Artificial intelligence is no longer just assisting with routine mathematical work. It is now producing counterexamples, helping formalize proofs, and pushing into open problems that once looked firmly outside the reach of machines.
That progress has created a complicated moment for mathematics. Some researchers see the arrival of AI in mathematics as the beginning of a more productive era. Others worry that the field could gain answers faster than it can preserve the human culture needed to understand them.
Open problems are starting to fall
In May 2026, OpenAI published a counterexample to the "Unit Distance Conjecture," disproving a problem in geometric graph theory that had been open since 1946. The conjecture was one of hundreds of open problems associated with Hungarian mathematician Paul Erdos.
AI had already contributed to other mathematical problems, including other Erdos problems. But many mathematicians viewed this result as the most important example so far. One week later, human researchers adapted the central proof technique and used it to disprove another major conjecture.
Since then, AI has become a steady presence in mathematical research news. Models are being used to find counterexamples, detect broader patterns, and help convert mathematical arguments into proofs that can be checked by machines.
Epoch AI, known for its demanding "FrontierMath" AI benchmark, recently announced the second solution in "FrontierMath: Open Problems," a test based on major unsolved questions. OpenAI's new Astra model also appears to be aimed at this kind of work, and the lab introduced it with ten solutions of varying difficulty.
Some researchers see a new kind of collaborator
For Abhishek Saha, a math professor at Queen Mary University of London, frontier AI models are already useful in serious research settings. He wrote on X, "At the moment, in my area of research mathematics, frontier AI models are at least as good as a solid and indefatigable PhD student."
Saha spent a full day using GPT-5.5 Pro on routine work that would previously have taken weeks. He described the shift as "increasingly playing the role of conductor, rather than doubling up as the whole orchestra."
That framing matters. In this view, AI is not replacing mathematical judgment. It is changing where that judgment is applied. The researcher spends less time grinding through every step and more time directing, evaluating, and integrating the output.
Saha also expects the profession to divide over adoption. He wrote, "Some will adapt soon, and find boundless possibilities," while others may remain skeptical until the end. His comparison was to "the folklore hero John Henry," the American folk hero who worked himself to death competing against a steam drill.
The golden age argument comes with a warning
Mathematician Trefor Bazett, writing in The Conversation, argues that AI and human ingenuity could help bring about a new golden age of mathematics. He points to an April 2026 paper from Carnegie Mellon University mathematicians that solved an open problem in Ramsey theory by combining SAT solvers, code generated by language models, and formal proof verification.
The paper connects its result to a "golden age" predicted in 2000 by Fields Medal winner Timothy Gowers. Gowers imagined computers handling routine checks while mathematicians focused on deeper ideas. As he put it, "In other words, computers would still do the boring bits for us, but these would not be quite as boring as they are now."
The Carnegie Mellon researchers believe that moment has arrived. They write, "We believe that we are now entering this golden age, thanks to the combination of several technologies."
But Gowers also warned that such a period might be brief. He wrote, "However, such a golden age, if it occurs, is unlikely to last for long." He further predicted that "during the next century computers will become sufficiently good at proving theorems that the practice of pure mathematical research will be completely revolutionized."
Now that this prediction appears to be taking shape, Gowers has mixed feelings. He says GPT 5.6 Pro twice solved a problem on its first attempt after he had spent considerable time on it. "It felt very strange and not particularly pleasant to have the rug pulled out from under my feet like that," he wrote on his blog, while also saying he was glad the problems were solved.
The limits are still real
The recent run of progress does not mean AI can solve all of mathematics. Bazett says AI performs much better in some areas than in others. Graph theory has proven especially suitable for machine-generated proofs, while other fields remain more resistant.
Epoch AI's benchmark shows where the boundary still sits. In the two hardest categories, "Major Advance" and "Breakthrough," AI has not solved a single problem. The six remaining Millennium Prize Problems, each carrying a $1 million prize from the Clay Mathematics Institute, remain out of reach for AI as well as humans.
OpenAI's Astra could not solve those problems either, although OpenAI researcher Noam Brown believes more computing power could change that. Bazett says students who spend years building mathematical skill have understandable reasons to fear replacement. But he adds, "Thankfully, we're not close to that yet."
Proof overload may become the next challenge
Mathematician Terence Tao offered a cautiously optimistic view in his talk at the 2026 International Congress of Mathematicians. He compared the current moment to the foundational crisis of the early 20th century, when paradoxes and incompleteness theorems forced mathematicians to rethink basic assumptions. That earlier crisis ultimately strengthened the field's foundations.
Tao argues that AI could force a similar reassessment of values and methods. If machines handle a large share of research tasks, mathematicians cannot measure progress only by how many open problems are solved.
Proofs still need to be checked, explained, understood by other mathematicians, and eventually absorbed into textbooks and broader theories. That work remains central even if AI makes discovery faster.
The risk is that proof scarcity could turn into proof overload. Results may arrive faster than people can review, process, and place them in context. In that world, mathematicians would still have a lasting role: deciding which results matter, how they should be presented, and what goals the field should pursue.
The Leiden Declaration on Artificial Intelligence and Mathematics reflects that shift in responsibility. More than 3,000 mathematicians have signed it, and it is backed by the International Mathematical Union. Rather than rejecting AI, it calls for transparency when AI tools are used, protection of authors' rights, and continued human responsibility for mathematical results.
The debate is no longer about whether AI belongs near advanced mathematics. It is already there. The harder question is how the field keeps human understanding, judgment, and culture at the center while the machines get better at producing answers.