August 6, 2026
What AI Actually Does in Math Research
I find AI quite useful when I am doing some kind of research. It connects topics across different areas of math. Since I am just an undergraduate student, I haven't seen several theorems and results yet. So several times, AI just points me toward a specific area, and then I go look at it, instead of scrolling through random math topics hoping something clicks. It also helps me learn any new topic much faster. I open the related book on one side and keep an AI chat tab on the other.
What I find more interesting is the intersection thing. A lot of important mathematics happened exactly at the boundary between two fields, someone who knew both noticed they were secretly the same thing. Algebraic topology, algebraic geometry, the whole relationship between geometry and physics. Those didn't come from going deep in one subject alone. They came from someone who had enough breadth to see the connection. AI can accelerate the breadth part. It can say, "this thing you're doing in algebra looks like what people in topology call X", and then you go deep yourself.
It's the same as having good problem solving skills and knowing the syntax. In a math context, knowing definitions, theorems, and finding a way to connect them. Sometimes AI just helps you think in a different direction.
But I still see that AI works in a brute-force way. I have seen mathematicians come up with strangely simple-looking formulas that work perfectly for specific tasks. AI doesn't do that. It forcefully engineers a formula. The great mathematicians: Euler, Ramanujan, Grothendieck, their results often look simple after the fact precisely because they found the correct framing. AI can't find an angle nobody has seen before. It interpolates between known framings.
What it can do is eliminate the drudgery around the edges, checking whether a theorem you half-remember actually exists, translating a definition from one notation to another, confirming a proof step. That's genuinely useful.
But the core insight, the "why does this even work", that still has to come from a human sitting with the problem long enough to see it differently. After some refinement though, AI can still be good.