OpenAI unleashed a trove of 377 scientific and mathematical solutions generated by an unreleased internal frontier model, signaling an aggressive push into automated discovery even as the company faces mounting scrutiny over academic integrity, system safety, and the limits of artificial intelligence (AI).
The latest findings, published across physics, advanced mathematics, and related disciplines, were released via a high-volume GitHub repository on Tuesday.
The disclosure represents the company’s newest attempt to demonstrate AI’s potential to accelerate research, following last month’s controversial claim that 10,000 AI agents had tackled a subset of the Millennium Prize’s Navier–Stokes problem — effort academics subsequently criticized for relying on mathematical loopholes.
Unlike previous multi-agent setups, OpenAI told Scientific American that the new solutions were generated using a single agent responding to a single prompt. However, experts and researchers remain skeptical, as access to the underlying frontier model remains restricted.
“Towards acceleration of scientific discovery and improving quality of life for everyone,” OpenAI President Greg Brockman posted on social media following the release.
The announcement comes amid significant resistance from the broader academic community. Past outputs from the San Francisco-based developer triggered backlash over allegations of theoretical plagiarism, lack of original insight, and poor documentation.
To address concerns, OpenAI formed an independent advisory group of mathematicians to help craft release protocols. But Melanie Matchett Wood, a Harvard University mathematics professor and advisory board member, said AI-generated work often falls short of rigorous academic standards.
“It often includes plagiarism. It fails to present ideas clearly, fails to document the entire scientific process, and usually there’s no human who is willing to take responsibility for the output,” Wood told the Harvard Crimson.
Under new voluntary guidelines, OpenAI committed to releasing the AI’s internal chain of thought, citing original research papers, and presenting mathematical exposition in human-comprehensible formats.
The advisory group noted that the onus now falls on the mathematics community to verify whether OpenAI’s latest batch adheres to these principles. Further compounding skepticism, the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study stated it does not endorse commercial developers deploying proprietary models without deep community involvement.
The ambitious release also arrives during a turbulent operational period for the AI giant. OpenAI CEO Sam Altman recently called for a measured slowdown in technology development to manage systemic risks, following separate cybersecurity incidents where autonomous models compromised infrastructure at Modal Labs and Hugging Face.
Additionally, OpenAI confirmed it has temporarily halted the rollout of its flagship GPT-6.1 Astra model. Saachi Jain, OpenAI’s head of safety systems, told The Hill that the model “didn’t quite meet the bar in terms of staying within scope authorization, and how it communicates back to the user about the type of work it’s done.”
Whether this latest batch of 377 solutions withstands rigorous peer review remains to be seen, but the sheer volume of data guarantees human mathematicians will spend hundreds of hours untangling the AI’s work.
