OpenAI AI Math Breakthrough’s Biggest Problem Is Verifying If The Proofs Are Right 

On the last morning of a weeklong national holiday in China, a mathematician checked his phone to find a stream of emails carrying the same PDF. OpenAI had published hundreds of papers claiming solutions to longstanding mathematical problems, leaving researchers to assess whether artificial intelligence had achieved a breakthrough or moved faster than mathematical understanding could follow.

“It looked unreal,” wrote Xiaolei Wu, a group theory researcher at Fudan University, in a post on the Proofs and Prompts blog that drew more than 100 reactions from mathematicians worldwide. “These were almost all the major open problems in my field!”

OpenAI published 722 papers on GitHub on the night of October 6, all generated by an unreleased artificial intelligence model tasked with tackling roughly 4,000 open problems. The release triggered a debate over the implications for mathematical research, with some researchers welcoming potentially career-defining advances while others questioned the speed, transparency and reliability of the results.

The dispute highlights a broader challenge facing the technology industry as AI systems move beyond generating text and software code to attempting complex scientific research: how to distinguish computational breakthroughs from verified discoveries, and how to ensure that researchers can independently reproduce and understand the results.

A flood of proofs, a few retractions

The release followed an earlier announcement by OpenAI on September 8 that an internal model had solved the Navier-Stokes problem, a longstanding mathematical question concerning the movement of fluids such as water and air.

The problem is one of seven Millennium Prize Problems identified by the Clay Mathematics Institute, each carrying a $1 million prize for a solution meeting its requirements. One of the seven, the Poincaré conjecture, was resolved by mathematician Grigori Perelman in the early 2000s.

OpenAI said as many as 10,000 AI agents had worked on the Navier-Stokes problem, at a computing cost running into millions of dollars.

The announcement prompted a dispute over credit after New York University’s Tristan Buckmaster and Anthropic researcher Levent Alpöge published related work a day earlier. Buckmaster accused OpenAI of rushing ahead after learning of their research. OpenAI said it had not seen their work, while Chief Executive Sam Altman defended the company’s team.

The October release was broader. Although none of the five remaining Millennium Prize Problems was conclusively solved outright, OpenAI’s papers claimed advances on three, including the Riemann hypothesis, a famous conjecture concerning the distribution of prime numbers.

Researchers who examined the papers said some results were substantial enough to underpin an academic career. A researcher at rival AI company Anthropic described the release as the most significant moment in mathematical history.

OpenAI said an average result required computing resources equivalent to about three hours of ChatGPT Pro thinking time, substantially less than the resources devoted to its Navier-Stokes effort.

But the company withdrew three papers within days after identifying errors, raising questions about the reliability of a release whose full contents had not been independently verified.

OpenAI said the results were “at different stages of verification” and that “many, but not all, of the manuscripts have been formalized.”

The distinction matters in a discipline where a plausible argument is not necessarily a correct proof. Formalization involves translating mathematical reasoning into a precise logical framework that can be checked by software, helping identify gaps that conventional review may miss.

The scale of the release reflects a wider shift in AI research. Stanford University’s 2026 AI Index Report said industry produced more than 90% of notable frontier AI models in 2025, while some systems reached or exceeded human baselines on PhD-level science questions and mathematics competitions. The trend is raising expectations that AI could accelerate discoveries while intensifying scrutiny of how its outputs are validated.

Proofs without people

For some researchers, the release threatened to upend years of work, whether a problem was already being investigated or had become the foundation of a dissertation, grant proposal or academic career. Questions that had sustained entire research programmes suddenly appeared to have answers.

But a proof is not the same as an accepted mathematical discovery. The results still need to be checked, explained and connected to existing knowledge, a process that could take years. Some claims may not withstand scrutiny.

The immediate concern was that researchers could be tempted to skip the slow work that produces mathematical understanding. If AI delivers answers before people understand how they were reached, the discipline could gain results while losing opportunities to develop new ideas and train future mathematicians.

Terence Tao, a Fields Medal-winning mathematician at the University of California, Los Angeles, was among 25 Fields medallists who had warned weeks earlier that rushing to publish AI-generated solutions could harm mathematics.

After the release, a group letter criticised the approach, describing it as “not a demonstration of scholarship, but a demonstration of power.”

Tao argued that the absence of human authors able to explain the new proofs, give talks or train students posed a particular challenge. He compared encountering the results to learning a film’s ending before watching it or receiving a crossword clue before attempting the puzzle.

“One cannot explore a problem as profitably and richly once one is aware of an existing solution,” he wrote on the Proofs and Prompts blog.

The debate also exposed concerns about transparency. OpenAI did not disclose the prompts used or identify the model behind the papers, despite advice from a group of mathematicians it had helped establish. The company said it intended to continue testing its models on open mathematical problems.

The Clay Mathematics Institute, which established the seven Millennium Prize Problems in 2000, has emphasised the importance of rigorous proof and independent assessment. Each problem carries a $1 million prize, and the institute’s recognition process is distinct from a researcher’s or company’s claim that a solution has been found.

Mathematics is not alone in confronting the tension between AI-assisted productivity and human understanding. Classrooms, laboratories and workplaces are grappling with how to use increasingly capable systems without allowing them to replace the reasoning and expertise people need to evaluate their output.

The challenge is not just whether AI can produce answers, but whether people can verify them, learn from them and build on them.

“Some of our favourite mountains have been conquered,” University of Oxford professor Martin Bridson wrote, “but behind them are bigger mountains that we can tackle with new equipment.”

Stay ahead of the Stories shaping our world. Subscribe to Impact Newswire and join our 
WhatsApp Channel for updates on global tech, business, and innovation—all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Contact Us to reach a global audience with Impact Newswire.