Terence Tao thinks mathematics is about to go through the kind of disruption that other professions have been bracing for. In a slide deck titled “Math 2.0,” dated October 2026, the Fields Medalist argued that AI has ended the era in which proofs were scarce, and that the field now has to decide what it is actually for. Tao has shared the deck after sharing a statement from the Association of Human Mathematicians that said that mathematicians hadn’t asked for the 300+ proofs that OpenAI dropped earlier this week, and that it would harm the field.
Tao begins with how mathematics has worked until now. Traditional “Math 1.0” assumed that solving an open problem and proving the solution correct was hard, rare and dependent on deep human expertise. Institutions, incentives and even the stated goals of the field grew up around that scarcity.
But mathematics has three properties that make it unusually easy for machine learning to attack. Proofs are objectively verifiable, and can even be checked mechanically in proof assistants like Lean. They are digital, requiring no physical experiment or real-world data. And much of the relevant literature is already online. Frontier AI labs have therefore prioritised math ahead of domains that lack one or more of those traits, and Tao says the result is an era of “proof abundance.”
He describes AI’s abilities as jagged. With enough compute, frontier models can now solve many open problems once thought difficult, increasingly without special expertise from whoever is prompting them. Yet they remain weak on tasks that are subjective, depend on real-world interaction, or lack data. Tao also cautions against trusting the headlines. Using the famous image of a bullet-riddled bomber from the survivorship bias story, he notes that failed AI attempts mostly go unreported, so social media and press releases show a skewed sample. Most problems at the real frontier remain unsolved, and since the space of problems is infinite, the frontier keeps expanding. He wants more transparent assessments that report resource use and negative results, and points to the First Proof project as one effort in that direction. Its second round, published in June, found that AI systems got roughly six or seven of ten research-level problems essentially right.
Tao has long been wary of how AI’s math results get framed. He has said that OpenAI used only upbeat snippets from a much longer conversation in one of its ads, omitting the balanced view of risk and opportunity he says he tried to convey. The Math 2.0 deck reads like that fuller picture. It also arrives amid fresh disputes over who deserves credit when AI is involved, such as the row over the Navier-Stokes problem, in which OpenAI said neither its researchers nor its agents saw a rival team’s work before it was released. Episodes like that show why careful reporting of what was done, by whom and at what cost matters.
The heart of the talk is Tao’s claim that problem-solving was only ever a proxy. In pure mathematics, he says, open problems act as lighthouses. They are less destinations than guides for exploring the surrounding landscape, and what mathematicians learn while attempting them, whether they succeed, fail or make partial progress, is often worth more than the answer. Reaching those lighthouses prematurely with automated tools can disrupt the exploration of paths not taken and sterilise the surrounding field.
He backs this with a model of the Math 1.0 ecosystem. Pure problems are solved in a renewable cycle that builds human understanding, and that understanding can then be safely transferred to applied problems, where judging whether a solution is truly aligned takes subjective judgement. Indiscriminate AI problem-solving breaks the cycle, leaving the field with answers but less understanding and making it harder to apply insights safely. Further blind optimisation of problem-solving alone, Tao writes, is now actively harmful to the long-term health of mathematics.
He illustrates the stakes with a thought experiment. An advanced AI is told to find a cancer cure that passes a stage 3 trial. It produces a cocktail whose effectiveness is confirmed in Lean and in the trial, though no one knows how it was found. Could it be exploiting a weakness in the trial process? Would you want at least one human expert to understand the mechanism before it went into your bloodstream? Verification, he suggests, is necessary but not sufficient, because someone still has to understand what has been verified. That echoes an earlier concern of his that AI-generated mathematics can look flawless while hiding subtle mistakes a human wouldn’t make.
Tao’s prescription is to expand the research frontier while decentering problem-solving. His examples include large-scale experimental mathematics, such as the Equational Theories Project, which settled over 22 million statements in universal algebra, and the 2026 Inverse Galois Problem challenge. He also proposes “open exposition problems” that seek motivated explanations of opaque proofs, with the Caltech Mathathon as an early experiment, along with living networks of results, large formal libraries like Mathlib, ablation studies that test whether theorems survive without key inputs, and open, interpretable math models. He says some of those models will be released soon.