Why artificial intelligence hasn't taken over mathematics — yet
August 26, 2026
Based on reporting from The Guardian → — simplified & explained by VAIIYA.
Last month, around forty top mathematicians gathered privately at OpenAI's headquarters. The purpose of the meeting: to discuss the future of their field in an era of rapid technological progress. The atmosphere around the meeting reflected clear concern. Many researchers fear for their careers and for the future of mathematics.
According to security technologist Bruce Schneier and mathematics professor Kasra Rafi, that panic is, at least in the short term, premature. While recent AI models achieve impressive results at the level of PhD-level researchers, the systems are still far from able to replace experienced academic mathematicians.
Surprising breakthroughs and clever combinations
The performance of top models in 2026 is undeniably remarkable. OpenAI previously reported that its latest model disproved the eighty-year-old unit distance conjecture from discrete geometry. Anthropic published results in the field of cryptanalysis, and had its Claude AI model make an attempt at proving the famous Riemann hypothesis.
Yet according to Schneier and Rafi, these successes mainly fall into two categories. On the one hand, there's finding counterexamples to theorems that humans were actually trying to prove. On the other, there's applying known techniques from one mathematical domain to problems in another domain.
An AI has the advantage of a gigantic working memory and can draw connections between fields that a human expert wouldn't quickly combine. For instance, the unit distance conjecture was disproven by using insights from algebraic number theory — an angle that human specialists simply hadn't thought of directly.
The lack of conceptual innovation
Making unexpected connections and searching enormous computational spaces is a form of creativity. It's comparable to the way AI systems beat grandmasters at Go or predict protein structures. But developing a genuinely new conceptual framework remains, for now, out of reach.
Much mathematical progress comes from identifying the essence of a problem and building an entirely new theory to understand it. Current AI models are extremely strong at searching and recombining existing knowledge, but weak at building deep, lasting new concepts.
Schneier and Rafi emphasize that this is a fundamental limitation of the current generation of artificial intelligence. AI can regroup existing ideas in an original way, but the system lacks the capacity for genuinely new intellectual structures.
Whether that boundary will hold is the question. Because many mathematical abilities are emergent properties of ever more powerful models, the authors expect that AI will eventually acquire this form of creativity too. Until then, developing profound new theories remains the exclusive domain of humans.