Mathematicians

Three parts of your work where AI is already doing real lifting, and what stays yours.

• Chalkboard1876–1940

  Chalkboard, printed journals, and hand computation (research university era)

  The professional mathematician of the late nineteenth and early twentieth centuries worked with chalk, paper, and the physical library. Proof was transmitted through journals -- the American Journal of Mathematics (1878), Annals of Mathematics (1884), and Transactions of the AMS (1900) -- and through in-person colloquia. Computation was done by hand or by mechanical calculation devices (the Brunsviga pinwheel calculator, introduced in 1878, was the workhorse of numerical computation for decades). The introduction of the chalkboard as a pedagogical and research tool replaced the recitation-based lecture hall; by 1900 the blackboard seminar was the standard format for communicating and critiquing new mathematical ideas in the research university. Nothing about this era automated the core task: proving theorems required sustained individual reasoning, and the only tool was the mind.

• Wartime applied mathematics1940–1965

  Wartime applied mathematics and electromechanical computation (AMP, ENIAC era)

  World War II transformed the practice of American mathematics for a generation. The Applied Mathematics Panel (1942-1945), led by Warren Weaver under the National Defense Research Committee, coordinated the work of hundreds of mathematicians on ballistics tables, submarine search theory, operations research, and cryptanalysis. Mina Rees, Hunter College professor, served as Weaver's technical aide and shaped the postwar funding landscape. Simultaneously, the war produced the first electronic computers: ENIAC (1945) and its successors were applied first to weapons calculations, then to hydrodynamics, number theory, and statistical sampling (Monte Carlo methods, invented by Ulam and von Neumann at Los Alamos). For the mathematician, the electronic computer was not yet a tool that could replace mathematical reasoning -- ENIAC was programmed by physicists and engineers following a mathematician's analytical model -- but it was the first device that could perform the numerical computations that had previously occupied weeks of hand calculation.

  Effect on the work

  The wartime AMP and postwar NSF/NDEA funding surge approximately doubled the professional mathematics workforce between 1940 and 1960, with the largest growth in government and industrial applied mathematics positions. The pace of this growth (348% in the single decade 1950-1960) was unprecedented in the history of the profession.

• Sputnik-era federal investment1957–1975

  Sputnik-era federal investment: NSF, NDEA, and the Cold War mathematics boom

  On October 4, 1957, the Soviet Union launched Sputnik I and triggered the most consequential peacetime investment in mathematics education and research that the United States had ever made. The National Science Foundation's budget grew sharply; the National Defense Education Act (1958) authorized over $1 billion in education spending, including 1,500 annual graduate fellowships in science, mathematics, and engineering. Congressional testimony in 1957 specifically identified an "acute shortage of mathematicians" driven by the demand of the electronic computer industry and the expanding national security apparatus. Universities hired rapidly: the 1960 Census recorded a 348% increase in mathematician employment since 1950. NSA became the single largest employer of PhD mathematicians in the United States. The era produced a generation of mathematicians trained primarily for academic and government research -- an employment structure that would prove fragile when defense budgets tightened and academic hiring froze in the late 1970s.

  Effect on the work

  The NSF/NDEA investment produced the 348% employment surge documented in the 1960 Census. This was also the era when the profession became formally institutionalized: the AMS Employment Register began in 1953, matching PhD mathematicians to university and government positions; by the late 1950s it was a physical binder passed from conference to conference.

• Mainframe1965–1988

  Mainframe and minicomputer symbolic algebra (REDUCE, MACSYMA, early CAS systems)

  The first computer algebra systems -- FORMAC (1962), REDUCE (1968), MACSYMA (1969, developed at MIT Project MAC) -- gave mathematicians the ability to perform symbolic manipulation by machine for the first time. MACSYMA could integrate, differentiate, factor polynomials, and simplify algebraic expressions in ways that had previously required hours or days of hand calculation. For pure mathematicians, these tools were interesting but not yet transformative; for applied mathematicians working in theoretical physics or celestial mechanics (where calculations often ran to dozens of pages of algebra), they were genuinely time-saving. The mainframe access barrier meant that in practice only researchers at well-funded universities and national laboratories could use them regularly.

• MATLAB1984–1993

  MATLAB (1984) and Mathematica (1988) democratize computation

  PC-MATLAB debuted at the IEEE Conference on Decision and Control in December 1984. Cleve Moler had written the original MATLAB at the University of New Mexico to give students easy access to LINPACK and EISPACK numerical libraries without writing Fortran; Jack Little rewrote it in C and cofounded MathWorks in California in December 1984. Then, on June 23, 1988, Stephen Wolfram shipped Mathematica 1.0 -- a system that could symbolically integrate, factor, and simplify in ways that had previously required weeks of manual work, running on a personal workstation rather than a mainframe. "When Mathematica 1.0 was released," Wolfram later wrote, "it was notable that one could routinely do integrals symbolically by computer, and it wasn't long before the system could do integrals better than any human." For the mathematician, this was a genuine shift: routine symbolic manipulation moved out of the proof workflow and into a machine. The surviving human task was knowing which computation to run, interpreting the output in the context of a proof, and recognizing when a symbolic result masked a subtle domain restriction. The effect was liberation of cognitive bandwidth for higher-level proof strategy, not displacement of the mathematician.

  Effect on the work

  No direct employment displacement -- mathematicians were already a tiny, credentialed workforce where the bottleneck was proof strategy, not computation speed. The effect was an expansion of what one mathematician could explore in a week, raising the pace of mathematical experimentation without reducing headcount.

• Soviet emigre influx1991–2005

  Soviet emigre influx and academic job market compression (post-Cold War)

  The collapse of the Soviet Union in 1991 brought many Soviet emigre mathematicians to US universities over the following decade. Soviet mathematics had developed deeply in areas where American work was sparse (topology, partial differential equations, mathematical physics) while avoiding others, partly due to ideological pressures and the Iron Curtain's restriction on intellectual exchange. The influx was documented in a 2012 NBER study by George Borjas and Kirk Doran, who found that American mathematicians working in fields where the Soviet emigres specialized saw reduced lifetime publication productivity -- not because their skills declined, but because the emigres took positions and wrote the papers that the existing cohort would otherwise have written. The Soviet contribution filled mathematical gaps but did not grow the total pie of mathematician employment. Simultaneously, Cold War defense spending cuts reduced government mathematician positions at NSA, RAND, and the national laboratories, while the academic market stagnated.

  Effect on the work

  The NBER study found statistically significant displacement effects on American mathematicians in overlapping fields: increased mobility to lower-quality institutions, reduced likelihood of writing "home run" papers. No net growth in mathematician employment accompanied the emigre influx -- the supply expanded but demand did not follow proportionally.

• arXiv2005–2022

  arXiv, Lean 4, and the formalization movement (proof assistants reach practical use)

  The arXiv preprint server (founded 1991 in physics, extended to mathematics in 1992) transformed how mathematical results circulated: the lag between completion and community availability shrank from months to days. By the mid-2000s arXiv was the default dissemination channel for new results in many subfields, and AI-assisted literature tools (Semantic Scholar from 2015, Elicit from 2021) built on the corpus to enable large-scale literature survey. Concurrently, the Lean theorem prover (version 4 released 2021) and its community-built Mathlib library reached practical usability: mathematicians began formalizing published proofs in Lean not as a curiosity but as a quality-control measure. Peter Scholze's "liquid tensor experiment" (2021) -- a community Lean formalization of a key step in his condensed mathematics program, completed in six months -- demonstrated that machine-checkable proofs of deep research-level mathematics were achievable. Lean 4 Copilot (2024) added LLM-generated tactic suggestions to the proof assistant interface.

• AI proof assistants2022–present

  AI proof assistants and reasoning models (AlphaProof, OpenAI o3, Mathematica LLM)

  In July 2024, Google DeepMind announced that AlphaProof -- a reinforcement-learning system trained on formal Lean 4 proofs -- had solved four problems from the 2024 International Mathematical Olympiad at a combined score equivalent to a silver medal, including the hardest problem (Problem 6, a number theory question). AlphaGeometry 2 solved the geometry problem. The proofs were machine-verified in Lean 4: every step was formally correct. In February 2025, the full research was published in Nature. OpenAI o3 achieved competitive scores on AIME and AMC benchmarks through extended chain-of-thought reasoning. Wolfram Mathematica 14 (2024) integrated an LLM Notebook Assistant that generates Wolfram Language code on natural-language prompts. Eloundou et al. (2023) rated Mathematicians at 100% LLM task exposure -- the highest of any occupation in their study -- because every task in the O*NET profile can in principle be time-reduced by an LLM. The key qualification is "in principle on a structured task": IMO and AIME problems are highly structured with known answer forms, while research mathematics involves identifying the right question in an unbounded solution space -- a task where AI tools as of 2025 require human direction at every inflection point. Fields medalist Terrence Tao noted in 2025 that AI tools are "genuinely useful for routine verification and literature search, but the creative direction of research remains entirely human."

  Effect on the work

  No statistically detectable employment effect yet -- the workforce is too small (2,220 OEWS jobs in 2024) and the AI tools are too recent for a clean signal. The structural bet is that AI proof-search tools will redirect mathematician effort from routine proof-checking and literature survey toward conjecture formation and research-question selection -- the same pattern that Mathematica and MATLAB produced in 1988-1995: augmentation, not displacement.