AINEWS
When one of the paper's authors is an LLM, the question of who owes the citation gets sharper, and OpenAI has already walked back part of its own pitch.
Steven Miller, a mathematician at Yeshiva University, says OpenAI's new paper reproduces an argument he co-authored in 2016 on packing spheres into higher-dimensional boxes and never cites him. Miller calls the omission "systematic" and "research misconduct" in comments to Scientific American.
The dispute sits inside a release OpenAI built for the math community. The company's late-July write-up lists 10 results it says its next major reasoning model, identified in Scientific American's reporting as "Astra," produced end-to-end, with the proofs and the model commentary checked by a small set of human mathematicians. Total compute, per OpenAI, came to about $2,000 in tokens. The paper runs nearly 250 pages, and a companion volume of reasoning walkthroughs is hosted alongside. The most quoted claim, that the results address problems "open and seen no progress on the main result for at least a decade," was softened inside a week after expert pushback; the live post now reads as a more measured version, and the same ten-proofs paper and the open GitHub repository sit next to it.
Miller's complaint is about a specific argument: a 2016 paper he wrote with a collaborator on how densely unit spheres can be packed in spaces with more than a thousand dimensions, a problem with a known answer in low dimensions and stubborn in the high. He says a key construction in that paper reappears in OpenAI's result without attribution.
"It's a systematic pattern, and it's research misconduct," Miller told Scientific American.
A second of the 10 results, in group theory, has drawn similar criticism from Francesco Fournier-Facio at the University of Cambridge and Andreas Thom at the Dresden University of Technology, both cited in the same piece. The pattern, not the model name, is what makes the complaint cut. Humans remain on the author list, and the humans are the ones who sign the citation page.
This is the part the standard "AI solves hard math" framing skips. Surfacing a new proof at this scale is a real engineering result. Posting the proofs and the walkthroughs openly, the way OpenAI has, is what researchers in the field have asked frontier labs to publish. None of that dissolves the authorship question. A paper that takes an argument from a 2016 article and calls it a new result is a paper with a missing citation, and that is true whether the drafter was a graduate student, a senior author, or an LLM.
OpenAI has already moved on at least one of its own framings. The original "no progress for at least a decade" wording survives in some early coverage and in screenshots, but the live blog post acknowledges the problems had not been "fully solved" and that progress had come in stages. The edit is small. It is also the kind of change a publisher makes only when the original wording has been pushed on, and it is the first time the company has adjusted a claim from this paper in public.
The company has not addressed the citation dispute on the record in the same post. An OpenAI spokesperson, in comments reported by Scientific American, said the company takes concerns about attribution seriously. The statement stops short of conceding the point. The fact that OpenAI has already revised one piece of its own copy makes the silence on Miller's complaint harder to wave off.
Underneath the dispute is a rule that does not change when one author is a model: a published argument that traces to prior work has to say so, and the humans on the author list are the ones who owe the check. Independent analyst Simon Willison, writing about the same release, flagged the press-release walkback as a small but real correction. That is the constructive lens the rest of the field can use: not "AI cannot do math," and not "AI is stealing proofs," but "the byline still has consequences."
The test case is the next paper. OpenAI says a larger run of AI-generated proofs is in the works. If the citation review on that paper is public, the company's first attempt will look like a learning step. If it is not, two named allegations will become a pattern, and the authorship standard the math community has enforced for a century will have to be re-fought for the LLM era.