Terence Tao, the UCLA mathematician who won the 2006 Fields Medal, has published a new essay arguing that artificial intelligence is on track to force mathematics through its most serious identity test in a century. The Decoder’s Matthias Bastian reported the essay’s central arguments on August 20, 2026. The threat Tao describes does not concern mathematical truth. It concerns the field’s implicit rules about who gets credit, what real understanding looks like, and whether a machine can claim authorship of a result.
Tao frames this against the crisis of 1900 to 1930. Russell’s paradox surfaced first; Gödel’s incompleteness theorems followed, exposing assumptions mathematicians had never bothered to state outright. That era hardened into a framework so rigorous it has lasted a hundred years. Tao’s working hypothesis is that AI is about to run a similar stress test, this time on the field’s values rather than its logic: “AI tools will, reasonably soon, become capable of performing a reasonable fraction of research-level mathematical tasks, with reasonable levels of success, quality, supervision, and cost.”
Tao cites the First-Proof Project as evidence. Its second round set four AI systems loose on ten previously unpublished research problems in a controlled test. Seven of the ten cleared the bar when at least one system produced a passing result, which graders defined as output nearly flawless or fixable with light edits, and each attempt cost somewhere between roughly ten and a few hundred dollars.
Tao’s deeper worry is what happens once that capability scales. He invokes Goodhart’s law: “When a measure becomes a target, it ceases to be a good measure.” Generative AI is unusually exposed to that trap, he argues, because it optimizes for the appearance of a good proof rather than the substance of one. Mathematics has long treated quotable, benchmarkable results as a substitute for deeper goals such as theory-building and mentorship, and Tao argues AI’s financial incentives make that substitution worse.
The practical risk, in Tao’s account, is that proofs stop being scarce and start accumulating faster than any mathematician can review, read, or absorb them. Dozens of machine-generated submissions already sit in the Erdős problem database, with not one human expert stepping up to verify them. A separate issue applies even to proofs that do get checked. Tao observes that human-written proofs retain a kind of productive friction: an awkward lemma, a shift in notation, a paragraph visibly rewritten more than once. A proof polished too smooth by AI erases that texture, leaving behind writing he calls “easy to read and hard to learn from.”
Tao also points to a policy anchor: the Leiden Declaration on Artificial Intelligence and Mathematics, a statement the International Mathematical Union backed when it came out in June 2026. His own test is stricter: “If the authors cannot convincingly demonstrate that they are able to give a clear, expert-level talk on their results, one that is correct and properly attributed, then the result should not be published.” A formally verified proof that nobody can explain in a seminar, in his view, is still incomplete. Training young mathematicians needs particular protection for what he calls the field’s “irreducibly human aspect,” since producing correct homework was never the actual goal of that training. He discloses his own AI use too: pulling literature searches, drafting diagrams, finishing text passages, and turning slide decks into paper drafts.
AI Insiders covered a related but narrower argument on August 14, when Fields Medallist Timothy Gowers mapped where large language models succeed and fail at research-level mathematics, a caution echoed by Peter Sarnak. That debate was about capability: how much of mathematics AI can currently do well. Tao’s essay assumes the capability question keeps resolving in AI’s favor and asks a harder one: whether the profession’s mechanisms for assigning credit and verifying work can survive contact with a machine that generates results faster than anyone can check them.
That question reaches past mathematics. Any field that leans on formal verification, code review, financial audit, scientific replication, hits the same bottleneck once checking, not generating, becomes the binding constraint on trust. For a working mathematician, Tao’s essay reads as an instruction to build refereeing capacity ahead of the submission volume, not after it arrives.
The Decoder’s Matthias Bastian reported on Terence Tao’s essay on August 20, 2026.
