Artificial intelligence is rapidly changing how mathematicians work, with recent breakthroughs in solving long-standing problems. Researchers are leveraging tools like ChatGPT to unearth forgotten proofs and even generate new solutions, marking a potential shift in mathematical discovery. The trend began last October when mathematician Mehtaab Sawhney used ChatGPT to find a solution to Erdős problem #339, a conjecture left unsolved for nearly two decades after Paul Erdős’s death.
AI as a Research Assistant
The use of AI in mathematics isn’t about replacing human researchers but augmenting their capabilities. LLMs excel at sifting through vast mathematical literature, identifying obscure references, and even combining existing theorems to produce novel proofs. In some cases, AI has independently constructed valid proofs with minimal human input. Since October, AI tools have helped resolve roughly 100 Erdős problems, turning them from “open” to “solved”.
Beyond Literature Search: Meaningful Suggestions
While initially used as a powerful search engine, LLMs now offer more than just retrieval. Mathematicians report that AI can provide valuable suggestions, guiding researchers towards solutions they struggled with independently. Andrew Sutherland of MIT notes that “mathematicians whose only experience with LLMs is with earlier models don’t yet fully appreciate this.”
Testing the Limits: First Proof Challenge
To rigorously test AI’s mathematical skills, the First Proof team launched a challenge, presenting eleven unsolved proof segments to LLMs. The goal: to determine if AI can generate valid proofs within a week—a timeframe shorter than many human mathematicians require. The experiment has already attracted hundreds of participants submitting AI-generated solutions, though verification remains a major hurdle. Lauren Williams of Harvard emphasizes that “verification is a problem because 90 percent of the time it will come up with a solution… it’s going to write something and sound confident about it.”
A Pivotal Year for AI in Mathematics
Despite the hype, the current impact remains limited. No major mathematics journal has published a peer-reviewed proof explicitly citing AI assistance, though this is expected to change in 2026. Ravi Vakil, president of the American Mathematical Society, recently co-authored a preprint documenting how Google’s LLM aided in solving a math problem relevant to his research, signaling a shift in academic practice.
The Erdős problems serve as an effective benchmark, but mathematicians recognize the need for more substantial tests. Carlo Pagano, collaborating with Google’s DeepMind, stresses that the focus should shift towards problems with broader implications.
Ultimately, AI’s integration into mathematical research is inevitable. Mathematicians are already adapting, with some even taking leaves from academia to join AI companies. This reflects a growing consensus that AI will fundamentally reshape how mathematics is done.
