OpenAI Publishes 700+ AI-Written Manuscripts on Open Problems, and Mathematicians Are Stunned

OpenAI has published more than 700 AI-written manuscripts that claim solutions to open mathematical problems. One math blog collected over 100 responses from researchers. They include admiration for new ideas, existential fear, and grief over a lost way of working. Fields Medal winner Hugo Duminil-Copin replied tersely: "I am paralyzed."

What exactly happened

The scale matters more here than any single result. Until now, AI breakthroughs in mathematics came one at a time: one problem, one proof, one caveat about how humans helped. Now hundreds of texts have landed at once. Each claims to close a question the community had worked on without success.

According to the source material, the manuscripts themselves claim solutions. That is not the same as confirmed results. The report does not say how many papers independent experts have already checked, or how many held up. It is not yet clear what share of the 700 texts is correct, what share contains gaps, and what share will turn out to be a restatement of known results. That determines how this plays out: as a scientific shift or as an avalanche of texts that nobody has the capacity to read.

Why the reaction was so sharp

The responses on the blog span the whole spectrum. Some researchers write about fascination: the AI proposes ideas they would not have come up with themselves. Others voice fear for their profession. Still others describe grief, and the word is telling. For a mathematician, a proof is valuable not only as a result but as a path to understanding. The title of the piece, "How Much Beauty Have We Lost?", hits exactly that point. If a theorem can come off an assembly line, what is in question is not its truth but the meaning of the manual work that used to give the profession its identity.

The paralysis Duminil-Copin describes looks less like weakness than a sober assessment. Work that took years can be done in hours, and it is still unclear what contribution will remain for humans.

Key figures and facts:

• more than 700 AI-generated manuscripts published by OpenAI;
• the manuscripts claim solutions to open mathematical problems;
• more than 100 responses collected by the math blog;
• reactions range from interest in the ideas to fear and grief;
• among those who responded is Fields Medal winner Hugo Duminil-Copin;
• the source report does not mention independent verification of all the manuscripts.

What to do in practice

The bottleneck is shifting from generating ideas to checking them. This is my analysis, not a statement from the authors, but the logic is straightforward. Mathematicians and departments should decide in advance who will read such texts and by what criteria. They need an expert verification queue, or the flood of manuscripts will drown in its own volume.

Journal editors will have to set policy: whether to accept AI-written papers, how to disclose the model's contribution, and how to tell a correct proof from one that merely looks convincing. Formal verification in systems like Lean looks like a logical tool for some of the results, although it will not cover all 700 texts quickly.

For those watching from the industry side, it helps to keep the difference between "published" and "confirmed" in mind. Until verification is done, the headline about 700 solved problems remains a claim. How much of it survives will show whether this is a new tool for science or just another round of loud promises.