Mathematics has always been the discipline most proud of its purity - exact definitions, airtight proofs, truths that hold regardless of who discovers them. Then artificial intelligence showed up, and suddenly the most “human” of intellectual pursuits is being shared with machines.
The impact is not a simple story of “AI is better at math now.” It’s a story of disruption, collaboration, and a quiet philosophical crisis about what proof even means. Let’s break it down.
🏆 The headline moments
AlphaProof and the IMO
In July 2024, DeepMind’s AlphaProof system (combined with AlphaGeometry 2) achieved a silver-medal level performance at the International Mathematical Olympiad, solving 4 out of 6 problems. The IMO is the most prestigious competition for pre-university mathematicians - problems there require genuine creativity, not calculation.
This was the moment many mathematicians stopped saying “AI can’t do real math.”
AlphaTensor and matrix multiplication
Earlier, AlphaTensor discovered a faster algorithm for multiplying 4x4 matrices - improving on a result (Strassen’s algorithm) that had stood for over 50 years. It searched a space of possibilities larger than the number of atoms in the universe and found something humans had missed.
Function-to-term breakthroughs
DeepMind’s work with funSearch found new results for the cap set problem, and AlphaEvolve in 2025 improved matrix multiplication and other combinatorial results further. The pattern is consistent: AI as a discovery engine for objects too large or strange for human intuition.
🤝 Three ways AI is changing mathematical practice
1. Conjecture generation
Large language models trained on mathematical literature have absorbed the “taste” of generations of mathematicians. Ask a modern LLM about a pattern in your data, and it will often propose conjectures that feel plausible - sometimes ones that turn out to be provable.
Terence Tao, arguably the most famous living mathematician, has been openly experimenting with AI-assisted research. He described using LLMs as a “colleague who has read a lot but is sometimes unreliable” - great for brainstorming, dangerous if trusted blindly.
2. Proof assistance and formalization
This is where things get philosophically interesting. Systems like Lean (a proof assistant) force mathematicians to write proofs so precisely that a computer can verify every step. Combined with AI:
- AlphaProof works in Lean - it generates proofs that are machine-verified, so no “hallucination” can slip through.
- Massive formalization projects are accelerating. The Liquid Tensor Experiment formalized a very technical theorem of Peter Scholze’s in Lean - work that would have taken a team of humans years.
- AI models are increasingly good at translating informal human proofs into formal Lean code, closing the gap between “mathematician intuition” and “machine-checkable truth.”
3. The end of “trivial but tedious”
A huge amount of mathematical work is unglamorous: verifying cases, checking boundary conditions, exhausting small configurations. AI is genuinely great at this. Theorems that were known to be “true for the first 10,000 cases but nobody proved it in general” are becoming either fully proven or fully refuted faster than ever.
⚠️ The problems AI brings
It’s not all celebration. Several real concerns have emerged:
Hallucinated proofs
LLMs confidently produce plausible-looking proofs that are flat-out wrong. Unlike code (which fails when you run it), a fake mathematical proof can look fine until a careful human finds the subtle error. This makes raw LLM output dangerous for mathematics unless paired with formal verification.
The review crisis
Mathematics papers increasingly rely on computational verification that referees cannot easily check. When an AI system produces a 500-step proof, how does a journal referee verify it? The peer-review system built for human-readable arguments is straining.
What happens to mathematical intuition?
If AI finds proofs by searching spaces humans can’t conceptualize, the proofs may be correct but incomprehensible. A machine-generated proof of a famous conjecture might tell us the answer without giving us the understanding we actually wanted. Many mathematicians care more about “why is this true” than “is this true.”
The career pipeline problem
Mathematics has traditionally trained young researchers by giving them tractable problems. If AI solves all the tractable problems, what’s left to train on? The field risks a gap between the problems humans can still do and the problems worth doing.
🔮 Where this is heading
My honest take on the next few years:
Formal verification becomes standard. Major results will increasingly be expected to ship with Lean proofs. AI makes formalization cheap, so it will happen everywhere.
Mathematicians become directors. The skill shifts from “can execute a long computation” to “can ask the right question and judge which machine-generated candidates are worth pursuing.”
New fields emerge at the boundary. “AI-assisted discovery” is becoming its own subdiscipline, with its own standards, tools, and conferences.
Education changes fundamentally. If AI can do the routine problem-solving we use to train students, mathematics education must pivot toward conceptual depth, formal reasoning, and taste - the things that remain distinctly human for now.
💭 Closing thought
The most interesting question isn’t whether AI will replace mathematicians. It’s whether mathematics itself stays the same discipline when machines can do parts of it.
Euclid didn’t stop being relevant when calculators appeared. Genuinely deep mathematics - the kind that reorganizes how we think - is likely to remain a human-AI collaboration for a long time. But the craft of doing mathematics is changing faster right now than it has in centuries.
If you work in a technical field, this is worth watching closely. The same disruption pattern will hit your discipline next.
What do you think - is AI-discovered mathematics “real” mathematics? Feel free to reach out through the contact page if you’d like to discuss.
