Technology media outlet scientificamerican published a blog post on August 6, reporting that several mathematicians accused OpenAI of academic misconduct, claiming that its AI model Astra failed to properly credit existing research for its mathematical results.
On July 31, OpenAI published a 249-page PDF detailing major advances achieved by its next-generation flagship model Astra (internal version) in several fields, including high-dimensional sphere packing, quantum game theory, group theory, and extremal combinatorics.
At Sol API rates, the total cost of the tokens required to find solutions to these problems would be approximately $2,000 (note: approximately 13,519 yuan at the current exchange rate).
However, after the paper was published, several scholars accused OpenAI of academic misconduct following their review.
Two core results failed to properly cite existing ideas from recent mathematical literature. Yeshiva University mathematician Steven Miller pointed out that, for the sphere-packing problem in spaces of 1,000 dimensions or higher, Astra presented a core mathematical argument. The argument first appeared in a paper published by Miller and his collaborators in 2016, but Astra presented it as its own work.
Another result addressed a longstanding problem concerning Soficity in group theory, proving that at least one group lacking this property exists. Cambridge mathematician Francesco Fournier-Facio and Andreas Thom, a mathematician at Dresden University of Technology, found that Astra’s key mathematical steps combined ideas from two papers published in 2016 and 2019.
In response to the criticism, an OpenAI spokesperson said in a statement that the company stood behind the correctness of the results and that they met generally accepted standards among human mathematicians, adding that it planned to make minor updates to the paper this week. Fournier-Facio emphasized that OpenAI is now fully involved in high-level research and must be subject to the same academic standards as humans.