Figure 1: Claude queries and workforce shares
Note: Fraction of US workforce and Claude
queries associated with major SOC occupation groups
This horizontal bar chart displays 24 major occupational groups with two bars each: red bars showing the fraction of US workforce and blue bars showing the fraction of LLM queries. Key findings include: - Computer and Mathematical: Shows the largest disparity with approximately 3% of the workforce but nearly 40% of LLM queries. - Management: Represents about 5% of workforce and approximately 14% of queries - Office and Administrative Support: About 12% of workforce with roughly 9% of queries - Educational Instruction and Library: Approximately 6% of workforce with about 10% of queries - Sales and Related, Transportation and Material Moving, Business and Financial Operations, Life Physical and Social Science, Healthcare Practitioners and Technical, Food Preparation and Serving Related, and Production: Range from 2-10% for both workforce and queries - Remaining occupational groups (Architecture and Engineering, Arts/Design/Entertainment/Sports/Media, Building and Grounds Cleaning, Community and Social Service, Construction and Extraction, Farming/Fishing/Forestry, Healthcare Support, Installation/Maintenance/Repair, Legal, Personal Care and Service, and Protective Service) each represent less than 5% of both workforce and queries The data demonstrates that LLM usage is disproportionately concentrated in computer and mathematical occupations compared to their representation in the overall workforce.
Figure 2: CPS and CES indexed levels
Note: Indexed levels of total CPS employment
(NSA) and CES employment (SA). Vertical line indicates ChatGPT release
date
This line chart displays two employment indices tracked from January 2014 through 2026: • Blue line: Total CPS employment (Current Population Survey, Not Seasonally Adjusted) • Red line: CES employment (Current Employment Statistics, Seasonally Adjusted) Key trends and patterns: 2014-2020: Both indices show steady growth from an index value of approximately 0.99 in January 2014 to a peak of approximately 1.07 in early 2020. The two lines track very closely together during this period with minor fluctuations. 2020 (COVID-19 impact): Both indices experience a sharp drop around March-April 2020, falling to approximately 0.90-0.92, representing the pandemic-related employment decline. 2020-2022: Both indices show strong recovery, climbing back to pre-pandemic levels by mid-2021 and continuing upward to approximately 1.06-1.07 by late 2022. 2022 on: The series move closer together The chart suggests that while both employment measures show overall growth, CES employment has grown slightly more steadily and to a higher level than CPS employment recently
Figure 3: Coder employment
Note: Employment level of coder occupations
based on O*NET programming skill. Dashed line shows ChatGPT release date
(November 2022.) Coder intensive industries are those where over 10
percent of workers are coders, non-coder intensive industries are the
remainder.
This figure consists of three line charts displayed side by side, each showing employment levels for coder occupations from January 2014 through 2026. All three charts include a solid blue line representing actual coder employment, a dashed gray line showing the 2016-2019 trend projection, and a vertical dashed line marking the ChatGPT release date in November 2022. Panel 1: All Industries (left chart) The y-axis ranges from 3,000 to 7,000 thousands of jobs. Coder employment shows steady growth from approximately 3,800 thousand jobs in early 2014 to around 6,000 thousand jobs by late 2022. The growth closely follows the 2016-2019 trend line throughout this period. After the ChatGPT release date in November 2022, employment continues to grow but more slowly. Panel 2: Coder Intensive Industries (middle chart) The y-axis ranges from 1,500 to 3,000 thousands of jobs. This panel shows employment in industries where over 10 percent of workers are coders. Starting at approximately 1,600 thousand jobs in 2014, employment grows to around 2,700 thousand jobs by late 2022, generally following the 2016-2019 trend. After the ChatGPT release, employment shows considerable fluctuation, ranging between approximately 2,500 and 2,700 thousand jobs through 2026, remaining below the upward trajectory of the 2016-2019 trend line. Panel 3: Non-Coder Intensive Industries (right chart) The y-axis ranges from 2,000 to 3,500 thousands of jobs. This panel displays employment in industries where 10 percent or fewer workers are coders. Employment grows from approximately 2,200 thousand jobs in 2014 to around 3,200 thousand jobs by late 2022, closely tracking the 2016-2019 trend. After ChatGPT's release, employment shows initially strong growth but flattens quickly, falling below the trend at the very end of the sample. Key Findings: The data reveals different post-ChatGPT employment patterns across industry types. While overall coder employment continued to grow after November 2022, it slowed noticeably.
Figure 4: Coder employment with counterfactual
Note: Employment level of coder occupations
based on O*NET programming skill. Dashed line shows ChatGPT release date
(November 2022.) Coder intensive industries are those where over 10
percent of workers are coders, non-coder intensive industries are the
remainder.
This figure consists of three line charts displayed side by side, each showing employment levels for coder occupations from January 2014 through 2026. All three charts include a solid blue line representing actual coder employment, a dashed red line showing counterfactual employment (what employment would have been without ChatGPT), and a vertical dashed line marking the ChatGPT release date in November 2022. Panel 1: All Industries (left chart) The y-axis ranges from 3,000 to 7,000 thousands of jobs. From 2014 through late 2022, actual coder employment (blue line) and the counterfactual projection (red dashed line) track very closely together, both growing from approximately 3,800 thousand jobs in early 2014 to around 6,000 thousand jobs by November 2022. After the ChatGPT release date, the two lines begin to diverge. The counterfactual line rises sharply to approximately 7,000 thousand jobs by 2026, while actual employment shows more modest growth with increased volatility, reaching approximately 6,000 thousand jobs by 2026. This represents a gap of roughly 1,000 thousand jobs between actual and counterfactual employment by the end of the period. Panel 2: Coder Intensive Industries (middle chart) The y-axis ranges from 1,500 to 3,500 thousands of jobs. In industries where over 10 percent of workers are coders, both lines track together from 2014 through late 2022, growing from approximately 1,600 thousand jobs to around 2,700 thousand jobs. After ChatGPT's release, a pronounced divergence occurs. The counterfactual line (red dashed) climbs steeply to approximately 3,600 thousand jobs by 2026, while actual employment (blue solid) stagnates and fluctuates between 2,500 and 2,700 thousand jobs. This creates the largest gap among the three panels—approximately 1,000 thousand jobs between actual and counterfactual by 2026. Panel 3: Non-Coder Intensive Industries (right chart) The y-axis ranges from 2,000 to 3,500 thousands of jobs. For industries where 10 percent or fewer workers are coders, the actual and counterfactual lines track closely throughout the entire time period. Both grow from approximately 2,200 thousand jobs in 2014 to around 3,300-3,400 thousand jobs by 2026. Right around the ChatGPT release the Coders line shoots higher, but is flatter immediately after. Key Findings: The counterfactual analysis suggests actual employment fell significantly below what would have been expected without ChatGPT, though the picture is more ambiguous for non coder intensive industries.
Figure 5: Normalized CES Employment
Note: Dashed line marks November 2022
This line chart displays four normalized employment indices (indexed to 1 in 2018) tracked from January 2014 through 2026. The chart includes a vertical dashed line marking November 2022, the ChatGPT release date. The four series are: • Solid blue line: Total private employment • Red dashed line: Information sector employment • Green dotted line: Coder-intensive aggregate employment • Orange dashed line: Software sector employment (NAICS 5132) Pre-Pandemic Period (2014-2020): All four employment indices show generally upward trends from 2014 through early 2020, though with different trajectories: • Total private employment (blue solid) starts at approximately 0.92 in 2014 and grows steadily to approximately 1.0 by early 2020 • Information sector (red dashed) begins at approximately 0.94 in 2014 and rises gradually to approximately 0.99 by early 2020, tracking closely with total private employment • Coder-intensive aggregate (green dotted) starts at approximately 0.85 in 2014 and shows steady growth to approximately 0.98 by early 2020 • Software sector (orange dashed) shows the most dramatic pre-pandemic growth, starting at approximately 0.67 in 2014 and rising steeply to approximately 1.0 by early 2020 Pandemic Period (2020-2021): Around March-April 2020, most sectors show a sharp employment decline: • Total private and information sector both drop to approximately 0.85-0.88 • Coder-intensive aggregate experiences a smaller decline to approximately 0.90 • Software sector shows a brief dip to approximately 0.95 but recovers quickly All sectors demonstrate recovery through 2021, returning to or exceeding pre-pandemic levels by late 2021. Software continues its pre-pandemic trend almost uninterrupted, coder-intensive is similar but more muted Post-ChatGPT Period (November 2022-2026): After the November 2022 ChatGPT release, marked by the vertical dashed line, the four sectors show dramatically divergent patterns: • Software sector (orange dashed) ceases its rapid growth, dipping before edging up in 2025 • Coder-intensive aggregate (green dotted) is similar • Information sector (red dashed) peaks in late 2022, then declines steadily to approximately 0.98 by 2026 • Total private employment (blue solid) shows steady modest growth from approximately 1.02 to approximately 1.05 by 2026 Key Findings: The data reveals striking heterogeneity in post-ChatGPT employment trends across sectors. The software sector experienced unprecedented growth prior and flattened sharply after.
Figure 6: Industry growth scatterplot
Note: Annualized CES growth of 3 digit NAICS
industries and the coding-intensive industry group.
This scatterplot displays employment growth patterns across industries, comparing pre-pandemic growth (2016-2019) on the x-axis to more recent growth (2020-2025m7) on the y-axis. Both axes show annualized average growth rates ranging from -10 to 10. The chart includes four elements: • Small blue dots representing individual 3-digit NAICS industries • A red dot representing the coding-intensive sector • A solid black line showing a 45-degree reference line (representing equal growth in both periods) • A dashed line showing the linear prediction/regression line Overall Pattern: The majority of industries cluster in the range of -5 to +5 on both axes, with the highest concentration between 0 and 5. The data shows a positive relationship between pre-pandemic and post-pandemic growth rates—industries that grew faster before the pandemic tended to grow faster after it as well. The best fit line is somewhat more shallow than the 45 degree line, suggesting mean reversion in industry growth rates. The coding-intensive dot sits almost exactly on the best fit line, only slightly below.
Figure 7: Negative occupation-specific shocks
Source: ACS, Deming, 2017, authors’ calculations
This figure contains four line charts showing employment levels from 1980 to 2020 for occupations affected by automation. Each panel compares observed employment (solid blue line) to counterfactual employment based on industry growth (red dashed line). Bank Tellers (upper left): Observed employment remains relatively flat around 400-450 thousand jobs throughout the period, while counterfactual employment rises from approximately 450 thousand in 1980 to over 800 thousand by 2020. The growing gap indicates significant automation-related employment suppression. Bookkeepers and Accounting and Auditing Clerks (upper right): Observed employment shows modest growth from approximately 1,500 thousand in 1980 to a peak of about 1,700 thousand in 2000, then declines to approximately 1,200 thousand by 2020. Counterfactual employment grows steadily to approximately 2,700 thousand by 2020. The divergence accelerates after 2000, suggesting substantial displacement. Telephone Operators (lower left): Observed employment declines sharply from approximately 250 thousand in 1980 to nearly zero by 2020, representing near-complete occupation elimination. Counterfactual employment rises to approximately 450 thousand by 2020, showing what employment would have been without automation. Data Entry Keyers (lower right): Observed employment rises from approximately 400 thousand in 1980 to a peak of about 550 thousand in 1990, then declines to approximately 350 thousand by 2020. Counterfactual employment grows steadily to approximately 620 thousand by 2020, with divergence beginning around 2000. Key Finding: All four occupations show substantial gaps between observed and counterfactual employment, demonstrating how technological change can reduce employment even in growing industries. The patterns vary—from gradual suppression (bank tellers) to near elimination (telephone operators)—but all indicate significant negative occupation-specific shocks.
Figure 8: Examples of industry shocks
Source: ACS, Deming, 2017, authors’ calculations
This figure contains four line charts showing employment from 1980 to 2020 for occupations affected primarily by industry-level changes rather than occupation-specific automation. Each panel compares observed employment (solid blue line) to counterfactual employment based on industry growth (red dashed line). Unlike Figure 7, these occupations show observed and counterfactual lines tracking closely together. Textile Sewing Machine Operators (upper left): Both lines decline together from approximately 700 thousand jobs in 1980 to approximately 280 thousand by 2020, reflecting the broader decline of the textile manufacturing industry in the United States. Petroleum, Mining, and Geological Engineers (upper right): Both lines decline from approximately 42 thousand in 1980 to a low of approximately 30 thousand around 2000, then rise sharply together to approximately 55 thousand by 2020, tracking the boom-and-bust cycles of the energy industry. Pharmacists (lower left): Both lines show steady parallel growth from approximately 150 thousand in 1980 to approximately 300 thousand by 2020, reflecting healthcare industry expansion and aging demographics. Guards, Watchmen, Doorkeepers (lower right): Both lines grow together from approximately 450 thousand in 1980 to approximately 1,000 thousand by 2020, tracking the expansion of the security services industry. Key Finding: In contrast to Figure 7's negative occupation-specific shocks, these occupations experienced industry shocks where employment changes reflected broader industry trends rather than occupation-specific displacement. The close alignment between observed and counterfactual employment indicates these occupations maintained their expected share of employment within their respective industries.
Figure 9: Positive occupation-specific shocks
Source: ACS, Deming, 2017, authors’ calculations
This figure contains two line charts showing employment from 1980 to 2020 for occupations that experienced positive shocks—where observed employment (solid blue line) substantially exceeded counterfactual employment based on industry growth (red dashed line). Legal Assistants, Paralegals, Legal Support (left panel): Observed employment grows dramatically from approximately 150 thousand in 1980 to approximately 600 thousand by 2010, then plateaus through 2020. Counterfactual employment shows much more modest growth from approximately 175 thousand to only 280 thousand by 2020. The widening gap indicates that these occupations expanded far beyond what industry growth alone would predict, likely due to increased delegation of legal work from lawyers to lower-cost paralegals. Programmers of Numerically Controlled Machine Tools (right panel): Observed employment initially declines from approximately 15 thousand in 1980 to nearly zero around 1990, then rises sharply starting in 2000 to reach approximately 100 thousand by 2020. Counterfactual employment remains relatively flat around 15-20 thousand throughout the period. This pattern reflects the occupation's transformation with the widespread adoption of computer numerical control (CNC) technology in manufacturing, creating substantial new demand for these specialized programmers. Key Finding: These positive occupation-specific shocks demonstrate how technological change and organizational restructuring can create substantial employment growth within specific occupations, even when their broader industries experience only modest expansion.
Figure A1: Bai-Perron Breakpoints
Note: Employment level of coder occupations
based on O*NET programming skill. Vertical lines show structural break
dates estimated via Bai-Perron tests. The dashed line shows ChatGPT
release date (November 2022).
This line chart displays coder employment across all industries from January 2014 through 2026, with multiple vertical lines marking statistically identified structural breaks. The chart shows two series: actual coder employment (solid blue line) and counterfactual employment (red dashed line). Structural Break Dates (Solid Vertical Lines): The Bai-Perron statistical tests identify three structural break points shown as solid black vertical lines at approximately: • Late 2017 • Early 2020 • Early 2022 ChatGPT Release Date (Dashed Vertical Line): A dashed vertical line marks November 2022, the ChatGPT release date.