OpenAI AI Math Proofs: 722 Manuscripts Spark Attribution Fight

What OpenAI released

On October 6, OpenAI published a large batch of mathematics generated by one of its AI systems: 722 manuscripts that the company says resolve long-standing open problems 12. The papers are organized into 372 "result families," which group related manuscripts together, and were produced by a frontier model that has not been released publicly 2.

The drop was not a surprise. In September, OpenAI said its model had "resolved more than 100 long-standing open problems across most areas of mathematics," but it did not say which problems or when the work would appear 2. The new release fills in some of those gaps. Alongside the papers, OpenAI included summaries of the model's reasoning, estimates of the compute involved, and figures on how many problems the system attempted 2. The company says the "average result" took roughly the equivalent of three hours of ChatGPT Pro thinking time 2.

According to AGMAI, the release covers solutions to "hundreds" of open questions 2. AGMAI is a newly formed, independent advisory group of prominent mathematicians assembled to help communicate the results responsibly 2.

Why mathematicians are pushing back

The scale of the release has not settled the debate. Coverage of the October 6 drop describes renewed anger among top mathematicians, including Fields Medalists, over attribution 1. That dispute had already been building after an earlier OpenAI claim related to the Navier-Stokes equations 1. The broader reporting describes a run of AI-driven results that have impressed and unsettled the mathematical community in roughly equal measure, raising questions about research ethics and academic conduct 2.

The two accounts emphasize different things. One leads with the backlash and treats the release as fuel for an ongoing revolt 1. The other spends more time on the release itself, including its structure, the advisory group, and the disclosed compute figures, while still presenting the controversy as central to the story 2. They agree on the basic facts and on the main tension: OpenAI is producing mathematical output at a volume and pace the field has never handled, and the field has not decided how to credit it or check it.

Reading the numbers

Some of OpenAI's disclosures look intended to answer criticism in advance. Publishing reasoning summaries, compute estimates, and attempt statistics gives outsiders a way to judge the work beyond headline claims 2. The attempt numbers matter in particular. A success count means more when readers can see how many problems the model tried and failed.

The framing still deserves caution. "Three hours of ChatGPT Pro thinking" is a description chosen by the company, and it is not a clear measure of cost or difficulty 2. Counting manuscripts or result families also says nothing on its own about how significant any single result is. An open problem can be long-standing because it is extremely hard, or because few researchers ever tried it. The phrase "hundreds of open questions" covers a lot of ground 2.

The attribution problem

The attribution fight may last longer than any individual proof. Academic mathematics depends on a reputation system. Credit determines careers, and correctness is established slowly through review by experts. A sudden release of hundreds of machine-generated papers puts heavy pressure on both. Mathematicians have to decide who deserves credit when a model builds on earlier human work, who carries responsibility if a proof contains an error, and whether a company should announce "breakthroughs" before the community has checked them.

AGMAI's existence suggests OpenAI knows these questions are sensitive. A panel of elite mathematicians helping to shape how results are communicated could support legitimacy 2. However, the panel was put together around this effort, and some critics may see it as giving OpenAI's claims an academic endorsement. That would be especially true for critics whose complaints about attribution began with the Navier-Stokes episode 1.

The takeaway

This release looks like a real escalation in what AI systems are being asked to do in mathematics. It also shows that the hardest problems here are increasingly social. OpenAI has moved from a vague September claim to a large, partly documented body of work, which is a meaningful step toward accountability 2. The volume, however, makes the main criticism harder to answer. Verification and credit assignment move at the pace of human experts, and 722 manuscripts arriving at once will strain that process 12.

The test that matters will not be the size of the release. It will be how many of these results hold up under independent review, and whether OpenAI and the mathematical community can agree on norms for crediting the human work these models rely on. Until that happens, the backlash is likely to grow with each new batch.