OpenAI removed three mathematical manuscripts generated by an internal artificial intelligence model just one day after presenting them as a historic advance upon detecting a calculation error that invalidated their conclusions. The company also reviewed fourteen other jobs from the same batch.
On October 6, 2026, OpenAI published 722 manuscripts organized into 372 families of related results, in a public repository on GitHub, the platform where programming code is usually hosted and which now served to disseminate mathematical research. The company said it sought to “push the frontier of human knowledge.”
A day later, OpenAI announced the withdrawal of three of those manuscripts due to a sign error that invalidated a central argument and the construction used in two derived works. Each withdrawn document now includes a note explaining “the gap” detected.
The company published 722 manuscripts organized into 372 result families on the GitHub platform on October 6, 2026. (REUTERS/Dado Ruvic/Illustration)
The company also revised fourteen other manuscripts with “evidence fixes, corrected assertions, clearer hypotheses and dependencies, and a correction to an outdated citation,” according to Retraction Watch.
In its own statement, OpenAI acknowledged that the batch includes “results in different stages of verification” and that not all of them have formalization in Lean, so it warned that some of the unverified results could have errors. The company said it will fund workshops and conferences to help the community interpret the findings, without specifying dates.
Dan Roberts, head of research at OpenAI, confirmed in a message on the social network OpenAI did not give a date for publicly releasing the model that produced the results.
OpenAI reviewed fourteen other papers to correct outdated citations, repair evidence, and refine hypotheses. (REUTERS/Dado Ruvic/Illustration/File Photo)
Alex Townsend, a mathematics professor at Cornell University, told the portal that collects retractions in the research community that he was not surprised that an error spread among several manuscripts.
“Given the skepticism towards AI in the mathematics community, I think OpenAI should have announced verified manuscripts first” with Lean, a programming language that allows mathematical proofs to be verified using a computer, he explained.
Kevin Buzzard, a mathematician at Imperial College London, reviewed thirty of the manuscripts linked to his specialty in number theory.
Dan Roberts confirmed the removal of the publications and stated that the company will keep the repository updated with new formalizations. (Illustrative Image Infobae)
As he told New Scientist, only seven seemed relevant and only one was formally Lean verified. “The acceptance of these results by the community is going to take time,” he warned.
Álvaro Lozano-Robledo, a professor at the University of Connecticut, told The Verge that even scrolling through the repository’s index took considerable effort.
“Just going through the entire list of summaries is overwhelming,” he said, referring to a nearly 40-page document that OpenAI itself had to publish to guide reading the material.
An independent panel of mathematicians issued recommendations to increase transparency in the dissemination of findings made with artificial intelligence. (Illustrative Image Infobae)
On September 8, 2026, OpenAI had announced that one of its models solved a problem linked to the Navier Stokes equationsone of the seven Millennium Problems that award a million-dollar prize for its resolution.
That announcement generated criticism from a group of mathematicians who considered that the company rushed to disseminate it “without time for adequate writing or to cite the previous work of other researchers,” according to a joint statement that gathered more than 8,000 signatures from mathematicians around the world.
That tension led to the formation of an independent panel of mathematicians, the Mathematics and Artificial Intelligence Advisory Group, which on September 29 of the same year published recommendations for AI companies to disseminate results with more transparency about the model used, the problems attempted and the cost of computing.
