This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.
Back to papersSummary Main Finding Notes Stanford Digital Economy Lab; 13% entry-level employment decline in AI-exposed occupations
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: This is a conceptual/agenda-setting paper, not an empirical study. Defines Transformative AI as AI enabling 3-5x increase in TFP growth. Proposes future methodologies including agent-based simulations, scenario planning, and TAI dashboards. No original data analysis conducted. Related to earlier working paper note mentioning Brynjolfsson 2025 on employment effects, but this is a separate conceptual piece.
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
[Claude classification]: Six stylized facts documented: (1) Employment decline for early-career workers in AI-exposed occupations; (2) Overall employment growth continues but stagnant for young workers since late 2022; (3) Declines concentrated in automative AI applications, not augmentative; (4) Results robust to firm-time fixed effects; (5) Adjustments visible in employment more than compensation; (6) Robust to excluding tech occupations and remote-work occupations. Uses Poisson event study regression with firm-time and firm-quintile fixed effects. Main outcome is monthly headcount. Also examines annual base salary (deflated). Paper argues codified knowledge (book learning) may be more replaceable by AI than tacit knowledge gained through experience, explaining age gradient. Sample restrictions: balanced panel of firms present Jan 2021-July 2025, excludes part-time workers, ages 18-70, requires job title observed (~70% of ADP sample). Authors note ADP sample not fully representative of US economy (overrepresents Northeast, manufacturing/services, faster-growing firms per Cajner et al. 2018).
Canaries in the Coal Mine: Six Facts about the Recent Employment Effects of AI
Brynjolfsson, Chandar, Chen
2025Stanford Digital Economy Lab
Observational labor marketCausal
LLM / Generative AIJunior / entry-levelSoftware / codingCustomer serviceAugmentation vs. substitution
Brynjolfsson, Chandar, and Chen use monthly payroll data from ADP (2018-2025) to examine employment changes in AI-exposed occupations by age group, finding that early-career workers (ages 22-25) in high-exposure occupations experienced a 13% relative employment decline after controlling for firm-time effects, while older workers and those in less-exposed occupations maintained stable or growing employment.
Early-career workers (ages 22-25) in the most AI-exposed occupations experienced a 13% relative employment decline (12 log points) from late 2022 to July 2025 after controlling for firm-time effects, while employment for workers aged 35-49 in the same occupations grew by over 9%, and employment declines were concentrated in occupations where AI is more likely to automate rather than augment human labor.
Primary Datasets
ADP payroll data
Secondary Datasets
None
- Key Methods
- Observational analysis of administrative payroll data with event study regression controlling for firm-time and firm-quintile fixed effects; merges occupational AI exposure measures (Eloundou et al. 2024 GPT-4 scores and Anthropic Economic Index automation/augmentation classifications) to individual-level employment records; compares employment trends across age groups and AI exposure quintiles.
- Sample Period
- 2018-2024
- Geographic Coverage
- US
- Sample Size
- 3.5-5 million workers per month (main balanced panel); covers over 25 million workers total across ADP clients; 250,000-350,000 workers aged 22-25 per month
- Level of Analysis
- Individual, Occupation
- Occupation Classification
- SOC
- Industry Classification
- None