AI Research Mathematician Career: What It Is and Who Hires

In May 2026, an OpenAI model disproved a conjecture Paul Erdős posed in 1946. Nine mathematicians published a paper confirming it. Two months later, multiple AI systems scored a perfect 42/42 on the International Mathematical Olympiad problems in Shanghai. Neither event replaced mathematicians. Both created work for them. The AI research mathematician career is the job of deciding which problem to hand a model, judging whether its output is a result or nonsense, and verifying it well enough that other mathematicians accept it. It pays like a research scientist, it needs a graduate degree, and the entry requirements start with things a 12-year-old can do for free this year.

Key Takeaways

• The role exists because verification became the bottleneck. As Thomas Bloom put it about AI-generated proofs: “It could be right. It could be nonsense. Who’s going to be able to check this?”
• BLS pay for the closest official categories (May 2025): mathematicians $126,710 median, computer and information research scientists $140,300. Lab-specific AI research roles pay well above both; those figures are estimates from job postings, not BLS data.
• Demand is small but growing fast. BLS projects mathematicians and statisticians up 10% from 2025 to 2035 (about 33,500 jobs total) and computer and information research scientists up 22% (about 38,600 jobs).
• The degree is real: a PhD for research roles, a master’s for verification and evaluation roles. Self-taught entry is rare here, unlike in software.
• What a 10–15-year-old can do this year, free: math circles or olympiad training, the Erdős Problems database, and learning Lean, a proof assistant that turns mathematics into code a computer checks.

What happened, and why it created a job

The Erdős unit distance problem asks how many pairs among n points in a plane can be exactly one unit apart. Erdős conjectured a ceiling in 1946. On May 20, 2026, OpenAI announced that a reasoning model had found a new family of point configurations exceeding that ceiling, using infinite class field towers from algebraic number theory. Noga Alon, Thomas Bloom, W. T. Gowers, Melanie Matchett Wood and five others published remarks confirming the result. Gowers said that if a human had written it, he’d have recommended acceptance without hesitation.

Then the hard part. The model was correct on roughly 50% of runs. OpenAI did not publish failure rates or compute time. And, as Quanta Magazine reported on August 3, 2026, the pace accelerated: OpenAI announced ten more advances on August 1, Google DeepMind resolved four Erdős problems and rediscovered nine forgotten solutions, and human-AI collaborations handled dozens more. Terence Tao co-authored one. Bloom’s worry moved from “can AI do mathematics” to “who reads the 200-page papers.”

That gap is the job. A lab that generates candidate proofs faster than anyone can check them needs people who can (a) choose problems where a result would actually matter, (b) read a machine-written proof critically, (c) formalize the key steps in a proof assistant so a computer verifies them, and (d) write the human-readable version other mathematicians will cite. That’s four distinct skills, and the people who have all four are rare.

What the job actually does, hour by hour

A day is not spent proving theorems by hand. Roughly:

Problem selection. Reading the literature to find questions that are open, well-specified, and consequential. The Erdős database is popular precisely because the problems are crisply stated and independently tracked.

Prompting and scaffolding. Setting up a model with the right context, sometimes with tool access to a proof assistant, and running many attempts. The 50% figure matters here: the job involves reading a lot of wrong proofs.

Triage. Deciding in twenty minutes whether an output is worth two days. This is the skill that decides whether a team ships results, and the hardest one to teach.

Formalization. Translating the critical lemmas into Lean or a similar system, where a computer checks every inference. Formalization is slow, unglamorous, and the only thing that settles disputes.

Attribution and writing. Wood’s objection is a real work item: models don’t cite, so a human has to trace which known results the proof depends on. In the Erdős case that meant identifying Golod-Shafarevich, Ellenberg-Venkatesh, and Hajir-Maire-Ramakrishna.

The AI research mathematician career path: high school to a working role

Stage	What to take or earn	What it costs	What you can do with it
High school	Algebra II, precalculus, calculus if available; competition math (AMC, AIME); one programming language; a proof-based elective or math circle	Free to low; AMC fees are small and often school-covered	Qualify for a strong undergraduate math program; start Lean on your own
Undergraduate	BS/BA in mathematics or math + computer science. Required: real analysis, abstract algebra, discrete math, number theory, one course in logic or formal methods	Varies widely; public in-state to private	Research assistant, data analyst, math-adjacent software roles
Certification (optional)	No license exists. The closest credentials are published formalizations: contributions to Lean’s mathlib library, visible on GitHub	Free	A portfolio that labs can actually inspect, which matters more than a certificate
Master’s	MS in mathematics, applied math, or computer science with a formal-methods focus	1–2 years	Verification engineer, evaluation researcher, technical staff at a lab
PhD	Doctorate in a field adjacent to what models are good at: combinatorics, number theory, graph theory, discrete geometry	4–6 years, usually funded	Research mathematician at a lab, university, or national institute
Day to day	Problem selection, prompting, triage of machine output, formalization in Lean, attribution, writing	—	Publishing results; deciding what the lab works on next

Two honest notes on this table. First, the PhD is not optional for the research version of the role, which is unusual in AI where self-taught engineers thrive. Second, the master’s-level verification roles are the fastest-growing and least-discussed entry point, because every lab producing proofs needs more checkers than provers.

Pay and demand, without the hype

There is no BLS occupation code for “AI research mathematician,” so anyone quoting a precise salary is estimating. Here is what is solid and what isn’t.

Role	Source	Median pay	Outlook
Mathematicians	BLS, May 2025	$126,710	+10% 2025–2035 (combined with statisticians); ~2,200 mathematician jobs
Statisticians	BLS, May 2025	$105,650	~31,300 jobs; master’s typical
Computer and information research scientists	BLS, May 2025	$140,300	+22% 2025–2035; ~38,600 jobs
AI research scientist at a frontier lab	Job postings and press reporting	Well above $200,000 total compensation	Estimate, not official data; small number of positions

The realistic read: this is a high-pay, low-headcount field. Roughly 2,200 people in the U.S. hold the job title “mathematician” according to BLS, and frontier labs employ a few dozen people in explicitly AI-plus-mathematics roles. The 22% growth in research scientist roles is where the volume is. Tell a kid this field is fascinating and competitive, not that it’s a safe bet.

One more caveat worth saying out loud to a teenager: the field is in flux. Alon’s reaction to AI solving Erdős problems was “Once AI started to solve them, there is no point anymore.” That’s a working mathematician saying a research program he cared about lost its meaning. Anyone entering this field is signing up for a decade where the job description changes more than once.

What a 10–15-year-old can do this year, for free

Join a math circle or start competition math

Competition problems are the closest thing to research that a 12-year-old can access: unfamiliar, no template, and you have to invent an approach. AMC 8 and AMC 10 are the standard on-ramps in the U.S. Our piece on math circles and olympiad training in the age of AI perfect scores covers how the training has changed since July 2026.

Learn Lean, seriously

Lean is a proof assistant: you write mathematics as code and the computer refuses to accept a wrong step. There is a free interactive Natural Number Game that teaches the basics in a weekend, and the mathlib library is open source on GitHub. A 14-year-old who contributes one small lemma to mathlib has a public artifact that a lab recruiter can read. Our explainer on how proof assistants turn math into code is the parent version.

Read the Erdős Problems database

Bloom’s database lists open problems with their status and history. Most are far beyond a teenager. That’s not the point. Reading twenty problem statements teaches what “well-specified open question” means, which is the first skill in the job.

Check an AI’s math, on purpose

Give a chatbot a competition problem you already know the answer to and grade its work. Then do it ten times with the same problem and count how often it’s right. Your kid will discover the 50% phenomenon themselves, which beats any lecture on AI reliability. Our guide on teaching kids what a 50% success rate means has the full exercise.

What not to do

Don’t push a kid into competition math if they hate contests. Research mathematics and timed competitions are correlated but not identical, and plenty of strong mathematicians were mediocre competitors. Reading proofs, writing proofs, and being stubborn about a single problem for a month matter more than speed under a clock.

What to Watch For Over the Next 3 Months

• Week 4: Your kid has written one Lean proof, however trivial, and can explain why a computer rejecting their step is useful rather than annoying.
• Month 2 red flags: They treat a chatbot’s math as authoritative. They want the answer rather than the argument. They quit a problem in under ten minutes. The third one is the most important: tolerance for being stuck is the trait this career selects for.
• Month 3 self-check: Ask them to explain the difference between “the AI found a proof” and “the proof has been verified.” If they can, they understand the actual job better than most headlines do.

Frequently Asked Questions

Is AI going to eliminate this career before my kid gets there?

It’s reshaping it rather than eliminating it, at least so far. Every AI mathematical result in 2026 needed humans to select the problem, check the output, formalize the steps, and attribute the ideas. The honest answer is that nobody knows what the role looks like in 2036, and that uncertainty applies to most technical careers right now.

Does my kid need to be a prodigy?

No, but they need to be genuinely stubborn. The distinguishing trait among research mathematicians is tolerance for sitting with an unsolved problem for weeks. Speed helps in competitions and matters less afterward. Our piece on when to accelerate a kid in math and when not to covers the difference.

What if they love math but hate programming?

Formalization in Lean is programming, so some coding is unavoidable in the AI-adjacent version of this career. But pure mathematics still exists as a career, and the university path is the same for both up through a bachelor’s degree. Let them find out at 19, not 13.

How competitive is it really?

Very. BLS counts about 2,200 mathematician positions in the U.S., and frontier-lab AI-mathematics teams are small. The wider field, computer and information research scientists at 38,600 jobs and growing 22%, is where most people with this training actually land, and it pays a median of $140,300.

Which undergraduate courses matter most?

Real analysis and abstract algebra, because they teach proof discipline. Then discrete mathematics, number theory, and combinatorics, because those are the areas where models currently perform best. Add one logic or formal-methods course, which is where Lean-style verification lives.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. OpenAI. (2026). “Model disproves discrete geometry conjecture.” May 20, 2026. https://openai.com/index/model-disproves-discrete-geometry-conjecture/
2. Alon, N., Bloom, T. F., Gowers, W. T., Litt, D., Sawin, W., Shankar, A., Tsimerman, J., Wang, V., & Wood, M. M. (2026). “Remarks on the disproof of the unit distance conjecture.” arXiv 2605.20695. https://arxiv.org/pdf/2605.20695
3. Quanta Magazine. (2026). “Why the legendary Erdős problems are falling to AI.” August 3, 2026. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/
4. Bureau of Labor Statistics. (2026). “Mathematicians and Statisticians.” Occupational Outlook Handbook, May 2025 data. https://www.bls.gov/ooh/math/mathematicians-and-statisticians.htm
5. Bureau of Labor Statistics. (2026). “Computer and Information Research Scientists.” Occupational Outlook Handbook, May 2025 data. https://www.bls.gov/ooh/computer-and-information-technology/computer-and-information-research-scientists.htm
6. Science News. (2026). “Mathematicians want guardrails on AI after it cracked an Erdős problem.” https://www.sciencenews.org/article/ai-guardrails-erdos-math-problem
7. Lean FRO. Lean theorem prover and mathlib documentation. https://lean-lang.org/
8. International Mathematical Olympiad. (2026). “IMO 2026, Shanghai, China.” https://www.imo-official.org/editions/2026/