How Mathematicians Use AI: 7 Real Research Workflows

Quick answer: Mathematicians use AI in seven main ways: finding references, talking through ideas, writing code for experiments, searching for constructions with tools like AlphaEvolve, drafting proof ideas, formalizing proofs in Lean, and drafting and editing papers. In each case the researcher supplies the question, AI returns a draft (a reference list, code, a construction or an argument), and the result is checked: by reading the source, testing the code, reviewing the proof or verifying it in Lean.

Sources checked November 22, 2026.

Headlines about AI solving famous problems hide a quieter story: how working mathematicians use AI on an ordinary Tuesday. Researchers have now documented AI use across almost every stage of research, not to replace their thinking but to speed up the parts that used to take weeks. Here are seven real workflows, each with what goes in, what comes out and how it is checked.

The 7 Workflows at a Glance

WorkflowResearcher suppliesAI returnsHow it is checked
1. Literature searchA question or statementCandidate referencesRead the original source
2. BrainstormingA proof ideaObjections, related techniquesHuman judgment
3. Code for experimentsA conjecture to testA program and its outputReview the code, test small cases
4. Construction searchA scoring ruleBetter examples (AlphaEvolve)Verify the construction exactly
5. Drafting proofsA precise statementA proposed argumentLine-by-line expert review
6. Formal verificationA proof and its statementLean codeLean’s checker, plus a check that the statement is right
7. Writing papersNotes and resultsLaTeX drafts, editsAuthor review, journal policies

Which Tool Does What?

  • Language models (chat assistants) generate text, arguments and code.
  • Coding agents can edit files and run programs.
  • Python and SageMath do the actual computation; they are not AI themselves.
  • AlphaEvolve searches over programs using an evaluator that scores answers.
  • Lean is a proof assistant, and Mathlib is its library of formalized mathematics.

1. Literature Search

Mathematics has a vast literature, and a result you need may sit in an old journal under a different name. Language models are good at matching ideas across wording; in an April 2026 Quanta interview, Igor Pak (UCLA) called ChatGPT “fantastic in finding the right references.” This cuts both ways: on the Erdős problems list, AI found forgotten papers that had already answered “open” questions (see Erdős problems solved by AI). The check: ask AI where to look, then read the source yourself, because AI can also invent citations.

2. A Brainstorming Partner

Explaining an idea out loud helps you find its gaps. Johannes Schmitt (ETH Zurich) uses language models as conversation partners: he describes an approach, asks what could go wrong and asks for related techniques. He continues despite frequent errors, because a partner who is often wrong can still spark the right idea.

3. Writing Code for Experiments

Before proving a statement, mathematicians test it on many examples, for example by computing the first thousand cases or searching a list of graphs. AI writes that code quickly. Geordie Williamson said an experiment that would have taken about two weeks a couple of years earlier now takes about 20 minutes. The check: read the code and confirm it on small cases you can do by hand.

4. Searching for Examples and Counterexamples

Some questions ask for the best possible construction: the most points with a property, the tightest packing, the largest set. Google DeepMind’s AlphaEvolve evolves programs that build candidates and scores them. In a November 2025 paper, Bogdan Georgiev, Javier Gómez-Serrano, Terence Tao and Adam Zsolt Wagner tried it on 67 problems; it matched the best known answers in most cases and improved several. In one case the team went from an AlphaEvolve construction, to a proof, to a Lean formalization. Gómez-Serrano estimated in April 2026 that he spends about two-thirds of his time using AI. In a preprint posted January 3, 2026, Jordan Ellenberg, Williamson and colleagues used AlphaEvolve to suggest patterns about Bruhat intervals, then proved a general result themselves. Since July 2026 AlphaEvolve has been generally available on Google Cloud.

5. Drafting Proofs

Reasoning models can now propose complete arguments. In October 2025, Uijeong Jang and Ernest Ryu (UCLA) posted a proof, developed with ChatGPT, that the iterates of Nesterov’s accelerated gradient method (introduced in 1983) converge, with Ryu checking each step. A January 2026 revision removed an auxiliary claim after an error was found in an AI-generated argument; the main result stood. That is exactly why review matters. In January 2026, Ravi Vakil and coauthors posted an algebraic geometry result developed with Google’s Gemini Deep Think and a specialized Gemini-based system, and their paper separates the AI’s contributions from the human-written exposition.

6. Formal Verification in Lean

Proof assistants such as Lean check every logical step. Writing Lean used to be slow; AI now helps translate a human proof into Lean code and fill in routine steps. Two questions remain for people: does the formal statement say what we mean, and what does the proof teach us? On November 21, 2026, the Lean and Mathlib communities certified both the statement and the proof of a Lean proof of the Riemann Hypothesis developed with Claude, a dated example of this division of labor (see AI math breakthroughs of 2026).

7. Writing and Checking Papers

AI drafts LaTeX, tidies notation and catches inconsistencies; Schmitt has described using it to draft sections he then rewrites. Penn computer scientists Michael Kearns and Aaron Roth, writing in March 2026, described using AI agents over three weeks in summer 2025 to help develop code and lemmas for a 50-page paper with Emily Ryu, while the exposition stayed human. They estimated that the generated proofs were “correct perhaps only three-quarters of the time,” an informal estimate from one project, not a general accuracy rate.

A Mini Example: Testing a Conjecture

Suppose you suspect that \(n^2 + n + 41\) is always prime. A researcher might ask AI for a short program to test \(n = 0\) to \(100\). The program reports primes up to \(n = 39\), then \(n = 40\) gives \(40^2 + 40 + 41 = 1681 = 41^2\). The check is easy to do by hand, and it shows the full loop: AI writes code, the computer runs it, and the human verifies and interprets the counterexample.

The Rules Experienced Users Follow

  • Verify everything. Read every step, or check it in Lean. Kearns and Roth compared AI to a smart but occasionally error-prone colleague.
  • Check the literature. A “new” AI result may be decades old.
  • Keep the understanding human. At ICM 2026, Terence Tao stressed that people must still explain results, and that mathematicians, not companies, should set the rules: “We set the rules on what’s acceptable or not.”
  • Learn without AI first. Kearns and Roth suggest early-career researchers build skills without AI before leaning on it.
  • Follow venue policies. Many journals now require disclosure of AI use and hold authors responsible; some restrict AI in confidential peer review.

What Students Can Borrow From This

The same habits work at every level: use AI to find resources, test ideas on examples and get feedback, but do the reasoning yourself and verify every answer. See Is Math Still Worth Learning? for study tips, and Can AI Do Math? for where AI still goes wrong.

Common Misconceptions

“Mathematicians just ask AI for proofs.” Most use is quieter: references, code, examples and feedback.

“AI output can be trusted if it sounds confident.” No. Even strong users report frequent errors; checking is part of the workflow.

“Disclosure makes any AI use acceptable.” Not quite. Venues set their own rules, and authors stay responsible for every claim.

Watch: Mathematicians and AI Tools in Video

If you learn better by watching, these videos are worth your time.

How Terence Tao uses AI with Lean programming language (Lex Clips). Terence Tao explains how AI and the Lean proof assistant fit into his work.
First Proof: A Nuanced Assessment of AI’s Ability to Solve Math Problems (Harvard Mathematics Department). A talk on the First Proof challenge, where mathematicians tested AI on unpublished research questions.
Google’s AlphaEvolve is making new discoveries in math… (Fireship). A fast overview of how AlphaEvolve evolves programs to find new mathematical constructions.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. In the 2025 AlphaEvolve paper by Tao and colleagues, how many problems were tested?
Found a bug? Let us know!
2. What did Kearns and Roth report about AI-generated proofs?
Found a bug? Let us know!
3. What is Lean used for?
Found a bug? Let us know!
4. Why must AI-found references be checked?
Found a bug? Let us know!
5. According to Geordie Williamson, an experiment that took two weeks now takes about how long?
Found a bug? Let us know!

Frequently Asked Questions

How do mathematicians use AI?

Mainly for literature search, brainstorming, writing code for experiments, searching for constructions with tools like AlphaEvolve, drafting proofs, formalizing proofs in Lean, and writing and checking papers. They verify everything AI produces.

Does Terence Tao use AI?

Yes. He coauthored a 2025 paper testing Google DeepMind’s AlphaEvolve on 67 problems, uses AI with the Lean proof assistant, and has argued that mathematicians should set the rules for acceptable AI use.

How accurate are AI-generated math proofs?

It varies. In one 2026 account, Penn researchers Michael Kearns and Aaron Roth found AI-generated proofs were correct perhaps only three-quarters of the time, an informal estimate from one project, which is why mathematicians review every proof or verify it in Lean.

What is AlphaEvolve used for in math?

AlphaEvolve evolves computer programs that construct mathematical objects, such as packings or point sets, and keeps the best ones. Mathematicians use it to improve bounds and find examples or counterexamples.

Can anyone use AlphaEvolve?

Yes. Google made AlphaEvolve generally available on Google Cloud on July 9, 2026. You supply a starting program and an evaluator that scores answers.

Should math students use AI like researchers do?

Borrow the habits: use AI to find resources and get feedback, but do the reasoning yourself and verify every answer. Researchers recommend building skills without AI first.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "How Mathematicians Use AI: 7 Real Research Workflows - Effortless Math"?

No one replied yet.

Leave a Reply