Tao said this could reverse patterns that have defined open science for centuries.
Why Tao wants the process judged, not only the answer
His proposed response is to classify some problems as “analysis-required.” Under that approach, a raw correct answer would not carry much weight unless it came with reasoning that illuminates nearby problems as well. Tao compared the idea to food banks that stopped accepting any donation that was merely edible.
He said the alternative would be to ban AI from mathematics, but called that option “technically infeasible.”
The proposal has not become policy anywhere so far. According to the report, the current behavior of major AI labs suggests that even this middle-ground approach may be hard to implement on technical grounds right now.