How AI Is Changing the Nature of Mathematical Research

AI is now changing not just software coding but also the way mathematical research is done. Researchers are experiencing a workflow in which giving AI just a high-level proof sketch is enough for it to organize this into rigorous theorems and proofs.

Over 3 weeks last summer, the authors used agentic AI tools to write a 50-page paper that would normally have taken months. The paper dealt with an optimization problem based on graph theory and machine learning concepts, and the AI accepted abstract descriptions like a "directed acyclic network" and translated them into formal mathematical language.

The key point is that AI does more than simply polish sentences—it converts incomplete intuitions into definitions, theorems, and proofs. In this process, the authors also mention an experience where the AI discovered a small auxiliary lemma on its own during a proof and simplified the logic, and they see AI as evolving to the point where it sometimes even contributes research ideas itself.

However, it is not yet a reliable collaborator. The authors estimate that AI-generated proofs are only about 75% accurate, and explain that productive collaboration is only possible if you can catch errors and iterate on fixes. If you follow incorrect answers as they are, you quickly hit a dead end, and it's easy to slide into AI research slop—content that looks smooth on the surface but is thin in substance.

The issue isn't just research productivity. If AI takes over the trial and error that junior researchers used to have to go through, it becomes harder for young researchers to learn intuition and good taste. The authors suggest that, just as people once learned to do addition by hand in the age of calculators, education may need at least one track where research is still learned the "old way."

Peer review is also taking a hit. AI makes it possible to mass-produce papers faster and more polished, adding further strain to an academic publishing system that is already overloaded. In particular, since a reviewer's role is not simply a consistency check but deciding where to direct the community's scarce attention, the authors argue that this function needs to be redesigned for the AI era.

As an alternative, they propose making more active use of automated verification tools such as formal verification, treating error detection in mathematical proofs somewhat like unit testing. The idea is that human reviewers should focus less on consistency checks—something machines are good at—and more on judging why a result matters and what new understanding it brings.

Ultimately, the authors see AI as bringing a sea change that simultaneously transforms the speed of research, education, and the review system. The remaining challenge, they argue, is not to block AI, but to redesign institutions so that AI amplifies human creativity and insight while preserving the rigor of science and the joy of research.