OpenAI’s 372 Mathematical Proofs Raise Concerns Among Experts
TEMPO.CO, Jakarta - OpenAI’s latest publication of mathematical results has raised concerns among experts, who argue the artificial intelligence lab is failing to properly verify its work while keeping its AI models out of reach for most mathematicians.
The scrutiny follows the company’s release of more than 370 mathematical proofs on GitHub at 6 p.m. EDT on Tuesday, October 6, covering topics such as algebra, theoretical computer science, and mathematical logic.
The announcement comes on the heels of OpenAI’s reported solution to the Navier-Stokes problem last month, which was a $1 million Millennium Prize challenge in fluid dynamics that is one of the most stubborn open questions in modern mathematics.
Institutions Urge Caution Over Private Models
The Institute for Advanced Study in Princeton, New Jersey, cautioned that it does not endorse leading AI laboratories tackling the hardest math problems with private models that most mathematicians cannot use.
The institute warned that AI systems can now generate mathematical arguments even when the researcher prompting the system cannot fully follow, verify, or stand behind the underlying logic. “Human understanding of mathematics remains of paramount importance,” the IAS asserted.
According to The Guardian, OpenAI said it would work with the institute to give mathematicians a voice in how the work moves forward, though it did not say it would stop testing its models on such problems.
Vetting remains a particular concern. As reported by The Guardian, New York University mathematician Tristan Buckmaster, who has worked on the Navier-Stokes problem, noted that mathematicians prompting the models may feed the system hints or key theoretical structures. In such cases, the AI may simply be executing standard algebraic expansions based on human guidance rather than discovering genuine mathematical breakthroughs on its own.
Questions over Vetting and Transparency
Scientific American reported that mathematicians will need months to sort through OpenAI’s 370-plus findings to determine whether the proofs bring new insights or mostly remix existing techniques. While many results have been validated using Lean, a programming language that validates a proof’s logic, experts emphasize that Lean verifies only logical consistency, not mathematical significance or originality; a proof can be 100% logically sound in Lean while remaining mathematically trivial.
The way OpenAI produced the results has also drawn attention. According to Scientific American, OpenAI said that almost every result came from a single prompt to a single AI agent, though some may have taken several attempts. OpenAI’s Navier-Stokes solution, by contrast, used a swarm of 10,000 AI agents that cost millions of dollars in compute power.
Questions about transparency have also emerged. An advisory group of mathematicians convened by OpenAI last month recommended that AI labs publicly disclose the model, complete prompt texts, and compute time behind each result. However, OpenAI is sharing only aggregate statistics and average compute times while withholding the specific prompt histories. The company told Scientific American that it is not bound by the group’s recommendations.
Concerns over Institutional Access and Equity
The advisory group further asked AI labs to grant mathematicians worldwide fair access to their models, warning that private internal models risk splitting the field between compute-rich corporate laboratories and compute-poor academic institutions.
OpenAI maintained that it is working to release its model as quickly and responsibly as possible. However, the company emphasized that it has no plans to slow down, arguing that math problems are a key test of whether its AI is getting smarter. OpenAI also added that many of the new results are not yet understood by OpenAI’s own mathematicians.
Daniel Litt of the University of Toronto offered a more positive view, saying that he sees no reason for OpenAI to keep the answers secret and expects the results to benefit the broader mathematical community.
Pandu Aryo (Intern)
Read: What We Know About Andrew Garfield’s OpenAI Drama ‘Artificial’
Click here to get the latest news updates from Tempo on Google News