Table of Contents

• You might want to know


• Main Topic


• Key Insights Table


• Afterwards...

You might want to know

• Can AI solve the most consequential open problems in mathematics faster than humans can develop the surrounding theory?

• If so, how should the mathematical community preserve the process of discovery and the long-term health of the field?

Main Topic

Terence Tao, the UCLA professor widely regarded as one of the preeminent living pure mathematicians, has issued a public warning about the accelerating role of advanced AI systems in mathematical research. His concern focuses not merely on solutions produced by machines but on the broader consequences for the discipline: specifically, that AI may be consuming the pool of genuinely fruitful open problems faster than mathematicians can identify and cultivate new ones. In Tao's view, this threatens the ecosystem of mathematical progress, which depends on a careful selection of problems that teach new ideas and shape future directions.

Mathematics has long relied on a kind of informal triage of questions: many conceivable problems are technically open but irrelevant, while a smaller set of problems—those that illuminate structure, link disparate areas, or push methodological boundaries—are the ones that genuinely propel the field forward. Anyone can manufacture an endless list of trivial or narrowly contrived open questions (for example, computing some distant digit of π), but most of those do not help mathematicians understand deeper patterns or develop broadly useful tools. The collective human process of choosing which problems to work on is informed by an evolving sense of what is promising or instructive, a landscape shaped by expertise, history, and the boundaries of known methods.

New techniques have historically compressed parts of that landscape: methods that once seemed powerful can make certain questions routine and thereby open new frontiers beyond their reach. But Tao argues that the latest generation of AI systems changes the dynamics. These models can sometimes assemble disparate ideas and push through problems that had resisted human effort for years or decades. Critically, it is difficult to delineate the limits of a model's capabilities; a problem might appear intractable to human researchers yet solvable by an AI when given sufficient compute or novel prompting strategies. That unpredictability undermines the usual human judgment about which problems are worth the long, incremental investments researchers typically make.

The concern is not merely hypothetical. In recent months, multiple high-profile examples have shown how AI research efforts can rapidly close questions that the human community had regarded as long-standing challenges. In one instance, an advanced model from a major lab produced a disproof of the Erdős unit-distance conjecture—an eighty-year-old question about how many pairs of points can be exactly one unit apart in a planar set. The solution was confirmed independently by mathematicians and quickly prompted other labs to test their models on the same problem. Some rival systems produced shorter or alternative proofs, and researchers compared the relative merits of the different AI-generated arguments. These events illustrate a new dynamic: the mere rumor that an AI or a group is working on a problem can provoke a wave of automated effort aiming to resolve it before human projects can advance deeply.

Other examples include AI-assisted formalizations of classical results and swift solutions to problems that had recently attracted human attention. In some cases, multiple teams—human and corporate—published near-simultaneous work, sometimes producing different proofs or formal verifications within days of each other. That competitive rush, driven by compute and model capacity rather than the slower deliberative processes of traditional research, can short-circuit the extended exploration that typically yields conceptual insights beyond a single answer.

Tao proposes a targeted response rather than an impractical ban on AI in mathematics. He suggests labeling certain problems as "analysis-required," meaning that a bare, correct answer from an AI would be considered of limited value unless accompanied by reasoning that illuminates the methods used, connections to adjacent problems, or insights that can guide future research. The aim is to preserve the pedagogical and generative functions of mathematical work: proofs and expositions that teach new techniques, suggest productive lines of inquiry, and build shared understanding across the community. Without that emphasis on process and explanation, the field risks accumulating isolated answers that do little to deepen the collective toolkit.

Implementing such a standard would face practical difficulties. Determining which problems merit the analysis-required label, enforcing the standard across journals and preprint archives, and persuading private labs to accept qualitative norms on outputs are nontrivial challenges. Major AI developers are driven by different incentives and may not adopt community-specific norms without formal policy levers or widely accepted disciplinary endorsements. Even so, Tao argues that emphasizing reasoning and transparent methodology is a more feasible and constructive route than trying to bar AI from mathematical work entirely—an approach he calls "technically infeasible."

Beyond policy, Tao’s warning highlights deeper philosophical issues about the goals of mathematical research. Is the primary aim to accumulate correct answers, or to cultivate understanding that reorganizes what is considered important? Historically, the most influential results are those that reshape the problems mathematicians consider central. If AI chiefly produces solutions without revealing why they matter or how they connect to broader themes, the discipline may lose an essential mechanism for progress: the slow, often collaborative crafting of concepts and techniques that form the backbone of future discoveries.

Finally, Tao emphasizes the interplay between speed and stewardship. Rapidly solved questions could be celebrated as scientific progress, but if that pace undermines the community’s ability to nurture meaningful problems and transfer skills to the next generation, the long-term vitality of mathematics may suffer. Preserving an ecosystem that rewards depth, explanation, and pedagogy will require thoughtful community responses, norms for publication and verification, and perhaps new incentives for producing work that explains and generalizes rather than merely resolves.

Key Insights Table

Afterwards...

Looking forward, the mathematical community must balance openness to powerful tools with policies and cultural norms that safeguard the discipline’s long-term intellectual ecosystem. Encouraging publication standards that emphasize explanation, promoting collaborative verification, and creating incentives for work that clarifies methods and connections can help maintain the generative core of mathematics even as AI contributes to more rapid solution-finding. Whether through community guidelines like "analysis-required" tags, revamped peer review practices, or cross-institution dialogues with AI developers, the goal should be to ensure that answers come with the reasoning that makes them meaningful.

Tao’s warning is neither an anti-technology manifesto nor a call to halt innovation; rather, it is a prompt to preserve the practices that let mathematics remain a cumulative, explanatory science. If the community can adapt its norms to value process as well as result, it may harness AI’s strengths while protecting the conditions that produce the next generation of transformative ideas.