Terence Tao is arguing that artificial intelligence may force mathematics into its biggest self-examination since the early 20th century. According to The Decoder, Tao makes the case in a new essay for the 2026 International Congress of Mathematicians, where he says the field should move beyond asking what AI systems can do and confront a harder question: what mathematical research is for. The analogy Tao draws is not to a routine tooling shift. The Decoder reports that he compares the present moment with the foundational crisis between 1900 and 1930, when Russell’s paradox and Gödel’s incompleteness theorems forced mathematicians to make explicit assumptions that had previously remained implicit. That earlier upheaval centered on the foundations of mathematical truth. Tao’s claim, as summarized by The Decoder, is that AI is testing a different layer: the field’s values and practices. Those values include what counts as a contribution, what gets rewarded, what it means to understand a proof, and whether work performed by a machine can be treated as mathematical work in the same sense as work performed by a person. Tao’s working hypothesis, according to The Decoder, is that AI tools will “reasonably soon” be capable of carrying out a meaningful fraction of research-level mathematical tasks with acceptable success, quality, supervision, and cost. The evidence cited in The Decoder’s account includes the First-Proof Project. In its second round, ten unpublished research problems were tested against four AI systems under controlled conditions. The Decoder reports that seven of the ten problems received at least one passing grade from at least one system, meaning the proposed solution was judged essentially flawless or required only minor revision, with costs running from tens to hundreds of dollars per problem. That is the mechanism behind Tao’s deeper concern: AI could separate goals that have historically reinforced one another. Solving problems, building theories, training young mathematicians, and sustaining a research community have been linked parts of the same system. If machines can generate substantial volumes of plausible or correct research output, those functions may no longer move together. Tao also points to Goodhart’s law, according to The Decoder: when a measure becomes a target, it stops being a good measure. In this framing, AI systems are especially likely to optimize for the appearance of a high-quality mathematical result, while the AI industry’s incentives favor quotable and benchmarkable demonstrations. The concern is not only that models might make mistakes, but that they could produce outputs that look like the things the field has learned to reward. One practical risk is proof abundance. The Decoder says Tao warns that AI-generated proofs could accumulate faster than experts can verify, read, or integrate into the discipline. As an example, the report says the Erdős problem database already contains dozens of AI-generated submissions that no human expert has volunteered to check. The Verge separately reports that the mathematics community is already in a state of debate after OpenAI published solutions to longstanding math problems. In a Decoder podcast summary, The Verge describes leading mathematicians as wrestling with an “existential crisis” over what mathematicians are for if frontier models can solve high-end abstract problems. That Verge item does not corroborate the details of Tao’s essay, but it supports the broader context: AI’s recent progress in mathematics is no longer being treated as a distant possibility by parts of the field. Who benefits: AI labs benefit if mathematics becomes a visible domain for demonstrating frontier-model capability, especially through benchmarkable or publishable successes. Mathematicians who can use AI tools while preserving rigorous human verification may also gain leverage. Who's exposed: Academic mathematics is exposed if its reward systems cannot distinguish between useful AI assistance, superficial polish, and contributions that genuinely advance understanding. Training pipelines for new mathematicians could also face pressure if problem-solving output becomes less scarce.