OpenAI's 722 AI-Generated Math Papers Split Mathematicians Between Awe and Alarm

What OpenAI published, and when

On October 6, 2026, OpenAI posted 722 manuscripts on GitHub, grouped into 372 families of related results, all produced by an internal model the company has not released publicly, according to xenospectrum.com and the AI-focused newsletter Latent.space. The company said the model had been set loose on roughly 4,000 mathematical problems, with the average accepted result costing the computational equivalent of about three hours of ChatGPT Pro reasoning time, Latent.space reported.

One manuscript addresses what researchers are calling a "Quasi-Riemann Hypothesis": a claim that no zeros of certain Dirichlet L-functions exist where the real part exceeds 7/8. Xenospectrum.com noted this does not solve the actual Riemann hypothesis, which concerns zeros at real part exactly 1/2 and remains an unsolved Clay Mathematics Institute Millennium Prize problem.

Praise and alarm among mathematicians

Reactions split sharply. Mathematician Levent Alpöge, described by Latent.space as affiliated with Anthropic, called the release "obviously the most significant moment in mathematical history." The Conversation reported that other mathematicians were far less impressed: one reviewer dismissed part of the batch as a "giant pile of turd dumped at our doorstep," and a senior mathematician said one paper was so unintelligible that it would have gone straight in the bin had it been submitted to a journal in the ordinary way.

The Conversation also reported that ChatGPT itself flagged at least one of the results in the batch as a "serious hallucination," and that several papers were later retracted or amended after errors were found.

The verification gap

Accounts of how much of the release has actually been machine-checked differ. Xenospectrum.com, tracking OpenAI's own formalization catalog as of October 7, counted 162 of the 722 manuscripts — about 22.4 percent — with results formalized in the proof-verification language Lean. The Conversation, by contrast, reported that OpenAI had formalized 300 main results using Lean-type software. Xenospectrum.com added that even a completed Lean formalization only confirms the logical steps of a proof, not that every explanatory claim in a paper has been independently verified.

Tech-insider.org reported that an advisory group involving Fields Medalist Terence Tao was set up to help manage how OpenAI's results reach the public, but that the group's role was explicitly to advise on release, not to independently verify that the results are correct. The same outlet cited a September claim that OpenAI's model, using roughly 10,000 AI agents working for about 88 hours, had produced a proof bearing on the Navier-Stokes existence and smoothness problem — another Millennium Prize problem — which the Clay Mathematics Institute still lists as active rather than solved, meaning no independent verification has been completed.

What mathematicians still dispute

Beyond correctness, tech-insider.org reported concern that OpenAI's publish-first, verify-later approach inverts how mathematics has traditionally worked, shifting the burden of slow, adversarial peer review onto a mathematical community that is already stretched thin. The Conversation reported additional unease over alleged data misuse tied to the earlier Navier-Stokes claim, raised by mathematicians Tristan Buckmaster and Levent Alpöge — allegations OpenAI has denied — as well as broader anxiety in the field about research being "scooped" by AI systems and about what the growth of such tools means for mathematics as a career.