In early August, top mathematicians gathered at the offices of ChatGPT-maker OpenAI in San Francisco for a summit on an existential question: What will be left for human math experts to do if artificial intelligence soon leaves them all behind?
“I think there’s a chance that we end up in a world where there’s no high-quality mathematics research, and human expertise in mathematics is totally lost,” said Daniel Litt, a professor at the University of Toronto. He gave a talk at the event titled “The End of Mathematics.”
Litt’s presentation at the day-long meeting sketched out a worst-case scenario in which AI churned out advanced research in such quantities that there was little incentive to keep human math experts around. He considers that extreme outcome unlikely but said mathematicians urgently need to adapt. “It’s important we actually do something,” Litt said.
Mathematicians belong to one of the first academic disciplines to be forced to grapple with the professional consequences of AI-generated discoveries that can rival the work of experts.
The OpenAI summit had been planned for months but took place shortly after the company released a list of 10 new and notable AI-generated results in math and computer science from multiple subfields that each usually require years of work to master. (The Washington Post has a content partnership with OpenAI.)
The struggle for the future of human mathematics is taking place at a time of widening concern among lawmakers, economists, business executives and the American public about the potential for AI to replace white-collar workers, disrupting the job prospects of millions of people.
Some mathematicians see their current situation as a preview of what may await other fields or sectors of the economy.
“I think the reason the AI companies have chosen — some of them — to focus on mathematics is because it’s a proxy for almost any activity that requires a lot of intelligence,” said Drew Sutherland, a mathematician at the Massachusetts Institute of Technology who participated in the OpenAI meeting.
“These capable intelligences are going to be applied elsewhere and to other fields. And maybe we’re just the canaries in the coal mine that it’s coming for the mathematicians first.”
OpenAI’s gathering of mathematicians, comprising about 40 people, was meant to get attendees talking about the future of education and research in their field given that at some point AI will become “robustly superhuman” at research-level math, said co-organizer Jacob Tsimerman.
Last month, he shocked many colleagues when he used the ceremony where he received the field’s closest equivalent to a Nobel Prize to announce that he would soon be joining OpenAI to work on AI safety.
“In terms of what the community of mathematicians wants, there’s not consensus,” said Tsimerman, who has not yet started at the company. “The point was much more to have a dialogue than to do any one thing in particular.”
Hey, ChatGPT, ‘do a breakthrough’
Leading AI companies have for years made lofty promises that at some time in the near future or long term, their models will advance and accelerate scientific discovery beyond human ability. The predictions served as corporate marketing for the importance of their technology, while also making math and science and the people who work in those fields into collaborators or targets.
In 2025, an advanced version of Google’s Gemini earned a gold medal score at the International Mathematical Olympiad, a global competition for top high school students. Both OpenAI and Anthropic, maker of the chatbot Claude, released new products in recent months geared toward advancing research for scientists and academics by assisting them with their work.
In May, OpenAI announced what many mathematicians credit as the first major AI-generated breakthrough. One of the company’s unreleased models provided an example that disproved the unit distance conjecture, a long-standing question about possible arrangements of dots on an infinite sheet of paper. It did so using a well-known but tricky technique from algebraic number theory, a different area of math.
At the time, Tsimerman said he would have accepted the answer into any journal “without hesitation.”
Since then, OpenAI and Anthropic have released a series of competing math results. People using their AI models to generate math findings have begun to bypass the field’s traditional pipeline for new ideas and discoveries, sometimes with casual flair.
Math research is usually collaborative, with discoveries chewed over in seminars, passed around experts and posted to the preprint site arXiv before being formally peer reviewed and published in a journal.
Last week, Levent Alpöge, a mathematician at Anthropic, chose to instead announce a new result with a 24,000-character post on the social platform X that contained only the symbols + and -.
+++-+-+-+-+–+–+++-++—-++–+—-+—–+++—++-+—+++++++-+–+-+-+–++-+–+–+—-+-+-+-+-++-++—+–++++–++-++++-+++++–+—++-+—+++++++++–+-+-+–++-+–+–+—–+++-+++-++-+-++++-++-+–+–+——++–+-++-+-+—+–+-+-+-+–++—+++++—–+++++++—+—+–+-+–+-+–+-++-++…
— levent (@__alpoge__) August 12, 2026
The wall of text was an AI-assisted rundown of pairs of perpendicular vectors pointing to the corners of a 668-dimensional cube — a construction human mathematicians believed to be possible but that no one had been able to prove out. Alpöge made the post with help from Claude. He and Anthropic did not respond to a request for comment.
In July, start-up founder Dmitry Rybin asked ChatGPT to “do a breakthrough” to disprove a hypothesis about network flows. Each time the model failed, he prodded it to keep going. ChatGPT eventually produced an example refuting the claim that he checked and tweeted out, he said, while watching a movie with friends.
“You drop a casual tweet saying you prove the big problem while cooking pasta and having a drink with your friends? Of course this is going to get viral,” said mathematician David Bessis, citing an example that is for now hypothetical. “The problem is, what happens after that? What are the incentives for people to sort out the mess and try to make sense of everything?”
Some mathematicians using AI translate their results into a programming language called Lean that allows a computer to check mathematical logic step by step. But Lean translations are lengthy and often pedantic — different from the way researchers write out proofs in formal papers intended to be legible to humans.
In June, concerns about AI-generated math results causing confusion over what has and hasn’t been proved in the field helped lead to the release of an open letter called the Leiden Declaration, signed by more than 3,000 mathematicians. It urges researchers to use AI tools responsibly, including by ensuring that their work is verified and cited properly when distributed in public. Others in the field have argued that mathematicians should avoid use of AI, and AI companies, altogether to preserve the role of humans.
‘Take a real stab … keep going’
Although many math researchers have been impressed by AI-generated mathematical constructions, others say they have fallen short of breaking open significant new areas of theory. Math is a vast subject, with many open problems that have not been directly tackled by humans for years — giving large language models an opportunity to try new arguments on previously identified problems.
The list of 10 new results released by OpenAI this month included a novel example of an mathematical object called a sofic group, a vast and infinite structure that cannot be approximated by any sort of finite representation. The new result bridged the gap between two papers published in 2016 and 2019 by mathematicians Gábor Kun and Andreas Thom. The pair quickly published a follow-up paper that simplified and built on OpenAI’s new finding.
“AI is problem-solving in a very clever way, in a significant way,” said Thom. “But the concepts that now emerged, they emerged, again, as part of the engagement of humans with the proof. So it’s not yet the AI that does it.”
Some recent breakthroughs have come from people without a formal math background. Last week, Anthropic announced that one of its employees had asked an internal version of Claude to “take a real stab” at the Riemann hypothesis, one of the field’s biggest open questions with a prize of $1 million on offer for a solution. After being repeatedly spurred on with prompts like “keep going,” the model eventually produced a new finding on a related problem.
Some mathematicians say the rush to find solutions via machine could jeopardize human understanding.
“Mathematics is about understanding results. It’s not just about proving the result,” said Bryna Kra, a math professor at Northwestern University, who attended the OpenAI meeting and was one of 16 organizers of the Leiden Declaration. “A proof that doesn’t get understood doesn’t become adopted as part of the literature,” she added.
I’m currently returning to Toronto from a summit on the future of mathematics, at OpenAI. @SebastienBubeck asked me to talk a bit about the future we’d all like to avoid, where humans are mathematically disempowered. @Jacob_Tsimerman advised us to try to prioritize detail over… pic.twitter.com/HRfyjlLCA8
— Daniel Litt (@littmath) August 11, 2026
Melanie Matchett Wood, a mathematician at Harvard University who helped prepare a human-written version of OpenAI’s unit distance result, said that top AI models seem to struggle to clearly explain their work. They tend to overexplain easy parts of an argument and breeze through the hard parts, she said.
“The top AI models are not quite able to understand what the difficult parts of the argument are, and exposit in a way that brings those out,” said Wood, who also attended the meeting at OpenAI.
“This is an entirely new kind of way for mathematics to be produced,” she added. “We don’t yet have any norms for how it should be done.”
Sébastien Bubeck, a research scientist and mathematician at OpenAI who helped organize its summit this month, said he was hopeful that humans can use AI to develop more powerful ways to do math. Attendees discussed other possible futures in addition to Litt’s worst-case scenario, he said. “This is not necessarily a future that we want,” he said.
Those other paths could see math become more like software engineering, where AI helps hundreds of people to work together on a common problem, Bubeck said, or like physics, where advanced problems are solved by amassing immense resources and AI models are used like particle accelerators.
Another possible future may see math become like museum curation, he said, with AI making discoveries that humans pick from. Or maybe the field should consider pivoting to AI safety.
“We have to put the people, the human, the mathematicians first,” he said. “Mathematics is only interesting insofar as mathematicians are learning from it.”
The post AI triggers a wave of anxiety among the world’s greatest math minds appeared first on Washington Post.
