Coder employment growth slowed sharply following ChatGPT's arrival, pointing to an occupation-specific shock rather than a broad industry downturn; coder jobs continue to expand but at a much-reduced pace.

Main Finding

Coder employment in the U.S. (occupations defined as coding- or programming-intensive using O*NET) experienced a clear deceleration in growth beginning around the introduction of ChatGPT (November 2022). After accounting for industry-level shocks with a novel within-/between-industry counterfactual, annual coder employment growth is estimated to be roughly 3 percentage points lower post-ChatGPT than pre-2022. Coders continue to add jobs but at a much slower pace than before.

Key Points

• Motivation and exposure

  • Real-world usage data from Anthropic’s Economic Index (AEI) show computer & mathematical occupations account for over one-third of Claude queries while being only ~3.4% of the workforce; most such queries are programming-related. This motivates focusing on coders as a highly LLM-exposed group.
  • Indirect exposure measures (e.g., Eloundou et al. 2024) and AEI broadly agree that coding occupations are among the most exposed, though the measures do not fully agree with each other.

• Definition of “coders”

  • Authors map ONET programming-skill ratings into CPS occupations to define coding-intensive occupations. After imputing missing ONET entries (and specially treating software developers), coding-intensive occupations cover about 3.7% of U.S. employment using a programming-skill threshold.

• Main empirical result

  • An event-study-style analysis around November 2022 shows a sharp slowdown in coder employment growth.
  • Controlling for industry-level employment dynamics (see Data & Methods), the slowdown remains large and statistically robust: roughly a 3 percentage-point reduction in annual growth attributable to an occupation-specific shock coincident with ChatGPT’s introduction.
  • Overall coder employment continues to rise, but more slowly than before 2022.

• Industry concentration

  • Coders are dispersed across industries, but ~40% of coders work in “Computer systems design and related services” (NAICS 5415), i.e., contract/software development firms — not primarily at large in-house tech employers.

• Measurement comparisons

  • The paper compares exposure measures (AEI vs. task-based assessments like Eloundou et al.) and documents notable disagreements in detail, but consistent ranking: coders rank highly exposed across measures.

• Caveats highlighted by authors

  • The 3% lower growth estimate is not a definitive causal estimate of LLM effects on coder employment. Possible confounders include aggregate labor demand shifts, changing task mixes within occupations, reallocation to other occupations, and the prior availability of coding-assistant tools (e.g., Copilot, GPT-3.5).
  • AEI excludes API/business license activity, which may under- or over-represent coder usage depending on user mix.

Data & Methods

• Data sources

  • Anthropic Economic Index (AEI) — task-level composition of user queries to Claude (Feb 2025 release used for motivation).
  • O*NET — occupational skill/task measures; programming-skill importance is the main metric.
  • CPS (Current Population Survey) monthly employment microdata (IPUMS occ2010 occupation codes).
  • Industry mapping uses longitudinally-consistent industry codes constructed from 2012/2017/2022 Census industry versions (authors avoid older ind1990 codes).

• Defining coders

  • Map O*NET programming-skill scores to CPS occupations; impute missing scores via embeddings + random forest; special-casing software developers based on additional evidence that they are coding-intensive.
  • Choose programming-skill threshold (≈2.76) to identify coding-intensive occupations, capturing ~3.7% of employment.

• Event-study and controls

  • Primary empirical approach: compare pre- vs post-Nov 2022 trends in monthly CPS coder employment (an event-study framing).
  • To separate occupation-specific shocks from industry-level shocks, construct a counterfactual (between-industry) series: the weighted sum of industry employment growth rates where weights are the pre-event distribution of coders across industries. Intuition: industry-level shocks scale employment homothetically within industries; occupation-specific shocks change within-industry composition.
  • Include this industry-level counterfactual as a control (analogous in form to a Bartik construct, but used here as a control rather than an instrument).
  • Validate the industry-control by applying it to historical examples of occupations known to have experienced either industry-wide or occupation-specific shocks; authors report that the control behaves as expected.

• Robustness and validation

  • Show results with and without the industry-level control; the occupation-specific slowdown is more evident after controlling for industry dynamics.
  • Compare exposure measures (AEI vs task-judgment measures) and report agreement/disagreement statistics.

Implications for AI Economics

• Evidence that LLMs are already producing measurable labor-market effects

  • The paper provides early empirical evidence consistent with a meaningful occupation-specific labor demand shock for a highly LLM-exposed occupation (coders) after widespread attention to ChatGPT.
  • This suggests that LLM diffusion can show up in occupation-level employment aggregates within a short time window.

• Mechanisms to consider

  • Complementarity vs substitution: coding assistants likely raise coder marginal productivity (complement) — if demand for coding services is inelastic, this can reduce employment; if demand is elastic, productivity gains could expand market size and increase employment over the longer run.
  • Short-run vs long-run dynamics: initial occupation-specific shocks may dominate short-run adjustments; longer-run effects depend on product-creation, new tasks/occupations, firm reorganization, and general equilibrium responses.

• Policy and research priorities

  • Need for firm-level data: firm controls and direct measures of AI adoption would help identify causal channels (substitution vs expanded demand, reallocation across firms).
  • Monitor industry concentration and displaced workers: high concentration of coders in contracting/IT services (NAICS 5415) implies sectoral vulnerability and specific labor-market adjustment needs (retraining, matching).
  • Improve exposure measurement: reconcile AEI-like real-usage metrics (including API/business usage) with task-based exposure measures; quantify how measurement differences affect empirical conclusions.
  • Study other margins: wages, hiring intensity, age composition (complementary to Brynjolfsson et al. 2025), occupational task mix within and across occupations, and longer-run creation of new coding-intensive products.

• Interpretation cautions

  • The estimated ~3 percentage-point slowdown should not be read as a definitive causal impact of LLMs on coder employment without further firm-level and causal identification work.
  • Early-stage evidence: effects observed so far could evolve substantially as LLM capabilities, adoption, and complementary organizational changes progress.

Summary takeaways: coders — a clearly highly LLM-exposed occupational group — have already shown a marked slowdown in employment growth coincident with the public emergence of ChatGPT. The paper’s within-/between-industry control strengthens the case that a component of that slowdown is occupation-specific (plausibly LLM-related), but important measurement and identification limitations remain.