# A Mathematician’s Warning: How AI Is Consuming the Very Problems That Drive Mathematical Progress
## The Unseen Cost of Solving Problems Too Quickly
Terence Tao, a professor at UCLA and one of the most celebrated pure mathematicians alive, has raised a profound concern about how artificial intelligence is reshaping the landscape of mathematical discovery. Best known for his Fields Medal awarded in 2006, Tao has warned that the rapid deployment of AI systems in mathematics is consuming open problems — the unsolved questions that serve as the lifeblood of the field — at a rate far faster than human researchers can identify new ones worth pursuing.
The warning isn’t about AI replacing mathematicians. It’s about AI eliminating the very challenges that push the discipline forward. In a post on a widely used mathematics-focused social platform, Tao articulated a distinction that many in the field are now grappling with: the difference between solving problems and actually advancing understanding.
“Anyone can invent infinite new math questions,” Tao wrote, noting that even the googol-th digit of pi technically counts as an open problem, yet almost none of such questions carry genuine significance. What matters, he argued, is knowing which questions are worth the sustained effort required to solve them — a judgment that has historically depended on a deep understanding of a problem’s difficulty, its connections to other areas of mathematics, and the potential insights its resolution might unlock.
## How AI Labs Are Rewriting the Rules
The concern is grounded in a series of startling developments. With the emergence of powerful reasoning models and frontier AI systems, major research labs have begun directing enormous computational resources at mathematical and scientific problems that have resisted human solvers for decades. Anthropic, OpenAI, and others have tested their models on challenges spanning quantum physics, applied mathematics, and medicine — with results that have occasionally stunned the broader research community.
The mathematical results have been particularly striking. In a notable case, an OpenAI model disproved a conjecture about unit distances between points on a plane — a question that had stood unresolved for 80 years. Within the same week, an Anthropic researcher independently ran the identical problem through the company’s unreleased top-tier model, working entirely offline to avoid copying the first solution. The result was a shorter, elegantly simple proof. A mathematician reviewing the work described it as “a bit worse” than the first version, but still a valid and notable achievement.
More recently, researchers at Anthropic formalized a proof of Fermat’s Last Theorem — a landmark result first established centuries ago — while OpenAI solved a problem that had remained open for 90 years, publishing a coauthored paper alongside an Anthropic researcher within hours of a human mathematician posting their own proof online.
## The Danger of Flattening the Difficulty Landscape
Tao’s core argument centers on what he calls the “difficulty landscape” of mathematics — the invisible terrain that tells researchers which questions are trivial, which require genuine effort, and which may be hopeless with current methods. Throughout history, new techniques have temporarily flattened parts of this landscape by solving previously intractable problems, but they have also naturally revealed new frontiers just beyond their own reach. This self-correcting dynamic has been one of the engines of mathematical progress for centuries.
AI, Tao argues, threatens to break this pattern entirely. When models can solve problems that once required years or decades of human insight, the difficulty landscape becomes effectively opaque. Nobody can say precisely where an AI system’s abilities begin and end, which means the careful process of choosing which problems to invest in — a process that has always been central to mathematical research — starts to break down.
As Tao put it, the indiscriminate use of powerful solution-extraction tools can achieve immediate results but at the cost of “sustaining the ecosystem for the next wave of progress, or in understanding the progress already obtained.” He also pointed to a troubling feedback loop: even the rumor that a research team is working on a problem now triggers massive AI-powered efforts to solve it, often before the human researchers have had time to explore its full potential.
## A Proposed Solution: Requiring Reasoning, Not Just Answers
Rather than calling for a blanket ban on AI in mathematics — something Tao acknowledges would be “technically infeasible” — he has proposed a more nuanced approach. He wants mathematicians to label certain problems as “analysis-required,” meaning that a correct answer generated by an AI would carry little weight unless it is accompanied by detailed reasoning that reveals something meaningful about related problems or broader mathematical structures.
Tao drew a comparison to food banks that stopped accepting donations merely because the food was edible, insisting instead on items that were actually nutritious. The analogy captures the spirit of his proposal: a solution is only valuable if it teaches us something, not just if it produces a correct result.
As of now, no institution or journal has formally adopted such a labeling system, and it remains unclear whether the research community would embrace it given how fiercely the AI race is already underway.
## Frequently Asked Questions
**What exactly is Terence Tao warning about?**
Tao is warning that AI systems are solving important open mathematical problems so rapidly that the field is losing its supply of meaningful challenges faster than researchers can identify new ones. The concern is not about AI replacing mathematicians, but about AI removing the very problems whose difficulty drives discovery and deepens understanding.
**What are “analysis-required” problems?**
These are mathematical questions that Tao proposes should be explicitly labeled as requiring more than just a final answer. Under this framework, a raw solution produced by an AI would be considered insufficient unless it includes the reasoning and explanatory steps that illuminate why the answer is correct and what it reveals about related problems.
**Has AI actually solved important math problems recently?**
Yes. In a short span of time, AI systems have tackled an 80-year-old conjecture about point distances on a plane, independently rediscovered a proof of the Erdős unit-distance conjecture, formalized a centuries-old proof of Fermat’s Last Theorem, and solved a problem that had been open for roughly 90 years.
**Why is mathematics different from other fields when it comes to AI?**
Unlike many scientific domains where AI can generate an essentially unlimited number of testable hypotheses, genuinely difficult mathematical problems are relatively scarce. Mathematicians have always chosen carefully which questions to pursue, and the scarcity of truly hard, meaningful problems makes the impact of AI-driven solutions disproportionately significant.
**Is anyone proposing a ban on AI in math?**
Some have floated the idea, but Tao describes it as “technically infeasible.” His alternative proposal is a labeling system that requires AI-generated solutions to include explanatory reasoning, ensuring they contribute to genuine understanding rather than merely producing correct outputs.
## Conclusion
Terence Tao’s warning strikes at the heart of how mathematics has progressed for centuries. The field has always relied on a delicate interplay between knowing which problems are worth solving and having the patience and insight to work through them. If AI removes that constraint — allowing solutions to appear faster than the community can assess their significance or extract deeper meaning — the long-term health of mathematical research could be at risk. Whether the proposed labeling system gains traction or the research community finds another way to preserve the pace of discovery, one thing is clear: the relationship between artificial intelligence and pure mathematics is entering uncharted territory, and the choices made now will shape the future of the field for decades to come.
Thank you for reading
