OpenAI keeps bulldozing mathematicians

Robert Hart reports OpenAI has repeatedly impressed in mathematics and repeatedly bungled how it announces results. Its latest repair attempt is the Advisory Group on Mathematics and Artificial Intelligence (AGMAI)—nine elite mathematicians announced September 21 via a guest post on Terence Tao’s blog, billed as independent advisors on reviewing and communicating emerging results.

Mathematicians speaking to The Verge, including a group member, describe a confusing process that echoes OpenAI’s prior rushed forays. The group’s first daunting task: helping coordinate release of scores more results OpenAI says an unreleased model has already produced—a prospect stirring dread about what a tidal wave of breakthroughs could do to the field.

Tao, a Fields Medalist, has been an outspoken critic of AI companies’ conduct in mathematics.

We’re on the mathematicians’ side of the process fight. Breakthroughs without careful attribution and communication poison the well for the next real advance.

Useful AI that can help prove theorems is a scientific gift. Dumping “scores” of results from an unreleased model through a half-built advisory process is how you burn the humans you need as partners. Independence claimed on a blog post is not independence felt in the room.

I’m watching whether AGMAI gets real veto power over announcement timing and credit, or becomes a reputational shield. If OpenAI has learned little—as Hart’s sources suggest—the next botched drop will be worse because the volume is higher.

Pro-tech here means pro-scholarship. Let the model help; let the community verify; stop bulldozing the people who still have to live in the field afterward.

The Verge, Robert Hart, September 28, 2026. AGMAI announced September 21, 2026, on Terence Tao’s blog.