When AI Became a Mathematician: The Year Machines Started Making Discoveries

0

For centuries, mathematics has been regarded as one of humanity’s highest intellectual achievements. Unlike games such as chess or Go, mathematics demands abstraction, creativity, logical rigour and the ability to discover entirely new knowledge. Many researchers believed it would remain one of the last frontiers resistant to artificial intelligence.

That belief is rapidly disappearing.

Over the past few years, artificial intelligence has progressed from solving textbook mathematics to making genuine contributions to frontier mathematical research. The pace of development has been extraordinary, raising an important question: are we witnessing the birth of AI as a mathematical collaborator rather than simply a computational tool?

The first major breakthrough came in 2024 with the release of AlphaGeometry, a neuro-symbolic AI developed by Google DeepMind. Unlike previous theorem provers, AlphaGeometry combined a large neural network with symbolic deduction, allowing it to solve International Mathematical Olympiad (IMO) geometry problems requiring genuine mathematical reasoning rather than brute-force computation. The system solved 25 out of 30 Olympiad geometry problems, approaching the performance of an average human gold medallist (Trinh et al., 2024).

The achievement was significant because geometry has traditionally been one of the most difficult areas for automated theorem proving. Successful solutions require intuition about auxiliary constructions, hidden relationships and elegant logical deductions—abilities long considered uniquely human.

Yet this was only the beginning.

Later in 2024, Google DeepMind announced another milestone with AlphaProof, which combined reinforcement learning with formal proof verification. Together with AlphaGeometry2, the systems achieved silver-medal standard at the International Mathematical Olympiad, solving four of the six official competition problems under contest conditions (Google DeepMind, 2024).

Only months later, AlphaGeometry itself evolved.

AlphaGeometry2 substantially expanded the original system’s capabilities by improving both its neural reasoning and symbolic search algorithms. The upgraded model solved approximately 84% of every geometry problem set at the International Mathematical Olympiad between 2000 and 2024, surpassing the performance expected of an average gold medallist (Chervonyi et al., 2025).

This progression was remarkable.

Within little more than a year, AI had moved from being competitive with elite human students to consistently outperforming them on one of the world’s most demanding mathematical competitions.

Perhaps even more impressive was what happened next.

In 2025, Google DeepMind introduced AlphaEvolve, a system designed not merely to prove theorems but to discover entirely new mathematical algorithms. Rather than generating proofs directly, AlphaEvolve used evolutionary search guided by large language models to create and refine algorithms automatically. The system improved matrix multiplication algorithms, optimised Google’s own data-centre scheduling systems, enhanced AI training procedures and discovered better solutions to numerous long-standing optimisation problems (Google DeepMind, 2025).

Most strikingly, AlphaEvolve tackled more than fifty open mathematical problems spanning geometry, combinatorics, number theory and mathematical analysis. According to DeepMind, it successfully rediscovered state-of-the-art solutions in around three-quarters of the problems and produced improved solutions in approximately one-fifth of them, demonstrating genuine mathematical creativity rather than simple memorisation (Google DeepMind, 2025).

Independent researchers have since extended this work even further. Mathematical exploration systems combining AlphaEvolve with formal theorem provers have shown the ability to discover new constructions, generate conjectures and assist mathematicians in advancing long-standing research questions, suggesting a future in which AI becomes an active research collaborator rather than merely a computational assistant (Georgiev et al., 2025).

These developments represent a profound shift in the philosophy of mathematics.

For centuries, mathematical progress has depended upon individual insight. Great mathematicians such as Euclid, Euler, Gauss, Noether and Wiles transformed entire disciplines through flashes of human creativity. AI introduces something fundamentally different. Instead of relying upon intuition developed over decades, machines can systematically explore millions of possible approaches simultaneously while formally verifying every logical step.

This dramatically changes the economics of discovery.

Problems that might occupy research teams for years can now be explored by AI systems in hours or days. Rather than replacing mathematicians, AI increasingly acts as an extraordinarily productive research partner capable of searching enormous mathematical landscapes beyond human cognitive limits.

The implications extend well beyond mathematics.

Many scientific disciplines ultimately reduce to mathematical optimisation. Physics, chemistry, biology, engineering, economics and computer science all rely upon discovering patterns hidden within immense solution spaces. If AI can systematically accelerate mathematical discovery, similar revolutions may soon occur throughout science itself.

Nevertheless, caution remains appropriate.

Success on benchmark problems does not necessarily imply human-like understanding. AI systems remain dependent upon carefully designed architectures, formal verification systems and vast computational resources. Human mathematicians continue to provide the intuition needed to formulate interesting questions, evaluate significance and connect discoveries across diverse branches of knowledge.

Indeed, many researchers argue that the future of mathematics will not belong solely to humans or machines.

Instead, it will belong to their collaboration.

Mathematicians may increasingly spend their time asking profound questions while AI explores enormous numbers of potential solutions, generates conjectures and constructs rigorous proofs. The role of the mathematician shifts from performing every calculation to directing discovery itself.

That possibility may prove to be the most important mathematical breakthrough of all.

Artificial intelligence is no longer simply learning mathematics.

It is beginning to create it.

References

Chervonyi, Y. et al. (2025) Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2. arXiv:2502.03544. Available at: https://arxiv.org/abs/2502.03544 (Accessed: 23 July 2026).

Georgiev, B. et al. (2025) Mathematical Exploration and Discovery at Scale. arXiv:2511.02864. Available at: https://arxiv.org/abs/2511.02864 (Accessed: 23 July 2026).

Google DeepMind (2025) AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms. Available at: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ (Accessed: 23 July 2026).

Trinh, T.H. et al. (2024) ‘Solving olympiad geometry without human demonstrations’, Nature, 625, pp. 476–482.