Mathematician · From Babylonian scribes to Fields medalists and AI-assisted proof: the profession that turns hard questions into permanent certainty, one theorem at a time.
Mathematics is the rare profession where the machines are coming for the prestige tasks first. Calculation fell decades ago; by 2024–25, AI systems reached silver- then gold-medal standard on International Mathematical Olympiad problems, and proof assistants like Lean verified arguments at the frontier of research — including, in 2022, a theorem its own author feared checking by hand. No serious mathematician now bets that routine lemma-proving stays human for long.
Yet the profession's reaction has been closer to excitement than dread, because its scarcest work sits upstream of everything the machines do: deciding which questions are worth years of a life, inventing the definitions that make a vague intuition provable, and judging which of infinitely many true statements mean something. Terence Tao, among others, describes the near future as mathematicians directing teams of tireless, semi-reliable machine collaborators.
A real share of the working day — literature search, verification, computation, well-posed problem solving — is automating now, and AI already proves competition-level results. But the tasks that define the career (choosing problems, forming concepts, judging significance, teaching humans) sit furthest from what current systems do. The likeliest outcome is fewer hours proving routine lemmas and more spent deciding what to prove: transformation of the job, not removal of the profession.
Scored from the tasks, not the job title. Lower is safer.
Jobs AI cannot take →A theorem prover can attack a stated conjecture; it cannot decide that a question is worth ten years, or notice that a different question is the important one. Problem choice is the profession's compounding advantage.
The field's deepest moves — Grothendieck's schemes, Noether's rings, Cantor's infinities — created the objects later work computes with. Concept formation, not deduction, is where mathematics actually grows, and no system yet does it.
Infinitely many true statements exist; almost all are worthless. Judging which results illuminate, which proofs explain rather than merely verify, and which directions are fertile remains a human consensus built over careers.
When mathematics certifies a cryptosystem, a drug trial's statistics or an aircraft control law, a named human must stand behind what the theorem actually says about the world. Machine-checked is not the same as correctly modeled.
The apprenticeship chain — advisor to student, hand to hand since the medieval universities — transmits taste and standards, not just content, and remains the way the profession reproduces itself.
Already gone: computer algebra systems have handled the field's calculation, integration and equation-solving since the 1980s, and no working mathematician regrets it.
Language models now retrieve relevant lemmas, known techniques and prior art across millions of papers faster than any human reading — eroding the advantage of encyclopedic memory that defined scholars like Erdős.
Checking correctness line by line — the referee's most tedious duty — increasingly belongs to formal systems like Lean and Coq, which do it with a rigor no tired human referee matches.
Given a precisely stated competition-style problem, AI systems reached IMO gold-medal standard in 2025. Extending that to routine research lemmas — stated by a human, proved by a machine — is the frontier being crossed now.
Peter Scholze's Liquid Tensor Experiment (2021–22) saw the Lean community verify a theorem he considered his most important and least checkable, converting many skeptics; major new proofs increasingly ship with machine-checked companions, and journals are beginning to expect it.
DeepMind's 2021 Nature paper generated conjectures in knot theory and representation theory that mathematicians Geordie Williamson and colleagues then proved; AlphaProof followed in 2024. The emerging division of labor: machines propose and grind, humans frame and judge.
The classification of finite simple groups consumed tens of thousands of journal pages and over a hundred authors; Polymath projects since 2009 prove theorems by open online crowds. Average authorship keeps rising — the solitary-genius workflow is quietly retiring.
AI labs, trading firms and cryptography teams now bid directly for research mathematicians, hiring the exact talent academia trains but cannot pay. The profession's center of gravity is drifting from the mathematics department toward industry, as physics' did in the twentieth century.
The mathematics of deep learning — optimization, probability, geometry of high-dimensional spaces — is thin and openly acknowledged as such; research mathematicians are hired to build the theory the empirical results are still waiting for.
Trading firms run on stochastic calculus, statistics and optimization, and recruit directly from mathematics departments at several times academic pay — the largest single exit ramp for PhDs in the field.
The post-quantum standards NIST finalized in 2024 rest on lattice problems from pure number theory; the world's communications security is now, quite literally, applied research mathematics with a deployment deadline.
The same proof-assistant skills entering pure mathematics are paid industrial work at chipmakers and cloud providers — Intel has formally verified floating-point units since the 1994 Pentium division bug cost it $475 million, and AWS proves properties of its core protocols.
The honest near-term picture: the parts of the job that look like solving — well-posed, self-contained, checkable — will increasingly be done with or by machines, exactly as calculation was. The parts that look like asking — framing, defining, judging, teaching — show no sign of automating, and they are the parts practitioners always identified as the real work.
The economics point the same way. Demand for mathematically trained people is rising everywhere machines get smarter, because someone must specify what the machines should do and verify what they did; the US Bureau of Labor Statistics projects double-digit growth for mathematical occupations through 2033. The title on the badge may say researcher, quant or scientist, but the training is the mathematician's.
The open question is cultural, not economic: whether a field whose identity was built on unaided human proof will treat machine-assisted theorems as fully owned knowledge. The four color controversy of 1976 previewed the argument; the Lean era is settling it in practice, one formalized proof at a time. The profession that survived the loss of its own name — computer — to a machine will likely survive sharing its proofs with one.
Derives and tests the mathematical laws governing matter, energy, space and time, from a lone chalkboard to a 3,000-author particle-collider paper.
AI-resistant 65 🧬The scientist who studies life itself, from Linnaeus naming species by hand to editing genomes with CRISPR, still testing every idea against a living organism.
AI-resistant 64 🧪The scientist who makes and measures matter itself — from Tapputi's Babylonian perfume still to today's robot laboratories, still the one who decides what the spectrum means.
AI-resistant 70 📊Finds patterns and builds predictive models from data — a 2008 job title built on three centuries of counting, testing and visualizing evidence.
AI-resistant 38 🔭The scientist who measures the universe, from Babylonian clay tablets to space telescopes, still deciding which flicker in the data is a discovery.
AI-resistant 70