What happens to mathematics when finding a proof is no longer the hardest part? For much of its modern history, research mathematics has been built around the difficulty of solving open problems. A significant theorem could occupy a researcher for years, or a community for generations. Finding a proof was rare enough that the discovery itself often justified the attention and prestige it received.
Artificial intelligence is beginning to challenge that assumption. In an October 2026 presentation titled Math 2.0, mathematician Terence Tao argues that mathematics is entering a new era in which powerful AI systems can generate proofs and solutions at a pace that the traditional research ecosystem was never designed to handle. His argument is more interesting than the familiar prediction that machines will replace mathematicians. It raises a deeper question: if proofs become abundant, what should mathematicians actually be trying to achieve?
The answer, in Tao’s account, has less to do with collecting theorems and more to do with developing understanding, improving explanations, connecting results, and using mathematics responsibly in the wider world.
From Math 1.0 to Math 2.0: the end of proof scarcity?
Tao describes traditional mathematical research as Math 1.0, a system organized around what he calls proof scarcity. To solve a difficult open problem, a mathematician typically needed substantial expertise, years of accumulated knowledge, and considerable time. It made sense for research institutions, journals, prizes, hiring decisions, and professional reputations to place great weight on the production of new results.
But mathematics is also unusual among intellectual fields. A mathematical proof, once precisely stated and formalized, can in principle be checked without conducting an experiment in the physical world. Proofs are highly digital: mathematical definitions, research literature, and proof-assistant libraries are increasingly accessible in machine-readable form. And a correct proof has an objective logical standard, even when mathematicians disagree about whether a result is interesting.
Tao identifies three characteristics that make mathematics especially hospitable to contemporary machine learning: objective verifiability, digitizability, and access to extensive high-quality data. These properties have encouraged frontier AI laboratories to invest heavily in mathematical problem-solving. The result, he argues, is a transition toward proof abundance.
This does not mean that every difficult theorem is now easy, that mathematical expertise has become irrelevant, or that the frontier of research is about to disappear. Tao is explicit that most problems at the current frontier remain unsolved even with large amounts of computation. Mathematics contains infinitely many questions, with no upper bound on their difficulty. AI may move the frontier forward, but it will not eliminate the frontier.
Why AI is exceptionally powerful in some mathematical tasks and surprisingly weak in others
A striking feature of current AI systems is how unevenly their abilities develop. Given access to powerful frontier models and enough computation, users can sometimes obtain solutions to problems previously thought difficult, and may occasionally do so without the specialized knowledge traditionally required to solve them. At the same time, these systems may remain unreliable on other tasks, including problems involving uncertain real-world conditions, limited data, or judgments that cannot be reduced to a mechanically checked answer.
Tao calls this profile jagged capability. It is important because impressive performance on a particular class of theorems does not demonstrate equal competence across mathematics, let alone science in general. A system that discovers an elegant combinatorial argument may still mishandle an apparently simpler task that depends on a poorly represented assumption.
There is another reason to be skeptical of sweeping narratives about AI solving mathematics: reporting bias. We regularly hear about the successful attempts because they make headlines. Failed experiments and expensive searches that lead nowhere are much less visible. In his presentation, Tao uses a familiar image of a bullet-marked airplane to illustrate why a selective collection of successes can create a misleading impression of what AI can reliably do.
A meaningful assessment would report not only the successful proof, but also how many attempts failed, what prompts and tools were used, how much computation was consumed, and whether the result was checked independently. Tao points to First Proof as one effort toward more transparent evaluation of AI’s mathematical research capabilities.
This distinction is especially relevant after major AI laboratories began announcing large collections of mathematical results. InsightArea recently examined OpenAI’s release of 372 families of AI-generated mathematical results, including the difficulty of evaluating them. Tao’s presentation provides a useful conceptual framework for why that evaluation matters.
Solving a problem is not the same as understanding mathematics
One of Tao’s most illuminating analogies concerns the role of open problems in pure mathematics. He describes them as lighthouses: they guide researchers through a landscape, but the journey can be more valuable than arriving at the destination. While trying to prove something, mathematicians develop techniques, notice unexpected relationships, refine conjectures, and learn which assumptions actually matter.
A proof that resolves a problem is a valuable outcome. Yet the paths explored while searching for it can generate far more mathematics than the final statement alone. A failed approach may become useful elsewhere. A partial result may introduce a new method. A clearer formulation may reveal that two subjects previously treated as separate are closely connected.
Now imagine an AI system that supplies the final proof immediately, but offers little insight into the route. We have obtained an answer, perhaps even a formally correct one, while missing much of the intellectual exploration that would normally accompany it. Tao warns that indiscriminate automation risks impoverishing the surrounding field by bypassing the development of human understanding.
This is not a claim that mathematicians should reject correct AI-generated proofs or artificially prolong research. Rather, it challenges the idea that the number of solved problems is an adequate measure of mathematical progress. When the cost of producing a theorem changes dramatically, the value of other activities—explaining, organizing, comparing, generalizing and connecting results—becomes easier to appreciate.
Why a formally verified proof may still not be enough
The distinction becomes even more consequential when mathematics is applied outside mathematics. In pure mathematics, a precisely defined statement can be checked for logical correctness. In applied mathematics, the situation is more complicated because the mathematical model is itself a representation of the world.
Even a flawless deduction from a model’s assumptions cannot guarantee that the assumptions describe reality accurately, that relevant variables were included, or that the resulting recommendation serves the goals we intended. In Tao’s diagrams, human understanding connects pure problems, applied problems, proofs and judgments about what counts as a useful application. Removing that central connection can make the system efficient at producing answers while increasing the chance that those answers are badly aligned with real needs.
Tao illustrates the issue with a deliberately hypothetical medical scenario. An advanced AI is asked to find a cancer treatment that passes a phase-three clinical trial. It proposes an unfamiliar chemical mixture, produces some form of formal certificate for its prediction, and the treatment appears to succeed under the stipulated test. Would we feel comfortable using it if no specialist understood how it worked or whether it might be exploiting weaknesses in the trial or the mathematical model?
The scenario is a thought experiment, not a report of an actual AI-discovered treatment. Its purpose is to expose the gap between optimizing a measurable target and understanding whether the outcome is genuinely safe or desirable. Formal verification establishes what follows from formal assumptions; it cannot, by itself, certify that those assumptions capture the full complexity of the world.
This concern extends beyond medicine. In engineering, finance, public policy and AI safety, the ability to inspect how a mathematical result interacts with the real system is often as important as the formal correctness of the calculation.
What would a more productive Math 2.0 look like?
Tao’s preferred response is not simply to place AI at the end of a pipeline that begins with a question and ends with a proof. He proposes using AI throughout the larger research process, especially where it improves the interplay between machine capability and human mathematical understanding.
1. Make difficult mathematics easier to understand
A theorem may be established yet remain difficult to explain. Tao suggests treating better explanations as a legitimate research objective: an open exposition problem, rather than an open problem about whether a theorem is true. AI could help find shorter proofs, intuitive reformulations, explanatory diagrams or alternative approaches that reveal why the conclusion follows.
His slides mention a revised Caltech Mathathon, designed around using open-source AI tools to improve proofs of existing results whose current demonstrations are unsatisfactory. This is a different ambition from racing to solve an unsolved theorem. It asks whether a known result can become more intelligible and useful.
2. Investigate whole landscapes instead of isolated questions
Traditional research often focuses on an individual conjecture: can we prove it, disprove it or establish a partial result? With automated assistance, it becomes increasingly realistic to investigate very large families of related statements and compare their structure systematically.
One example in Tao’s presentation is the Equational Theories Project, a collaboration that brought human and automated reasoning together to settle more than 22 million true-or-false statements in universal algebra. The significance is not just the remarkable number. It is the opportunity to map a large conceptual space and see how individual claims depend on, or differ from, one another.
Another example is the Inverse Galois Problem challenge. Tao describes a first stage in which participants located degree-24 polynomials associated with roughly 25,000 possible Galois groups at that degree, followed by a plan to extend the work to higher degrees. Such efforts can turn one-off discoveries into structured mathematical data that researchers can search, compare and build upon.
3. Build living networks of mathematical knowledge
Mathematical papers contain chains of ideas, definitions, lemmas and dependencies. They are usually published and read as separate documents, even though their intellectual content is highly interconnected. Tao points to the Integrated Explicit Analytic Number Theory Network as an example of a living structure that combines many related results and can be updated as the state of knowledge changes.
With capable AI tools, researchers might maintain richer networks of mathematical knowledge instead of relying entirely on static literature searches. For a reader coming from computer science, the analogy is useful: isolated papers resemble separate code files, while an explicit network of dependencies makes it easier to understand how changing one assumption affects the rest of the system.
Large formal mathematical libraries such as Mathlib are another part of this infrastructure. They make mathematical definitions and proofs more reusable and machine-checkable. As these libraries improve, AI systems can potentially contribute to more ambitious projects with better verification and clearer connections to existing work.
4. Explore what happens when assumptions change
Tao also suggests mathematical versions of ablation studies. Suppose a theorem’s existing proof relies on a sophisticated piece of theory. Can we still prove it without that machinery? Is there a more elementary argument? Can we identify which assumptions are truly indispensable?
Related work involves exploring the neighborhood of a theorem: weakening a hypothesis, changing a parameter or removing a condition, then investigating how much of the conclusion survives. This can reveal something a stand-alone proof does not: why a theorem has the shape it does.
These questions are natural candidates for human–AI collaboration. Machines may search large families of variations, while mathematicians develop interpretations and decide which patterns reveal a meaningful principle.
AI mathematics needs transparent evidence, not just spectacular headlines
There is a productive tension in Tao’s position. He takes AI’s mathematical abilities seriously enough to describe a genuine disruption of research, but he also resists equating selected successful demonstrations with comprehensive mathematical competence. The field needs to understand both what the systems can do and where they fail.
That means collecting negative results, measuring computational resources, improving independent checks and asking whether newly generated proofs introduce transferable mathematical ideas. It also means avoiding the assumption that a model can safely solve any applied problem merely because it performs well on precisely specified formal tasks.
In other words, mathematical research is developing a second evaluation problem alongside the first. We still need to know whether a proposed theorem is true. Increasingly, we also need to know what the proof teaches us, how it was produced, what it cost, what remains unexplored, and how it relates to other knowledge.
The implications extend to software engineering and scientific research
Although Tao’s slides concern mathematics, the argument resonates with other technical fields. AI coding tools can produce implementations more rapidly than developers can read or reason about them. Yet a working program is not sufficient evidence that its design is maintainable, secure or appropriate to a user’s needs. Code generation becomes more valuable when it also helps engineers understand architectural trade-offs, discover invariants and analyze failure modes.
The analogy is not exact: unlike a mathematical proof, the correctness of a real software system often depends on changing environments, incomplete requirements and complex human expectations. That difference reinforces Tao’s broader point. The usefulness of automated output depends on how well the surrounding process preserves understanding and judgment.
In science more generally, faster hypothesis generation will matter only if researchers can identify the hypotheses worth testing, design experiments that genuinely discriminate between alternatives, and understand what the observations imply. Producing answers is one activity within a larger process of knowledge creation.
Will AI replace mathematicians?
Tao’s presentation does not establish that human mathematicians are becoming obsolete. If anything, it proposes that the most useful human contribution may change. As machines become increasingly capable of producing formal arguments, people may devote more effort to posing fruitful questions, choosing meaningful definitions, designing research programs, explaining results, and testing mathematical ideas against the real world.
There will still be difficult unsolved problems. There will also be new kinds of work: organizing huge bodies of results, building reliable formal infrastructures and extracting conceptual insights from a volume of proofs no individual could comfortably read. AI can assist in each of these activities, but the presentation argues against reducing mathematics to a scoreboard of completed conjectures.
Tao closes with a striking suggestion: the main bottleneck to Math 2.0 may be imagination, rather than computing technology or even institutional culture. We have spent decades optimizing research around the scarcity of proofs. Now we have to imagine what mathematical research should look like when producing a proof is, in some circumstances, substantially easier.
The future of mathematics will not be measured only by how many problems AI solves. It will also be measured by whether those solutions help humans understand more.
Frequently asked questions
What is Math 2.0 according to Terence Tao?
Math 2.0 is Tao’s proposed vision of mathematical research in an era of increasingly abundant AI-generated proofs. It emphasizes human understanding, new explanations, organized mathematical knowledge and meaningful human–AI collaboration, rather than optimizing solely for the number of solved problems.
Does Tao believe AI can solve every mathematical problem?
No. His presentation stresses that AI capabilities are uneven, most frontier mathematical problems remain unsolved, and the range and difficulty of mathematical questions are effectively unlimited.
Why isn’t a computer-verified proof sufficient for every application?
A formal proof checks what follows from specified premises. Real-world applications additionally depend on whether the premises and models accurately describe the situation, and whether the resulting solution meets the goals that matter.
What can AI do for mathematics besides proving new theorems?
It can help explain existing results, investigate variations of theorems, organize dependencies across papers, expand formal proof libraries, and discover reusable patterns and methods.
Sources and further reading
Primary source: Terence Tao, Math 2.0, October 2026, a 29-slide presentation analyzed for this InsightArea article. The examples, terminology and structure of the argument above are drawn from that presentation; comparisons with software engineering are editorial interpretation.
- First Proof — a project cited by Tao in his discussion of transparent AI mathematics evaluations.
- Mathlib documentation — a formal mathematics library mentioned among the tools that could support Math 2.0.
- OpenAI’s 372 AI Math Results: What Has Actually Been Proved? — a related InsightArea analysis of AI-generated mathematical results and verification.
Written by Costin Liculescu for InsightArea, exploring the connections between mathematics, computer science, artificial intelligence and scientific reasoning.
Comments are closed.