Two of OpenAI’s 10 Astra math results face prior-work attribution dispute

Two of the most prominent results in OpenAI’s package of 10 AI-generated mathematical advances are now under scrutiny for how they credit prior work. Mathematicians reviewing the roughly 250-page paper produced around the company’s unreleased Astra model found that crucial ideas had appeared in recent literature. According to Scientific American’s August 6 report, OpenAI has already changed promotional language that said the problems had seen no progress on their main results for at least a decade.

The two disputed results

The first concerns sphere packing in dimensions of 1,000 and above. Astra’s proof improves the best estimate for how densely spheres can be packed, but Yeshiva University mathematician Steven Miller said its central argument first appeared in a 2016 paper he wrote with a collaborator. Miller objected that the model’s proof presented the method as its own and characterized the omission as a research-integrity problem.

The second result concerns soficity, the question of whether every mathematical group can be faithfully approximated by simpler groups. Astra constructed at least one group without that property. Francesco Fournier-Facio of the University of Cambridge and colleagues then traced its pivotal step to a combination of ideas in papers from 2016 and 2019. Their criticism does not make the resulting theorem worthless. Andreas Thom, a co-author of the 2019 paper, described the construction as creative and elementary. The narrower issue is that recent human progress was essential to the result, contrary to OpenAI’s initial framing.

A $2,000 computation still needs costly scholarship

OpenAI said the total token cost for all 10 results was only $2,000. That number highlights how inexpensive model inference can be compared with traditional research, but it leaves out the labor required to establish novelty. Literature review, attribution, expert checking and revision remain separate parts of publication. A model can locate real papers and combine methods while still failing to explain which contribution is new and which came from earlier researchers. Responsibility for that distinction ultimately rests with the people and institution publishing the work.

An OpenAI spokesperson told Scientific American that the company takes responsibility for the correctness of the results and is applying standards expected of human mathematicians. OpenAI also said it planned small updates to the paper under normal academic practice. The next test is therefore visible and concrete: whether the revised manuscript properly traces the two arguments, and whether the other eight results survive the same literature review. As AI systems enter high-level mathematics, proof correctness is only one gate; novelty claims and citations have to withstand expert scrutiny too.