A top mathematician on OpenAI's 700 papers: a watershed, atrociously written
The math world's reaction to AI has moved from "is it real" to "what do we do now". Frank Calegari, a number theorist at the University of Chicago, published a guest post on Terence Tao's blog on November 1 titled "Mathematics has changed. Let's act like it."
The trigger is OpenAI's October 6 release of more than 700 AI-generated math papers. Calegari doesn't dismiss them, calling the release a watershed and says AI is undeniably capable of groundbreaking contributions. Then comes the other half: the papers are "atrociously written", they violate the community's expectations about attribution, and some of them are wrong.
The essay turns that into six proposals: evaluate all work as if it is AI-assisted; hire through in-person interaction; demand more from journals and abandon those that don't adapt; emphasize breadth in graduate training; give better seminars; and take more ownership of AI results. The common thread: when proofs get cheap, the scarce skills are choosing problems, building understanding and explaining results.
This is the clearest statement yet of a pattern that shows up well beyond math. Agents can produce output faster than any field can verify, credit or digest it. The bottleneck moves to review, attribution and taste, and institutions like journals and hiring committees are where that bottleneck sits.
Read it: https://terrytao.wordpress.com/2026/11/01/mathematics-has-changed-lets-act-like-it/
← Back to all articles
The trigger is OpenAI's October 6 release of more than 700 AI-generated math papers. Calegari doesn't dismiss them, calling the release a watershed and says AI is undeniably capable of groundbreaking contributions. Then comes the other half: the papers are "atrociously written", they violate the community's expectations about attribution, and some of them are wrong.
The essay turns that into six proposals: evaluate all work as if it is AI-assisted; hire through in-person interaction; demand more from journals and abandon those that don't adapt; emphasize breadth in graduate training; give better seminars; and take more ownership of AI results. The common thread: when proofs get cheap, the scarce skills are choosing problems, building understanding and explaining results.
This is the clearest statement yet of a pattern that shows up well beyond math. Agents can produce output faster than any field can verify, credit or digest it. The bottleneck moves to review, attribution and taste, and institutions like journals and hiring committees are where that bottleneck sits.
Read it: https://terrytao.wordpress.com/2026/11/01/mathematics-has-changed-lets-act-like-it/
Comments