Economists

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

• Paper1776–1890

  Paper, correspondence, and printed statistical tables (pre-institutional era)

  The political economist of the 18th and 19th centuries worked entirely with pen, paper, printed data tables, and correspondence networks. Quantitative work existed, most notably in the British census tradition and in the work of American statistician Francis Amasa Walker, but it was slow, largely hand-tabulated, and accessible only to scholars with access to major libraries. The economist's main tool was argument: deductive reasoning from first principles, illustrated with whatever statistical abstracts were available from government publications. The founding of the AEA in 1885 and the proliferation of university economics departments did not change the tools; it changed the institutional setting in which those tools were applied.

• Mechanical calculators1890–1950

  Mechanical calculators, adding machines, and punch-card tabulation (Hollerith era)

  Herman Hollerith's punch-card tabulation system, used for the 1890 US Census, was the first mechanical data-processing technology that reached practicing economists and statisticians. By the 1920s and 1930s, Marchant, Friden, and Monroe mechanical calculators had become standard equipment in university economics departments and federal agencies. The New Deal era brought large-scale statistical work into government for the first time, with national income accounts being constructed by hand tabulation in the Department of Commerce (Simon Kuznets' work from 1934 onward). The mechanical calculator era made it possible, with sufficient labor, to run regressions and tabulate cross-sections that would have been prohibitive by hand, but the computation time for a single coefficient estimate could run to days.

  Effect on the work

  Punch-card tabulation machinery required a class of "computing girls" (largely women with mathematical training employed as human computers) to operate the equipment. The economist who designed the analysis was separate from the workers who ran the calculations, a division of labor that persisted until electronic computers arrived in the 1950s.

• Mainframe computers1950–1975

  Mainframe computers and early econometrics packages (IBM 704, FORTRAN, Cowles Commission era)

  The arrival of mainframe computers at major universities and federal agencies in the 1950s transformed what an economist could compute. The Cowles Commission at the University of Chicago (and later Yale) had been developing simultaneous-equation econometric methods since the 1940s; by the mid-1950s those methods could be estimated on machines rather than by hand. IBM's mainframe, introduced in 1954, was the first widely available machine with floating-point hardware; FORTRAN, released by IBM in 1957, gave economists a practical programming language. By the mid-1960s mainframe-based regression packages were available at most research universities. The speed difference from mechanical calculators was roughly a thousand-fold, which enabled regression models with dozens of variables and opened the door to Monte Carlo simulation. This era separated the "theorist who could calculate" from the "theorist who could compute at scale" and created the first genuine specialization in econometrics as a distinct subfield.

  Effect on the work

  Mainframe access was expensive and mediated through university computing centers, so it remained a shared, scheduled resource rather than a personal tool. Junior economists and graduate students still spent significant time on laborious hand calculations for coursework; the mainframe handled the largest research tasks. The era eliminated the class of "human computers" that punch-card tabulation had required.

• Personal computers1975–1995

  Personal computers, SAS, RATS, and TSP (desktop econometrics era)

  The personal computer transformed economics from a field where computation required institutional access to one where an economist's individual workstation was the primary research tool. SAS (1976), RATS (Regression Analysis of Time Series, 1983), and TSP (Time Series Processor) became standard desktop packages for applied economists in the 1980s. Stata was first released in 1985 and by the early 1990s was displacing SAS for applied microeconomic research. The desktop era fundamentally changed the scope of empirical economics: a single researcher could now run hundreds of regressions in an afternoon, test robustness across samples, and iterate on specifications in ways that mainframe-queue-based computing had made prohibitive. This is the era that produced the "credibility revolution" in empirical economics, in which the focus shifted from large structural models to quasi-experimental identification strategies, partly because computing had made it feasible to explore datasets exhaustively enough to find natural experiments.

  Effect on the work

  The desktop econometrics era substantially reduced the proportion of the economist's time spent on computation logistics (scheduling mainframe access, submitting batch jobs, waiting for printouts) and increased the proportion spent on specification design and interpretation. It did not reduce demand for economists; it increased it, by expanding the scope of questions empirical economics could address.

• Internet data access1995–2015

  Internet data access, WRDS, FRED, administrative microdata, and R (big-data micro era)

  The internet made large economic datasets publicly accessible at negligible marginal cost for the first time. The Federal Reserve Bank of St. Louis launched FRED (Federal Reserve Economic Data) in 1991 with web access from 1994; Wharton Research Data Services (WRDS) gave university researchers online access to Compustat, CRSP, and other financial databases from the mid-1990s onward. The US Census Bureau expanded remote access to restricted microdata through its Federal Statistical Research Data Center network (first RDC opened 1994). R was released as open-source in 1995, eventually displacing SAS and RATS for much academic work. The combination of cheap data access and open-source statistical software enabled a generation of empirical economists to work with administrative datasets containing millions of observations, tax records, and matched employer-employee files that had been inaccessible to the previous generation. Geographic information systems, satellite data, and digitized historical records expanded further what "data" an economist could work with.

  Effect on the work

  The administrative-microdata era created strong demand for economists who combined statistical skill with knowledge of specific data sources (the LEHD, SIPP, NLSY, PSID, etc.) and who could navigate the institutional processes for data access. It did not threaten economist employment; it expanded the range of researchable questions and, consequently, the demand for economists to answer them.

• Machine learning libraries2015–2023

  Machine learning libraries, cloud compute, and Python for economics (causal ML era)

  The mid-2010s brought machine-learning methods into the economist's toolkit, initially as predictors for treatment assignment (lasso for variable selection in IV research) and later as first-stage tools in double machine learning (DML) and causal forest methods. Python, with scikit-learn (2011) and later PyTorch and TensorFlow, became a second primary language alongside R and Stata for economists working on high-dimensional data problems. Cloud computing (AWS, Google Cloud) made access to datasets too large for a local machine routine. The "causal machine learning" literature, exemplified by Chernozhukov et al. (2018) on double/debiased ML and Athey-Imbens work on causal forests, combined the credibility revolution's identification discipline with the predictive power of ML. This era expanded what applied economists could do without threatening the core causal-identification judgment that distinguishes a trained economist from a data scientist.

• Large language models2023–present

  Large language models and AI-assisted research tools (Stata 19 AI, GitHub Copilot, NotebookLM, Bloomberg AI Assist)

  The 2023-2026 AI wave has reshaped the junior-economist task stack more than any tool since the personal computer. GitHub Copilot generates R, Python, Stata, and Julia code from plain-English specifications; Jupyter AI provides in-notebook LLM assistance; Stata 19 (2025) offers AI-assisted do-file generation; NotebookLM and AlphaSense accelerate literature triage and synthesis. Brookings (2023) cites a "10x leverage factor" for research-active economists who adopt AI for literature triage and data assembly. The tasks that have remained robustly human are causal identification design, formal theory development, and policy-mandate testimony, all of which require institutional knowledge and accountability that no LLM can provide. The IMF (2024) identifies a growing demand for economists who can validate and contextualize AI-generated economic analysis in policy settings, suggesting that AI tools may increase demand for senior economists even as they reduce the routine coding burden at junior levels.

  Effect on the work

  BLS projects employment growth for economists through 2034, broadly flat headcount with hundreds of openings per year driven primarily by replacement demand. Net effect: stable or slowly growing headcount, with strong upward pressure on wages for AI-fluent practitioners.