Elicit: AI's Impact on Employment (public)
AI's Impact on Employment (public)
Research reportView only
Create alertChat
September 12, 2025
What is the estimated impact of AI on employment?
AI's impact on employment is mixed, with studies showing negative effects on low-skill jobs but positive effects on high-skill positions, and among 40 studies, 20 found positive impacts, 15 negative, and 11 neutral.
Abstract
Artificial intelligence exhibits a heterogeneous impact on employment. Several studies report declines in low‐skill and routine roles. For example, one analysis notes a 21% drop in job posts for writing and coding tasks, and another finds a 2% decline in freelance jobs and a 5.2% reduction in earnings. In contrast, other work documents gains among high‐skill workers—estimates include a 2.6–4.3% increase in high‐skill job shares and a complementary effect on skills that can be as much as 50% larger than substitution effects. Some papers report no aggregate change as sectoral shifts offset one another, with short-term displacement giving way to longer-term job transformation.
Overall, among 40 studies the direction of AI’s impact is positive in 20 studies, negative in 15, and neutral in 11. The effects vary by worker type, industry, and region, with establishment-level declines often offset by new opportunities in services and technology-driven sectors.
Methods
We analyzed 40 sources from an initial pool of 999, using 8 screening criteria. Each paper was reviewed for 5 key aspects that mattered most to the research question. More on methods
Papers identified with Elicit search
n = 999
Papers screened using: AI-Employment Relationship, Quantitative Employment Measures, Empirical Methodology, AI Technology Focus, Empirical Data Presence, AI-Specific Content, Comparison or Baseline, Publication Type
n = 999
Papers screened out
n = 959
Papers included for extraction
n = 40
Press enter or space to select a node.You can then use the arrow keys to move the node around. Press delete to remove it and escape to cancel.
Press enter or space to select an edge. You can then press delete to remove it or escape to cancel.
Paper search
Using your research question “What is the estimated impact of AI on employment?”, we searched across over 126 million academic papers from the Semantic Scholar corpus. We retrieved the 999 papers most relevant to the query.
Screening
We screened in sources based on their abstracts that met these criteria:
- AI-Employment Relationship: Does this study examine the relationship between artificial intelligence technologies and employment outcomes?
- Quantitative Employment Measures: Does this study measure quantitative employment impacts (such as job creation, job displacement, unemployment rates, wage changes, or productivity effects)?
- Empirical Methodology: Is this an empirical study using one of the following methodologies: randomized controlled trial, quasi-experimental study, natural experiment, longitudinal study, cross-sectional study with control groups, systematic review, or meta-analysis?
- AI Technology Focus: Does this study examine specific AI technologies (such as machine learning, automation, robotics, natural language processing, computer vision, or other AI-powered systems)?
- Empirical Data Presence: Does this study contain empirical data rather than being purely theoretical, conceptual, or opinion-based?
- AI-Specific Content: Does this study focus on AI technologies rather than solely on general digitization or computerization without specific AI components?
- Comparison or Baseline: Does this study include comparison groups or baseline measurements (i.e., is it more than just a case study or descriptive study without controls)?
- Publication Type: Is this a full research article rather than a conference abstract, editorial, letter, or news article?
We considered all screening questions together and made a holistic judgement about whether to screen in each paper.
Data extraction
We asked a large language model to extract each data column below from each paper. We gave the model the extraction instructions shown below for each column.
- Study Design:
Identify the primary research design used in the study. Classify as:
- Quantitative analysis (specify type: cross-sectional, panel data, time series, etc.)
- Econometric modeling
- Empirical analysis
Look in the methods or data section. If multiple design elements are used, list all in order of prominence. Be specific about the analytical approach (e.g., two-way fixed-effect model, two-stage least squares method).
Examples of acceptable answers:
Panel data analysis across 30 provinces
Cross-country econometric analysis using Labour Force Surveys
Establishment-level analysis of online job vacancy data
Data Sources and Sample Characteristics:
Extract the following details:
- Geographic coverage (countries/regions)
- Time period of data collection
- Total number of observations/data points
- Specific data sources (e.g., Labour Force Surveys, provincial panel data, online job vacancy databases)
Look in the methods, data description, or sample section. If multiple data sources are used, list all. If exact sample size is not clear, provide the closest available estimate.
Ensure to note any specific sampling strategies or data collection methods.
- AI Exposure Measurement Method:
Describe how the study measured AI’s impact on employment:
- Specific metrics or indicators used
- How AI exposure was quantified
- Any occupational or sectoral classification methods
Locate this information in the methodology or measurement sections. If multiple measurement approaches are used, list them in order of importance.
Look for specific indicators like:
- AI occupational impact measure
- Robotics installation rates
- AI-related job vacancy analysis
- Task-based AI compatibility assessment
Provide precise details about the measurement approach.
- Key Employment Outcomes:
List the primary employment-related outcomes examined:
- Employment growth
- Job numbers
- Hours worked
- Skill composition changes
- Wage impacts
Extract from results and discussion sections. Ensure to capture both quantitative metrics and qualitative observations about employment changes.
Include specific numerical findings if available, with units and statistical significance where mentioned.
- Primary Findings on AI-Employment Relationship:
Summarize the study’s main conclusions about AI’s impact on employment:
- Overall directional effect (positive, negative, neutral)
- Conditional effects or nuanced findings
- Mechanism explanations for observed impacts
Extract from results, discussion, and conclusion sections. Focus on the most prominent findings that directly address the AI-employment relationship.
Capture any heterogeneous effects across different worker groups, industries, or skill levels.
Results
Characteristics of Included Studies
Study
Study Focus
Methodology
Sample/Data
Time Period
Full text retrieved
Demirci et al., 2025
Impact of generative AI on online freelancing platforms
Quantitative time series analysis
Global freelancing platform, Google Trends
8 months post-ChatGPT
No
Damioli et al., 2022
AI technologies and employment (supply side)
Panel data, system Generalized Method of Moments (GMM)
3,500+ global AI-patenting firms
2000–2016
No
Acemoglu and Restrepo, 2020
AI and jobs in US online vacancies
Establishment-level panel data, econometric modeling
US online vacancies (Burning Glass)
2010–2018
Yes
Georgieff and Hyee, 2022
AI and employment, cross-country
Cross-country econometric analysis, fixed effects
23 Organisation for Economic Co-operation and Development (OECD) countries, labor force surveys
2012–2019
Yes
Acemoglu et al., 2022
AI and jobs in US online vacancies
Establishment-level panel data, econometric modeling
US online vacancies (Burning Glass)
2010–2018
Yes
Schumann et al., 2021
AI and employment, cross-country
Cross-country econometric analysis
23 OECD countries, labor force surveys
2012–2019
No
Kanagarla, 2024
AI and employment transformation, multi-sector
Mixed-methods: econometric, interviews, case studies
327 sectors, 142 interviews, 47 case studies
2020–2024
No
Coelli and Borland, 2019
Critique of automation risk predictions
Empirical analysis, regression
US Occupational Information Network (O*NET), Bureau of Labor Statistics (BLS), 701 occupations
~2010
Yes
Shen and Zhang, 2024
AI, robots, and employment in China
Panel data, two-way fixed effects, Two-Stage Least Squares (2SLS)
30 Chinese provinces
2006–2020
Yes
Bughin, “AI and Future of Work”
Corporate AI use and labor demand
Cross-sectional survey, econometric
3,000 firms, 10 countries
Not stated
No
toof
Pageof
View 30 more rows
Summary of Methodologies:
- Panel data methods: 25 out of 40 studies used panel data approaches, including fixed effects, Generalized Method of Moments (GMM), Two-Stage Least Squares (2SLS), Dynamic Ordinary Least Squares (DOLS), Fully Modified Ordinary Least Squares (FMOLS), threshold regression, and related methods.
- Cross-sectional or survey-based methods: 3 studies.
- Time series analysis: 3 studies.
- Empirical or econometric methods (unspecified): 5 studies.
- Mixed-methods (econometric, interviews, case studies): 1 study.
- Event study design: 1 study.
- Dynamic Computable General Equilibrium (CGE) model: 1 study.
- Vector Error Correction Model (VECM): 1 study.
Summary of Samples/Data:
- US data: 8 studies (including online vacancies, O*NET, BLS, patents, Stack Overflow, and worker transitions).
- Chinese data: 6 studies (provinces, firms, sectors).
- OECD country data: 3 studies.
- European or EU country data: 2 studies.
- Global samples: 4 studies (AI-patenting firms, freelancing platforms).
- G7 countries: 1 study.
- South African data: 1 study.
- Brazilian data: 1 study.
- Taiwanese data: 2 studies (electronics firms, industrial sectors).
- Samples covering 10, 16, 17, 28, 30, or 33 countries: 7 studies.
- Online labor platforms or freelancers: 2 studies.
- General firm samples (not country-specific): 2 studies.
- No mention found for sample/data: 6 studies.
Summary of Time Periods:
- 1980s coverage: 1 study (1986–2017).
- 2000s coverage: 7 studies (2000–2010).
- 2010s and early 2020s: 19 studies (2010–2023).
- 2020–2025 or immediate post-ChatGPT era: 4 studies.
- No mention found for time period: 6 studies.
Effects of AI on Employment
Overall Employment Effects
Study
Effect Direction
Effect Size
Statistical Significance
Demirci et al., 2025
Negative
21% decrease in job posts (writing/coding), 17% (image creation)
No mention found
Damioli et al., 2022
Positive
Significant positive impact of AI patents on employment
Significant
Acemoglu and Restrepo, 2020
Negative (establishment), Neutral (aggregate)
Reduced non-AI hiring; no aggregate effect
No mention found
Georgieff and Hyee, 2022
Neutral
No clear relationship; positive in high computer use jobs
No mention found
Acemoglu et al., 2022
Negative (establishment), Neutral (aggregate)
Reduced non-AI hiring
No mention found
Schumann et al., 2021
Neutral
Employment growth in most occupations; positive in high computer use
No mention found
Kanagarla, 2024
Mixed
23.4% reduction (middle-skill), 31.7% increase (new jobs)
No mention found
Coelli and Borland, 2019
Neutral
No added value of automation risk for forecasting
No mention found
Shen and Zhang, 2024
Positive
Significant positive coefficients for job growth
Significant
Bughin, “AI and Future of Work”
Neutral
Mixed expectations (growth/reduction)
No mention found
toof
Pageof
View 30 more rows
Summary of Effect Directions (across 40 studies):
- Positive effect direction: (AI associated with increased employment or job growth): 20 studies
- Negative effect direction: (AI associated with decreased employment or job loss): 15 studies
- Neutral effect direction: (no clear relationship or no significant change): 11 studies
- Mixed, complex, or threshold effects: (effects varied by country, over time, or by gender): 4 studies
- Split effects by subgroup: 10 studies
- By skill level: 7 studies (positive for high-skill, negative for low-skill)
- By sector: 1 study (positive in services, negative in industry/agriculture)
- By gender: 1 study (positive for men, threshold effect for women)
- By employment vs. wage outcome: 1 study (negative for employment, positive for wages)
Statistical Significance:
- 18 studies reported statistically significant effects.
- We didn’t find mention of statistical significance in the available full texts or abstracts for 22 studies.
Key Patterns:
- Both positive and negative effects are frequently reported, often depending on skill level, sector, or other subgroups.
- Statistical significance was reported in less than half of the studies.
Differential Effects by Worker Type
A consistent pattern across these studies is the polarization of employment outcomes:
Low-skilled and routine workers:
Disproportionately negatively affected, with job losses, reduced hours, and displacement documented in multiple studies (Bonfiglioli et al., 2024; Hui et al., 2024; Giwa and Ngepah, 2024; Marguerit, 2025; Liang et al., 2025; Wang, 2024).
High-skilled and STEM workers:
Often benefit from AI adoption, with increased employment shares, wage premiums, and new job creation (Albanesi et al., 2023; Mäkelä and Stephany, 2024; Chen and Wang, 2024; Colombo et al., 2024).
Gender and other groups:
Some studies report nuanced effects, such as more severe negative impacts on low-skilled female employment (Liang et al., 2025) or threshold effects for women with disabilities (Abid et al., 2024).
Freelancers and knowledge workers:
Generative AI may disproportionately affect even high-quality freelancers (Hui et al., 2024).
Sectoral and Firm-Level Variations
Service sector:
Generally sees positive or neutral effects, with job creation and skill upgrading (Damioli et al., 2024; Algül, 2024; Gu et al., 2022).
Industrial/manufacturing sector:
More likely to experience job losses due to automation (Chu et al., 2020; Algül, 2024; Wang, 2024).
Firm size and age:
Younger firms and those in the service sector benefit more from AI-driven job creation (Damioli et al., 2024; Damioli et al., 2021).
Regional differences:
Effects vary by region, with more negative impacts in certain provinces or countries depending on digital infrastructure and labor market institutions (Liang et al., 2025; Bonfiglioli et al., 2024).
Temporal Dynamics and Adjustment Processes
Short-term displacement:
Several studies document immediate job losses following AI adoption, especially in automation-prone roles (Demirci et al., 2025; Hui et al., 2024).
Long-term adaptation:
Over time, new job categories emerge, and skill upgrading can offset initial losses (Kanagarla, 2024; Mäkelä and Stephany, 2024).
Role of reskilling and policy:
Proactive reskilling and comprehensive AI strategies are associated with better employment outcomes and higher worker retention (Kanagarla, 2024; Adhikari et al., 2024).
Synthesis and Limitations
- Evidence base: Based on the included studies, the evidence base appears extensive and methodologically diverse, with most studies using quantitative methods and large datasets.
- Heterogeneity: The studies vary widely in context, measurement approaches, and outcomes, which complicates direct comparison and generalization.
- Measurement: Several studies rely on proxies for AI exposure, for example, patents or job postings, which may not fully capture the complexity of AI’s labor market effects.
- Relevance: The rapid evolution of AI technologies, especially generative AI, may make findings from older studies less applicable to current policy debates.
- Context-dependence: The estimated impact of AI on employment is highly context-dependent, with evidence of both displacement and augmentation.
- Variation by subgroup: The net effect varies by sector, skill level, and institutional context, highlighting the need for targeted policy interventions and ongoing monitoring.
References
J. Bughin\ (2020).How Firms will affect the Future of Work
Download BIBDownload RISDownload TXT
Report
Status
Gather sources
999 sources found
Details
Screen sources
40 sources included
Details
Extract data
200 data points extracted
Details
Generate report
Save PDF
BIBLaTeX, ZoteroRISZotero, MendeleyTXTAPA BibliographyPDFPDFDOCXMicrosoft Word
Chat
Got some follow-up questions?
Sign up or sign in to chat with this report.
Back
Artificial Intelligence and Jobs: Evidence from US Commuting Zones
Alessandra Bonfiglioli, G. Gancia, Ioannis Papadakis, Rosario Crinò
Social Science Research Network·
2024·
13 citations
SourceDOI
Plain textPDFFigures
Searching for PDF
Unable to find PDF from source
Back
Study Design
Panel data analysis using two-stage least squares method with shift-share instrument
Data Sources and Sample Characteristics
- Geographic coverage: Entire mainland of the US - Time period of data collection: 2000 to 2020 - Total number of observations/data points: 722 commuting zones (CZs) - Specific data sources: Decennial Census (2000), American Community Survey (ACS) (2010 and 2020)
AI Exposure Measurement Method
- Specific metrics or indicators used: Growth in AI-related occupations, specialized software requirements. - How AI exposure was quantified: Measured by changes in employment in AI-related occupations relative to total employment in each commuting zone. - Occupational or sectoral classification methods: O*NET database's "Hot Technologies" section to identify AI-related occupations.
Key Employment Outcomes
- Employment growth: Negative impact on overall employment, especially in low-skill and production jobs. - Job numbers: Decrease in job numbers for low-skill and production workers; increase for high-wage and STEM occupations. - Skill composition changes: Negative for low-skill, positive for high-skill workers; significant impact on middle-aged workers. - Wage impacts: Positive for top wage distribution workers, indicating potential wage polarization.
Primary Findings on AI-Employment Relationship
- Overall directional effect: Negative - Conditional effects or nuanced findings: Negative for low-skill and production workers; positive for high-skill workers and those in STEM occupations - Mechanism explanations: AI adoption concentrated in services but affects both services and manufacturing; contributes to increasing inequality - Heterogeneous effects: Negative for middle-age workers, low-skill workers; positive for high-skill workers and STEM occupations
We study the effect of Artificial Intelligence (AI) on employment across US commuting zones over the period 2000-2020. A simple model shows that AI can automate jobs or complement workers, and illustrates how to estimate its effect by exploiting variation in a novel measure of local exposure to AI: job growth in AI-related professions built from detailed occupational data. Using a shift-share instrument that combines industry-level AI adoption with local industry employment, we estimate robust negative effects of AI exposure on employment across commuting zones and time. We find that AI’s impact is different from other capital and technologies, and that it works through services more than manufacturing. Moreover, the employment effect is especially negative for low-skill and production workers, while it turns positive for workers at the top of the wage distribution and for those in STEM occupations. These results are consistent with the view that AI has contributed to the automation of jobs and to widen inequality.
1Introduction
Arti…cial Intelligence (AI) is often viewed as one of the most transformative and disruptive technologies of recent times. Thanks to improvements in machine learning techniques and the growing availability of vast amounts of digital data, the last two decades have witnessed a tremendous increase in the use of AI applications, which include web search engines, targeted advertising, recommendation systems, self-driving cars, generative or creative tools and chatbots. A pressing policy question is how these advances will a¤ect labor markets and especially employment.
On the one hand, intelligent tools promise to enhance human capabilities and create new demand for certain skills. For instance, AI can substitute existing capital (Bresnahan, 2019) and boost labor productivity (McKinsey Global Institute, 2017) . On the other hand, the fear is that AI will surpass workers in a growing set of tasks, making them redundant. For instance, AI-assisted machines can be used to automate jobs (Acemoglu, 2022, Acemoglu andJohnson, 2023). Whether AI will complement or substitute workers is therefore an empirical question, for which there is still very little systematic evidence. This paper is one of the …rst attempts at …lling this gap by studying the e¤ect of AI on employment across US Commuting Zones (CZs) over the period 2000-2020. Our analysis covers the years of the rise of the digital economy. All the major companies involved in big data collection were founded by the early 2000s: Amazon (1994), Google (1998), LinkedIn (2003) , Facebook (2004) , Twitter (2006) . There is also evidence that the di¤usion of AI technologies speeded up after 2010 (e.g., Taddy, 2019 , Alekseeva et al., 2021 , and Acemoglu et al. 2022 ). Yet, precise metrics for quantifying AI adoption are at present lacking. To overcome this limitation, we recognize that using AI necessitates skills that only professionals in a narrow range of occupations possess. Taking advantage of a novel section of the O*NET database, we classify AI-related occupations as those whose job postings most frequently require specialized software used for machine learning and data analysis. Then, we detect AI adoption from the growth in the relative importance of these AI-related occupations.
A second challenge in identifying causal e¤ects is that AI adoption might be correlated with other shocks hitting a CZ. To overcome this problem, we use a shift-share instrument that combines industry-level AI adoption for the US with CZ-level employment shares across industries. Then, guided by a simple model, we identify CZs more exposed to AI as those specialized in industries that experienced faster growth in AI-related occupations.
With this data at hand, we …rst document some patterns about the change in the employment share of AI-related occupations across industries, CZs and time. This preliminary analysis con…rms that AI-adoption varies signi…cantly and that it accelerated after 2010. We then estimate the e¤ect of AI-adoption on employment using 2SLS stacked …rst-di¤erences models for the decades 2000-2010 and 2010-2020. To control for other characteristics of a CZ that may in ‡uence AI adoption and employment, we include a wealth of …xed e¤ects and covariates. We also control for other technologies such as ICT and industrial robots, and follow various approaches to account for underlying trends and unobserved shocks. In all cases, we estimate robust negative e¤ects of AI exposure on employment across CZs and time. Interestingly, we …nd that the 2SLS coe¢ cient is more strongly negative than its OLS counterpart, which is consistent with the view that contemporaneous shocks may induce an upward bias. 1 inally, we dig deeper into the e¤ect of AI adoption. We start by comparing AI with other shocks studied in the literature. Then, we explore the mechanisms through which the e¤ect of AI unfolds. Di¤erently from industrial robots, we …nd that AI adoption is concentrated in the service sector, but it exerts a negative e¤ect on employment both in services and in manufacturing. Finally, we study how the employment response to AI varies by gender, age, skill and occupation. We show that the negative employment e¤ects are largest for low-skill and production workers, while they turn strongly positive for workers at the top of the wage distribution. This suggests that AI adoption may have contributed to increasing inequality. This paper is related to the large and growing literature on the labor market e¤ects of new technologies and especially automation. Particularly close is the seminal paper by Acemoglu and Restrepo (2020), which studies the e¤ects of industrial robots, from the International Federation of Robotics (IFR), on US CZs. We di¤er in several important respects. First, we focus on an entirely di¤erent technology, the adoption of AI, using a novel measure. Second, our measure of AI adoption has several advantages compared to their proxy for automation.
The IFR data, …rst used by Graetz and Michaels (2018), is available for 19 broad sectors only, while our variable has a much …ner variation (188 industries). Moreover, the IFR data are not available at the CZ level, which makes it impossible to test the mechanisms through which US-level exposure operates across localities. In terms of results, we …nd evidence that AI adoption, similarly to robots, displaces workers. However, we also …nd that AI adoption, unlike robots, operates mostly through the service sector.
A recent strand of the literature studies the evolution of occupations that are at risk of being displaced by AI. Various measures have been proposed: the AI occupational impact measure by Felten et al. (2018Felten et al. ( , 2019Felten et al. ( , 2021)); the Suitability for Machine Learning (SML) index by Brynjolfsson et al. (2018Brynjolfsson et al. ( , 2019)); and Webb's (2020) AI Exposure score. Combining these measures with information from job postings in the US, Acemoglu et al. (2022) …nd that, consistently with our approach to measuring AI adoption, establishments with AI-suitable tasks increase their demand for AI-related skills and reduce their overall hiring. Albanesi et al. (2023) study instead the evolution of occupations exposed to AI in a panel of European countries. None of these papers detect a negative relationship between displacement risk and employment at the occupation or industry level, suggesting that existing measures of displacement risk might be correlated with other shocks. Our paper is the …rst to identify negative aggregate e¤ects of AI adoption on employment. Our results are consistent with Hui, Reshef and Zhou (2023), who document a negative impact of generative AI on the employment of free-lancers in an online labor market, and with Grennan and Michaely (2020), who …nd negative e¤ects for …nancial analysts.
While most of the literature focuses on occupations that compete with AI, a few recent papers study instead the jobs that are involved in the creation of AI. Hanson (2021) selects AIrelated occupations based on keywords such as "computer", "data" or "software" in job titles, while we use speci…c software requirements included in job postings. More importantly, he uses his classi…cation to study a very di¤erent question, namely, the determinants of regional specialization in AI-related activities and not the labor market e¤ects of AI. Alekseeva et al. (2021), instead, provide some descriptive evidence on the demand for AI skills in the US across occupations and industries. Finally, relative to using vacancies to measure the demand for AI-related skills (as in Acemoglu et al., 2022), occupational data have the advantage of being available for a broader sample.
The remainder of the paper is organized as follows. Section 2 presents a simple model of the e¤ects of AI on employment to guide the empirical analysis. Section 3 describes the data and the main variables. Section 4 discusses the empirical speci…cation and the identi…cation strategy. Section 5 presents some stylized facts and preliminary evidence. Section 6 contains the main empirical results. Section 7 considers possible threats to identi…cation. Section 8 compares the results to other technologies and explores the adjustment mechanisms. Section 9 concludes.
2AI and Local Labor Demand
This section presents a simple partial-equilibrium, task-based, model of the e¤ects of AI on the demand for labor across CZs similar to Acemoglu and Restrepo (2020). The role of the model is to show that AI can automate jobs or complement workers, to illustrate how to estimate its e¤ect using variation in local exposure to AI and to clarify how to measure it. The analysis is kept deliberately simple to make the link with the empirical section as transparent as possible. Partial equilibrium is adopted both for simplicity and for consistency with the econometric speci…cations.
2.1A Simple Model of AI and Employment
The economy consists of C commuting zones. Each CZ c 2 C has identical preferences over I industries. For simplicity, trade is free across CZs and we denote with p i the price of the output of industry i 2 I. Each industry produces output by combining a speci…c capital with a continuum of tasks indexed by z 2 [0; 1], each of which can be performed by AI or labor.
The production function for industry i in CZ c is:
where x ci (z) is the output of task z and K ci is CZ c's endowment of the speci…c capital used in industry i. Di¤erences in the endowment of speci…c capital generate di¤erences in the industrial composition of employment across CZs.
Denote with ! c and w c the cost of AI and the wage in CZ c. We assume that ! c < w c .
We identify the capabilities of AI with the set of tasks it can perform. Workers can perform all tasks. However, since AI is cheaper than labor, workers will not be employed in tasks that can be performed by AI. Assuming that AI can be used in the set of tasks [0; ic ], we will have x ci (z) = A ci = ci , where A ci is the AI input used in CZ c in industry i. Labor performs the remaining tasks
, where L ci is employment in CZ c in industry i. Substituting these quantities into (1), industry output becomes:
The demand for AI services and for labor from industry i in CZ c are:
and
Substituting ( 3)-( 4) into (2) yields:
We de…ne an improvement of the AI technology as an increase in the set of tasks it can perform: d ci > 0. We also allow technological progress in AI to boost productivity of all factors: d' ci =d ci = 0. This assumption may capture phenomena such as the use of AI as a general purpose technology complementing all existing factors or the creation of new tasks.
These are mechanisms that have been emphasized in part of the literature. Since AI services are cheaper than labor and may raise productivity, industry output necessarily expands with
where c w c =! c > 1 is the cost saving of AI relative to labor.
The partial-equilibrium e¤ects on labor demand follow from di¤erentiating (4):
This equation illustrates the two key possible e¤ects of AI. The …rst term captures the negative e¤ect on labor demand when AI displaces workers in some tasks previously performed by humans. The second term is the increase in labor demand when AI raises productivity and hence total revenue of the industry. Eq. ( 6) clari…es that the displacement e¤ect of AI is the only one responsible for any negative impact on labor demand.
Aggregating the industry-level implications yields the e¤ect on local labor demand:
2.2 AI and Local Labor Demand: Empirical Specification Equations ( 5) and (7) summarize the e¤ects of advances in AI on labor demand, through the displacement of workers and the increase in productivity. In order to derive an estimation equation for employment in the most transparent way, we now make some additional simplifying assumptions. First, we assume that the supply of labor is fully elastic and set the wage as the numeraire, w c = 1. For instance, this would be the case if labor is freely mobile across CZs or if there is a binding wage ‡oor. We discuss later what happens when labor supply is not fully elastic. Second, we assume that AI services are also in fully elastic supply.
In particular, the AI input is a bundle of intermediate inputs I ic (such as computing power and data) and specialized workers H ci (e.g., programmers and data scientists), which must be combined in …xed proportions:
Third, as in Acemoglu and Restrepo (2020), we consider an initial equilibrium in which ci 0. This assumption is appropriate to study the adoption of an entirely new technology such as AI.
Under these assumptions, we can detect technological progress in AI from the increase in AI-speci…c workers. Di¤erentiating (3) and rearranging yields:
This equation shows that technological progress in AI in industry i can be measured by the increase in AI employment over raw labor. Finally, using ci 0, w c = 1; (5) and ( 8) into (7) yields:
Equation (9) shows that the e¤ect of AI can be estimated by regressing changes in employment on changes in AI-related jobs at the CZ level.
However, the main concern is that shocks to AI employment in a CZ might be correlated with other local shocks that have a direct e¤ect on employment. For instance, an increase in local demand may trigger AI adoption and simultaneously raise labor demand. Ideally, we want to use changes in AI technology that are exogenous to other labor market shocks in CZ c. To do so, we adopt a classic Bartik design and instrument dH ci =L ci with its national counterpart, dH i =L i , where dH i is the change in AI-related jobs in industry i in the US. We defer a detailed discussion of the identi…cation strategy and the possible threats to identi…cation to Section 4.
Before proceeding, we pause to brie ‡y discuss what happens when labor supply is not fully elastic. In this case, employment in CZ c must satisfy the market-clearing condition P i2I L ci = L c . Adding a labor supply equation of the form L c = w c , it is possible to solve simultaneously for L c and w c . An increase (decrease) in labor demand will then translate into an increase (decrease) in both employment and wages.
3Data and Variables
Our sample consists of 722 CZs covering the entire mainland of the US. 2 The time period of our analysis spans the last two decades, given that the surge of AI is a recent phenomenon. Speci…cally, we observe each CZ at the endpoints of each decade, i.e., in the years 2000, 2010 and 2020. We now present the data sources and explain the construction of the main variables. Descriptive statistics are provided in Appendix Table B1 .
3.1Employment, Population and Other Characteristics of CZs
For each CZ, we measure employment, population and other characteristics using micro-level data from two sources: the decennial Census for the year 2000 and the American Community Survey (ACS) for the years 2010 and 2020. Both data sources are extracted from IPUMS (Ruggles et al., 2013). To increase sample size, we follow Autor, Dorn and Hanson (2013) and Acemoglu and Restrepo (2020) in measuring 2010 variables using pooled …ve-year ACS data for 2011 (2007) (2008) (2009) (2010) (2011) . Similarly, we measure 2020 variables from pooled …ve-year ACS data for 2021 (2017-2021). 3 We construct total CZ-level employment using sample weights, considering working-age individuals (age 16+) who are not unpaid family workers, do not reside in institutional group quarters, and have reported being employed over the previous year (Autor and Dorn, 2013) .
In later sections, we also consider disaggregations of employment along various dimensions. Speci…cally, we distinguish between employment in private and public sectors as well as in primary, secondary and tertiary industries. We also disaggregate employment between highand low-skill workers, production and non-production workers, male and female workers, occupations at di¤erent points of the wage distribution, and by age. 4 Using data from the 2 CZs are clusters of counties with strong commuting ties within them and weak commuting ties among them (Tolbert and Sizer, 1996). As such, CZs may approximate local labor markets. The CZs in our sample are the same as in Autor and Dorn (2013), Autor, Dorn and Hanson (2013) and Acemoglu and Restrepo (2020). 3 The Census and the ACS are 5% and 1% samples, respectively, of the US population, and are representative at the level of micro-regions known as Public Use Microdata Areas (PUMAs). We map PUMAs to CZs using a crosswalk developed by Autor and Dorn (2013) . 4 The public sector comprises transportation, communication, other public utilities, public administration, and armed forces. The private sector is made up of all non-public sector industries. The primary sector comprises agriculture, forestry, …sheries and mining. The secondary sector comprises construction and manufacturing. The tertiary sector consists of the public sector as well as wholesale trade, retail trade, …nance, insurance, real estate, business and repair services, personal services, entertainment and recreation services, professional and related services. High-skill workers are those who have completed at least a bachelor's degree, low-skill workers those who have completed less than a bachelor's degree. Production occupations comprise construction, extraction, installation, maintenance, repair and other production occupations.
Census and the ACS, we also construct CZ-level population and numerous proxies for initial demographic and industrial composition (details in Section 4).
3.2Measuring AI Adoption within CZs
Precise metrics for quantifying AI adoption are presently lacking. To make progress, we build a novel proxy for AI adoption in each CZ by exploiting a speci…c feature of AI technologies. Speci…cally, adopting these technologies demands a wide spectrum of specialized software, which is essential for a …rm to: execute existing machine learning algorithms, which extract patterns from large datasets by exploiting results from statistics and data science; combine and adapt these algorithms to solve complex problems speci…c to the distinct needs of the …rm; generate, update and assemble the input datasets in real time; train the algorithms and govern their learning process. In turn, operating this specialized software necessitates a distinct set of skills, which are primarily possessed by professionals engaged in a narrow range of occupations pertaining to the domains of computer science, mathematics and statistics.
Our proxy for AI adoption leverages the speci…c requirement of specialized software and advanced technical skills that is typical of AI. In a nutshell, we …rst identify a set of AI-related occupations using data on the software knowledge required to workers in each job. Then, following the theoretical model, we proxy for AI adoption in a CZ using the increase in the relative importance of AI-related occupations in that locality.
To identify AI-related occupations, we take advantage of a novel section of the O*NET database called "Hot Technologies". The latter reports the software requirements that are most frequently included in all current employer job postings in the US. The list of software includes 157 titles, spanning from software with general applications like Microsoft Excel, to advanced programming languages like Python and C++. With the help of computer scientists, we narrow down the list to 54 software that are normally used for data collection and generation; for execution and adaptation of machine learning algorithms; and to feed these algorithms with large (structured and unstructured) datasets. The list of software is reported in Appendix Table B2 .
Using the "Hot Technology" section of O*NET, we identify occupations for which each software is "in demand". These are occupations to which knowledge of a given software is typically required in job postings. This yields 82 occupations, de…ned according to the 2018 version of the Standard Occupational Classi…cation (SOC). We re…ne the list by applying two sequential …lters. First, we restrict to occupations for which at least two software are "in demand". This excludes 21 occupations that use a single software in their daily activities.
Examples are "Special E¤ects Artists and Animators", who use Python, or "Commercial and Industrial Designers", who use JavaScript. Second, we use occupational de…nitions from the SOC classi…cation, together with information on each occupation's main tasks provided in O*NET, to further exclude occupations whose main activities fall outside of the domains of computer science, mathematics and statistics. This …lter eliminates occupations like "Actuaries". While both Python and SQL are "in demand" for this job, the main role of an actuary is to "analyze statistical data, such as mortality, accident, sickness, disability, and retirement rates and construct probability tables to forecast risk and liability for payment of future bene…ts."
The …nal list of AI-related occupations comprises 19 titles, which are listed in Table 1 . Using correspondence tables from the US Bureau of Labor Statistics, we track these occupations back in time across the revisions of the SOC classi…cation occurred during the sample period. Then, we use information on each worker's SOC occupation (available in the Census and in the ACS) to match our consistent set of AI-related occupations with the micro-level data. With this information, we measure employment in AI-related occupations in each CZ and year. Finally, following (9), we construct our proxy for AI adoption as follows:
where L AI c 0 and L AI c 1 denote employment in AI-related occupations in CZ c in the …rst year ( 0 ) and last year ( 1 ) of each decade t, while L c2000 is total employment (across all occupations) in CZ c in 2000. Hence, for each CZ, AIado ct measures the decadal change in the relative importance of AI-related occupations, as proxied by the employment of these professions relative to initial employment in the CZ.
Before proceeding, we pause to discuss two potential limitations of this variable. First, not all employment in AI-related occupations needs to be devoted to AI adoption. Second, employment growth in AI-related occupations may re ‡ect a wider usage of ICT independently of AI. We believe both issues to have a modest in ‡uence on AIado ct . On the one hand, the list of AI-related occupations is obtained starting from a restricted list of specialized software, and is further narrowed down by our sequential …lters. Second, because ICT have rapidly spread all over the US during previous decades, the scope for further di¤usion of ICT over our sample period is likely to be limited.
Nevertheless, in the empirical analysis, we speci…cally tackle these two issues. In a robustness check, we construct AIado ct using a single AI-related occupation, "Data Scientists", whose tasks are limited to the typical domains of AI. 5 The resulting indicator is certainly too narrow to capture the actual size of AI adoption in the US, but the patterns it delivers are identical to those of our broader proxy, both qualitatively and quantitatively. Second, we present robustness checks controlling for various proxies for ICT exposure. Controlling for these variables does not a¤ect our main results. Interestingly, we also …nd ICT to have smaller, and sometimes opposite, e¤ects on employment compared to AI.
4Empirical Specification and Identification Strategy
In this section, we present the empirical speci…cation and illustrate our identi…cation strategy.
4.1Regression Equation
Guided by the theoretical model (eq. 9), our empirical analysis relates changes in employment to AI adoption across CZs. To this purpose, similarly to Autor, Dorn and Hanson (2013) and Acemoglu and Restrepo (2020), we estimate speci…cations of the following form:
where c indexes a CZ, d denotes the Census Division to which the CZ belongs, t stands for time, and " cdt is an error term. 6 We estimate (11) by stacking two …rst-di¤erences corresponding to changes over 2000-2010 and 2010-2020, the two decades spanned by our sample. The main outcome variable, (L=P ) cdt , is the change in the employment-to-population ratio of CZ c over decade t. 7 To shed light on the mechanisms, we also consider alternative outcomes, such as unemployment and non-participation rates, and disaggregate employment across different sectors. To study heterogeneity, we further split employment across occupations, skill levels, genders and age groups. The main explanatory variable is AIado cdt , the proxy for the adoption of AI technologies in CZ c over decade t introduced in Section 3. Given the de…nition of AIado cdt in (10), eq. ( 11) is a changes-on-changes regression.
The speci…cation in (11) includes a wealth of …xed e¤ects and covariates, which control for other characteristics of the CZ that may in ‡uence AI adoption and employment. Specifically, we control for Census Division …xed e¤ects, d , to absorb heterogeneous trends in labor market conditions across groups of contiguous states. We also include decade …xed e¤ects, t , to soak up time-varying shocks hitting all CZs simultaneously. Moreover, we control for a large set of covariates, included in the vector X cdt . The latter contains two types of variables. First, it includes several proxies for initial characteristics of each CZ: (i) size, measured by log population; (ii) demographic composition, proxied by the population shares of female, college-educated, White and old individuals (age 65+); and (iii) composition of economic activities, captured by the share of manufacturing in total employment, the share of females in manufacturing employment, the share of light manufacturing in total manufacturing employment, and the employment share of workers in routine-intensive occupations. 8 These variables account for heterogeneous trends across CZs characterized by di¤erent initial demographic and industrial structures. Second, the vector X cdt includes proxies for two main labor market shocks considered in the recent literature: the change in import competition from China (Autor, Dorn and Hanson, 2013) and exposure to industrial robots (Acemoglu and Restrepo, 2020). Both variables are measured in each decade. 9
6 Census Divisions are de…ned by the US Census Bureau and subdivide the country into nine groups of states: New England, Middle Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, and Paci…c. 7 While the model yields sharp predictions for non-AI employment, we prefer to focus on total employment to be more general and conservative, also given that AI employment is still small as shown in Section 5. 8 Light manufacturing is made up of textile mill products; apparel and other …nished textile products; paper and allied products; printing, publishing and allied industries. The employment share of routineintensive occupations is de…ned as the share of hours worked in occupations with routine-task intensity at the top tercile of the distribution (Autor and Dorn, 2013) and is constructed using data from the 2000 Census. 9 Import competition from China is a Bartik measure of changes in imports from China per worker across industries, as in Autor, Dorn and Hanson (2013). Exposure to industrial robots is a Bartik measure of The coe¢ cient of interest is . Our empirical design implies that this coe¢ cient measures the relative di¤erence in the growth of employment relative to population across CZs that have similar initial conditions, face similar trade and automation shocks, but experience di¤erent rates of AI adoption over time. Although the rich set of controls and …xed e¤ects absorb most observable confounders, the OLS estimate of need not have a causal interpretation, due to unobservable factors that may simultaneously a¤ect AI adoption and employment. We now turn to discussing the key identi…cation concerns and illustrate our strategy to estimate causal e¤ects.
4.2Identification Strategy
Variation in AIado cdt across CZs could re ‡ect CZ-speci…c unobservables that also in ‡uence employment. In particular, demand shocks in a CZ may lead …rms to hire more workers and use more AI technologies, inducing a spurious positive correlation between AIado cdt and (L=P ) cdt . To account for this issue, we construct an instrument that is meant to isolate variation in AIado cdt not due to demand shocks within CZs. Following a long tradition initiated by Bartik (1991) and Blanchard and Katz (1992), and recently applied to the automation literature by Acemoglu and Restrepo (2020), we use a shift-share (Bartik) instrument, which combines nation-wide industry shifts with local industry shares. The instrument is constructed as follows:
where L AI i 0 and L AI i 1 denote employment of AI-related occupations in industry i, at the national level, in the …rst year ( 0 ) and last year ( 1 ) of decade t, respectively; L i2000 is total employment in industry i, at the US level, in 2000; and cdi1980
The intuition behind this instrument is the following. As technological progress lowers the cost of AI and/or increases its productivity, AI adoption grows, especially in industries whose activities are more amenable to the use of these technologies. These industry-level AI adoption shifts have asymmetric e¤ects across CZs, depending on historical di¤erences in their industrial specialization, as captured by the employment shares cdi1980 . In the baseline version of the instrument, we measure these shares in 1980, i.e., 20 years before the changes in robot density (the number of installed robots per worker) across 29 sectors, and is constructed as in Acemoglu and Restrepo (2020) using data on robot installments from the IFR. beginning of the sample period. Because AI was largely inexistent at that time, this choice mitigates the concern that the industrial composition of a CZ might have been in ‡uenced by the anticipation of AI developments in future decades. We keep the employment shares constant to avoid endogenous and serially correlated changes in AIexp cdt in the context of our stacked …rst-di¤erences speci…cation.
To compute the industry-level shifts, we use the micro-level data from the Census and the ACS, which contain information on each worker's industry of employment. The employment shares are computed using data from the County Business Patterns, which we slightly aggregate to match the industry detail of the micro-level data. The …nal sample includes 188 NAICS industries, mostly at the 3-or 4-digit level, spanning all sectors of the economy.
It is worth noting that our instrument exploits signi…cantly larger cross-industry variation than the typical Bartik measures used in the recent automation literature. For instance, the proxies for exposure to industrial robots based on data from the IFR aggregate industry-level shifts for 19 broad sectors.
The identifying assumption behind our approach is that the industry-level shifts are exogenous to shocks occurring in individual CZs. We believe this to be a reasonable assumption, given that most CZs are tiny relative to the overall size of the US economy. Moreover, our speci…cation controls for a large set of …xed e¤ects and covariates, which absorb a wealth of potential observable confounders. Yet, our identi…cation strategy would be endangered in two cases. First, if some contemporaneous shocks remained that correlate with the outcome (L=P ) cdt and the instrument AIexp cdt . Second, if some remaining underlying trends at the CZ-level in ‡uenced employment independently of AI.
In Section 7, we use various approaches to account for underlying trends and contemporaneous shocks more ‡exibly. We …nd that these potential confounders are unlikely to drive the results. Moreover, we implement a falsi…cation test showing that current AI adoption does not explain past changes in employment. This further suggests that our results are driven by period-speci…c AI adoption rather than secular trends that predate the rise of AI. This extensive sensitivity analysis gives further credibility to our results but might not eliminate all concerns with a violation of the exclusion restriction. Hence, we also use an approach developed by Conley, Hansen and Rossi (2012) to study how strong a violation would have to be for inference on to become uninformative about the causal e¤ect of AI. We …nd that inference would remain informative even under implausibly large violations. We now illustrate the main patterns of AI adoption emerging from our data and provide preliminary evidence on the relationship between AI and employment in the US. Figure 1 The process of AI adoption is not uniform across industries. Rather, our data reveal a substantial degree of heterogeneity across the 188 industries in our sample, a crucial aspect for the identi…cation strategy. Speci…cally, the standard deviation of the industry-level AI adoption shifts is 0:3%, almost …ve times the mean, and the di¤erence between the industries at the top and bottom percentile of the distribution is 1:2%. Another distinguishing feature of AI adoption is the type of industries that are most interested by the phenomenon. Table Table Industries are classified according to the NAICS classification. The second and fourth columns report the AI adoption shift in each industry (the term in round brackets in eq. 12) averaged between the decades 2000-2010 and 2010-2020.
2 lists the top ten and bottom ten industries in terms of the average shifts over the sample period. The top industries consist of advanced services such as information services; professional, scienti…c and technical services; business services; and …nancial services. They also comprise some utilities (electricity) and some of the most advanced branches of the public sector (national security and international a¤airs; executive o¢ ces and legislative bodies).
Conversely, the bottom industries include traditional manufacturing sectors like beverage, apparel and structural clay products; utilities other than electricity; retail services; and the public administration.
Interestingly, AI adoption involves very di¤erent industries compared to the adoption of industrial robots. Indeed, Acemoglu and Restrepo (2020) show that automation is mostly concentrated in manufacturing, especially in highly mechanized industries such as automotive;
in the chemical and pharmaceuticals sectors; and in heavy activities such as production of metal products. This suggests that AI adoption and automation are distinct shocks, which are likely to have di¤erent implications for labor demand. Our empirical results will support this view.
We now discuss regional variation. Figure 2a shows the average value of AIado cdt , computed between the two decades, in each CZ; Figure 2b On the contrary, AIexp cdt is relatively low in Northern states and in most of the Midwest.
Third, the spatial variation in AIado cdt largely resembles that in AIexp cdt . This suggests that the combination of nation-wide industry-level AI adoption shifts with spatial historical di¤erences in industrial structure is a good predictor of actual AI adoption across CZs in the US. Fourth, there are notable exceptions to this pattern. For instance, some CZs in Northern states score higher in the ranking of AIado cdt than they do in terms of AIexp cdt . This is consistent with the idea that AI adoption could be driven by CZ-speci…c factors, independently of actual exposure to these technologies. This is exactly the endogenous variation in AI adoption our identi…cation strategy aims to get rid of.
6Main Results and Robustness
In this section, we present the baseline results and perform an extensive sensitivity analysis to assess their robustness. The baseline estimates of ( 11) are reported in 2) and (3) imply a negative 2SLS estimate of (coe¤. 0:671, s.e. 0:157), as shown in column (4). Also in this case, the 2SLS coe¢ cient is more negative than its OLS counterpart, consistent with OLS estimates being upward biased due to unobserved CZ-speci…c shocks.
The last four columns of Table 3 report estimates from our preferred speci…cation, which includes the full list of controls described in Section 4. The results are con…rmed, suggesting that our evidence depends neither on initial di¤erences in CZ characteristics nor on other important shocks occurred over the sample period, namely, the increase in Chinese import competition and the di¤usion of industrial robots. The 2SLS coe¢ cient reported in column (8) implies that a 1 standard deviation (s.d.) higher AIado cdt causes a reduction in (L=P ) cdt by 1 percentage point (p.p.), roughly 0:56% of a s.d.. To have a sense of the magnitude of the e¤ect, if the CZ with average AI adoption over the sample period (0:004) had hypothetically had no AI adoption at all, its employment-to-population ratio would have grown by 0:6 p.p. more, i.e., 20% faster than observed growth. While these numbers cannot be interpreted as a counterfactual exercise, they neverthess suggest that AI adoption has contributed to slowing down job creation in the US over the last two decades.
6.2Robustness Checks
We study the robustness of the baseline results along three dimensions: the presence of in ‡uential observations; the use of di¤erent corrections for the standard errors; and the adoption of alternative de…nitions for the main variables.
6.2.1Outliers
In Table 4 , we report 2SLS estimates of obtained on various sub-samples, which exclude extreme observations in (L=P ) cdt or AIado cdt . We start by dropping observations for which (L=P ) cdt is above or below the sample mean by more than two standard deviations (column 1), or for which (L=P ) cdt falls in the top or bottom 1% (column 2) or 5% (column 3) of the distribution. In the last three columns, we use the same approaches to exclude extreme observations in AIado cdt . In all cases, the coe¢ cient remains negative and very precisely estimated. In fact, excluding extreme observations either leaves the point estimate close to the baseline or makes it larger, suggesting that our results are not driven by outliers. The sample consists of 722 CZs observed over two decades, 2000-2010 and 2010-2020. Each observation is a CZ x decade pair. The subscripts c , d and t denote CZs, Census Divisions and decades, respectively. The dependent variable is the change in the employment-to-population ratio in each CZ over each decade. AIado cdt and AIexp cdt are the measures of AI adoption and AI exposure defined in eq. ( 10) and ( 12), respectively. The specifications in columns ( 1)-( 3) exclude observations for which the change in the employment-to-population ratio is above or below the sample mean by more than two standard deviations (column 1), or falls in the top or bottom 1% (column 2) or 5% (column 3) of the distribution. The specifications in columns ( 4)-( 6) exclude the corresponding observations of AIado cdt . All specifications include Census Division fixed effects, decade fixed effects and the control variables used in Table 3 Excluding Outliers in D(L/P) Excluding Outliers in AIado
6.2.2Inference
In Figure 4 con…dence intervals are very similar to the baseline one, suggesting that our results are not sensitive to the speci…c approach used for inference.
6.2.3Variables De…nitions
In Table 5 , we consider alternative ways of de…ning the main variables. We start by tackling the concern that our de…nition of AI-related occupations might be too broad, including work unrelated to AI. To this purpose, we reconstruct both AIado cdt and AIexp cdt using a narrow
Hollow circles correspond to the baseline 2SLS coefficient on AIado cdt (Table 3 , column 8). Confidence interval (1) refers to standard errors corrected for clustering at the state level. Confidence intervals ( 2)-( 7) correspond to standard errors corrected using the alternative clustering schemes indicated on the horizontal axis. Confidence interval ( 8 1) is negative and very precisely estimated also in this case. Quantitatively, the e¤ect is essentially identical to the baseline: a 1 s.d. higher AIado cdt would cause a fall in (L=P ) cdt of 1 p.p., or 0:56% of its s.d..
In column (2), we reconstruct AIexp cdt using industry shares for 1990 rather than 1980. It is reassuring that hardly changes, suggesting that AIexp cdt exploits long-lasting di¤erences in the industrial specialization of CZs, rather than period-speci…c developments in their industrial structure that could result from transitory shocks. In column (3), we adjust AIado cdt for cross-industry di¤erences in the growth rate of employment within each CZ, and AIexp cdt for cross-industry di¤erences in the growth rate of employment at the national level. While the …rst-stage relationship is slightly weaker, our main results are qualitatively unchanged and quantitatively stronger. This suggests that our evidence is not driven by changes in industries'relative size but re ‡ects variation in AI intensity within industries.
In column (4), we estimate (11) for the log change in employment, adding the log change in population among the controls. The coe¢ cient implies that a 1 s.d. higher AIado cdt would The sample consists of 722 CZs observed over two decades, 2000-2010 and 2010-2020. Each observation is a CZ x decade pair. The subscripts c , d and t denote CZ, Census Divisions and decades, respectively. The dependent variable is: the change in the employment-to-population ratio in each CZ over each decade (columns 1-3); the log change in employment in each CZ over each decade (column 4); and the change in private sector employment (column 5) or public sector employment (column 6) relative to population in each CZ over each decade. AIado cdt and AIexp cdt are the measures of AI adoption and AI exposure defined in eq. ( 10) and ( 12), respectively. In column (1), AIado cdt and AIexp cdt are based on a narrow definition of AI-related occupations, which only includes "Data Scientists". In column (2), AIexp cdt is constructed using CZ-level industry shares for 1990 rather than 1980. In column (3), AIado cdt and AIexp cdt are adjusted for cross-industry differences in employment growth rates at the CZ and national level, respectively. All specifications include Census Division fixed effects, decade fixed effects and the control variables used in Table 3 . The specification in column ( 4 lower employment by 1:3%. Finally, in columns ( 5) and ( 6), we revert to the employment-topopulation ratio as the dependent variable but split the numerator into private and public sector employment. The results show that the e¤ect of AI adoption is entirely concentrated in the private sector, where the bulk of the phenomenon is currently taking place.
7Threats to Identification
Identi…cation requires that, after controlling for Census Division …xed e¤ects, decade …xed e¤ects, initial characteristics of CZs and shocks to trade and automation, no unobservable remains that correlates with the instrument and in ‡uences employment across localities. As mentioned in Section 4, the exclusion restriction is threatened by two types of confounders, namely, underlying trends and contemporaneous shocks that might a¤ect local labor markets independent of AI adoption. We now use alternative approaches for accommodating the possible in ‡uence of these confounders and study how the coe¢ cient is a¤ected.
In Table 6 , we deal with contemporaneous shocks. To start with, we allow for the possibility that CZs with similar changes in employment or similar AI adoption might be hit by similar unobservable shocks. To accommodate them, we divide CZs into ten bins corresponding to the deciles of (L=P ) cdt or AIado cdt over 2000-2020. We then interact a dummy for each bin with decade dummies. We add these interactions either individually (columns 1 and 2) or jointly (column 3). These interactions absorb shocks hitting all CZs with comparable changes in our key variables. Accordingly, identi…cation exploits the remaining variation across CZs falling in the same bin within each decade. In column (4), we take a complementary approach and exclude CZs with the highest values of AIado cdt . These are CZs falling in the top decile of the distribution in a given decade. The coe¢ cient remains negative, precisely estimated and in the same ballpark as the baseline estimates across all speci…cations.
A related concern is that the industry shifts that make up the instrument might be driven by shocks occurring in speci…c CZs. If these shocks also a¤ected the local labor market, the exclusion restriction would be violated. We believe this concern to be assuaged by the small size of most CZs relative to the US as a whole. Nevertheless, in column (5), we estimate (11) using a "leave-one-out" instrument, in which the industry-level shifts are computed after excluding the CZ to which the instrument refers. Our main results are preserved also in this case.
Unobserved shocks may also play out at the industry level. For example, industries with large AI adoption might experience other shocks that are relevant to the labor market of some CZs. Our speci…cation already controls for trade and automation shocks, by means of Bartik The sample consists of 722 CZs observed over two decades, 2000-2010 and 2010-2020. Each observation is a CZ x decade pair. The subscripts c , d and t denote CZs, Census Divisions and decades, respectively. The dependent variable is the change in the employment-to-population ratio in each CZ over each decade. AIado cdt and AIexp cdt are the measures of AI adoption and AI exposure defined in eq. ( 10) and ( 12), respectively. In column ( 5), AIexp cdt is constructed excluding the CZ to which this variable refers. In column (6), AIado cdt and AIexp cdt are constructed excluding industries whose shifts (the round bracket terms of eq. 12) fall in the top decile of the distribution. All specifications include Census Division fixed effects, decade fixed effects and the control variables used in Table 3 . The specification in column (1) controls for full sets of interactions between decade dummies and dummies for deciles of the change in the employment-to-population ratio over 2000-2020. The specification in column (2) controls for full sets of interactions between decade dummies and dummies for deciles of AIado cdt over 2000-2020. The specification in column (3) jointly includes the two sets of interactions used in columns ( 1) and ( 2). The specification in column ( 4 measures of changes in Chinese import penetration and in robot density. Moreover, in the next section, we will enrich the speci…cation with Bartiks for di¤erent forms of technical change, and our results will turn out to be insensitive to these additional controls. Nevertheless, we now use two complementary approaches to further address concerns with industry-speci…c shocks.
In a …rst exercise, reported in column (6), we reconstruct both AIado cdt and AIexp cdt excluding industries whose AI adoption shifts fall in the top decile of the distribution. The coe¢ cient of interest remains close to the baseline estimate. Interestingly, this exercise also serves two additional purposes. On the one hand, it further helps mitigating concerns with outliers. On the other hand, because some of the industries with the largest shifts could be AI-producing sectors, excluding them may better isolate the e¤ect of AI adoption. In a second exercise, shown in column (7), we augment the speci…cation with a Bartik measure of changes in log industry employment. This variable serves as a synthetic proxy for all shocks resulting in di¤erential changes in employment across industries. Controlling for this measure is largely inconsequential for the results.
We now discuss underlying trends. Our speci…cation controls for linear trends both across Census Divisions-through the Census Division …xed e¤ects, d -and across CZs with similar initial characteristics within each Census Division-through the start-of-period controls included in the vector X cdt . In Table 7 , we allow for richer sets of …xed e¤ects to accommodate underlying trends more ‡exibly. In column (1), we replace the Census Division …xed e¤ects with state …xed e¤ects, thereby allowing for heterogeneous linear trends across individual states rather than across groups of them. The coe¢ cient of interest is essentially unchanged.
In column (2), we go one step further by taking advantage of our stacked …rst-di¤erences design, which allows us to include CZ …xed e¤ects. In this case, identi…cation relies on a di¤erent source of variation, namely, changes in (L=P ) cdt and AIado cdt within each CZ between the two decades. This speci…cation fully exploits the acceleration in AI adoption occurred after 2010, but is extremely demanding because the sample comprises only two observations for each CZ. 11 The estimate of is largely una¤ected. In a last exercise, we replace the Census Division and decade …xed e¤ects with Census Division decade …xed e¤ects (column 3) and state decade …xed e¤ects (column 4), thereby allowing for non-linear trends.
The main results are preserved.
In the remainder of this section, we use two additional approaches to further raise trust in our 2SLS estimates. In column (5) of Table 7 , we implement a falsi…cation test by regressing changes in the employment-to-population ratio prior to the beginning of the sample (over 1980-1990 and 1990-2000) on current AI adoption (in 2000-2010 and 2010-2020). We include the same controls as in the baseline speci…cation, and instrument AIado cdt using AIexp cdt . If current AI adoption explained past changes in employment, our evidence could re ‡ect labor market trends that antedate the rise of AI, or long-lasting CZ characteristics that jointly a¤ect innovation and employment. Reassuringly, however, the coe¢ cient on AIado cdt is very small and statistically not signi…cant. In column (6), we include the pre-sample change in the employment-to-population ratio among the controls. Consistently with column (5), this variable has no bearing on the coe¢ cient . Overall, these results help strengthening the view that our evidence captures the e¤ects of current AI adoption rather than other time-varying confounders.
In a second approach, we change perspective and start from the premise that our instrument might be correlated with the error term due to some confounding factor. Then, following Conley, Hansen and Rossi (2012), we study how strong a violation of the exclusion restriction would have to be for inference on to become uninformative about the causal e¤ect of AI adoption. We illustrate the approach of Conley, Hansen and Rossi (2012) in Appendix A. Here, we present the main results, which are summarized in Figure 5 . The latter plots the 90% con…dence interval around corresponding to di¤erent violations of the exclusion restriction, as captured by the parameter . When = 0, we are in the benchmark case in which the exclusion restriction is satis…ed. When > 0, instead, the exclusion re- The sample consists of 722 CZs observed over two decades, 2000-2010 and 2010-2020. Each observation is a CZ x decade pair. The subscripts c , d and t denote CZs, Census Divisions and decades, respectively. Except for column (5), the dependent variable is the change in the employment-to-population ratio in each CZ over each decade; in column ( 5), the dependent variable is the change in the employment-to-population ratio in each CZ over the pre-sample decades 1980-1990 and 1990-2000 . AIado cdt and AIexp cdt are the measures of AI adoption and AI exposure defined in eq. ( 10) and ( 12), respectively. All specifications include Census Division fixed effects, decade fixed effects and the control variables used in Table 3 . The specifications in columns ( 1)-( 4) control for state fixed effects, CZ fixed effects, Census Division x decade fixed effects and state x decade fixed effects, respectively. The specification in column ( 6 Overall, these …gures suggest that even substantial, and likely implausible, relaxations of the exclusion restriction would leave inference informative about the employment e¤ect of AI adoption.
8Additional Evidence
In this section, we dig deeper into the e¤ect of AI adoption. We start by comparing AI with other shocks studied in the literature. Then, we explore the mechanisms through which the
8.1AI Adoption and Other Shocks
Our model shows that AI adoption a¤ects labor demand in two ways: by replacing workers in some tasks (displacement e¤ect) and by increasing e¢ ciency (productivity e¤ect). Accordingly, the e¤ects of AI adoption should di¤er from those of other shocks that do not have a displacement e¤ect. We now compare AI adoption with various shocks of this type. The results are reported in Table 8 . In columns ( 1)-( 6), we augment (11) with proxies for di¤erent shocks; in column (7), we include all these proxies together. The speci…cation in column (1) includes a proxy for capital deepening, which is a Bartik measure of the change in capital intensity (capital-to-labor ratio) across industries. 12 The coe¢ cient is largely unchanged, suggesting that our results are not capturing the e¤ect of capital deepening. At the same time, the coe¢ cient on the new control is very small and positive, consistent with the e¤ect
In column (5), we consider other productivity-enhancing shocks. To remain agnostic about the precise nature of the shock, we add a broad proxy, which is obtained as a Bartik measure of the change in log industry value added. This proxy enters with a positive and statistically signi…cant coe¢ cient, but its inclusion is inconsequential for our coe¢ cient of interest. Hence, AI adoption di¤ers from productivity-enhancing shocks, and has distinct e¤ects on the labor market. Finally, in column (6) we compare AI adoption with o¤shoring.
To this purpose, we add the initial employment share of o¤shorable occupations, constructed using data from Autor and Dorn (2013) . The coe¢ cient is largely unchanged. Moreover, while o¤shoring could also displace workers, the results show that AI adoption has markedly di¤erent labor market e¤ects. 14
8.2Channels
So far, our evidence highlights a negative impact of AI adoption on employment across CZs.
We now explore some of the mechanisms underlying this e¤ect. The results are reported in Table 9 . We start by analyzing the sectors that contribute the most to the overall e¤ect.
As shown in Section 5, the leading industries in terms of AI adoption belong to the service sector, while manufacturing is still lagging behind. It is conceivable, therefore, that AI adoption in services might currently have larger e¤ects on labor demand than AI adoption in manufacturing. To study the role of the two sectors, we split both AIado cdt and AIexp cdt in two separate variables, constructed as in (10) and (12) on the subsets of manufacturing and non-manufacturing industries, respectively. We then estimate (11) using the sector-speci…c variables in place of the aggregate variables. The results are reported in column (1) for manufacturing and in column (2) for non-manufacturing. Consistently with the di¤erent speed of AI di¤usion in the two sectors, the estimates show that the service sector makes up the lion's share of the overall e¤ect.
A related issue has to do with the sectors that experience the largest changes in employment as a consequence of AI adoption. The previous results do not imply that the e¤ect is entirely concentrated in the service sector, because industries are linked to each other by upstream or downstream relationships and because workers may move across sectors in response to shocks. To investigate the employment e¤ects in di¤erent sectors, we divide the numerator of the dependent variable into employment in primary, secondary and tertiary industries. We also consider employment in manufacturing, which makes up the secondary sector together with the construction industry. We then estimate (11) using employment in each branch of
8.3Heterogeneity
In columns (1) and (2) of The largest negative e¤ect is found for middle-age workers, which account for almost 88% of the e¤ect on overall employment. The impact of AI adoption is also negative, albeit much smaller, on younger workers, while it is small and imprecisely estimated on older workers. One possible interpretation of these results is that younger workers are better equipped to cope with the challenges posed by AI, as they are capable of swiftly updating their skills. Older workers are instead shielded from the AI shock possibly due to their tendency to hold more stable jobs. Notably, these advantageous traits are often less pronounced among middle-aged workers.
A second important dimension of heterogeneity is education. As shown in columns (1) and (2) of Table 11 , the negative e¤ect of AI adoption is entirely concentrated on lowskill workers. For high-skill workers, the estimated coe¢ cient is positive albeit imprecisely estimated. Occupations also play a role. As shown in columns (3)-( 6), the e¤ect of AI adoption is always negative for production workers. On the contrary, for non-production workers, the e¤ect is negative only among low-skill ones. These results strengthen the view that the implications of AI adoption di¤er from those of industrial robots, whose e¤ects are mostly concentrated on production workers and largely independent of education (Acemoglu and Restrepo, 2020). This di¤erence may be due to two factors. On the one hand, the two technologies have spread in di¤erent industries: while AI is prevalent in advanced service industries, automation is more prominent in highly mechanized manufacturing sectors. On the other hand, the skill requirement of AI technologies is higher than that of industrial robots.
In conclusion, we consider another dimension of heterogeneity, which is the focus of an important stream of literature. It is well documented that, both in the US and in other industrialized countries, the labor market has undergone a process of job polarization: employment The most accredited explanation for job polarization is routine-biased technical change, i.e., the di¤usion of technologies that substitute labor in standardized and codi…able activities, while complementing it in abstract and manual-intensive tasks. Our interest is to study whether AI adoption has contributed to job polarization over the last two decades. The previous results suggest that this might not be the case: as previously shown, the e¤ects of AI adoption on overall employment are di¤erent from those of software and computers, which are the main culprits of job polarization according to the literature (e.g., Autor and Dorn, 2013) .
To provide direct evidence, in Table 12 , we estimate (11) using as the dependent variable the change in employment, as a share of population, for occupations in each quintile of the initial wage distribution. The results are inconsistent with a contribution of AI adoption to job polarization. Indeed, there is no evidence of a U-shaped pattern in the estimated coe¢ cients. Rather, the coe¢ cients follow a seemingly monotonic trend: they are strongly negative for the bottom two quintiles, mildly negative for the third and fourth quintiles, and strongly positive at the top of the distribution. This pattern is broadly in line with our previous evidence by skill group, as long as an occupation's position in the wage distribution is correlated with its skill requirements. Overall, these results suggest that AI adoption poses threats to low-and medium-skill workers, while o¤ering new opportunities to the most educated individuals. Hence, AI adoption may be a source of rising inequality in the US labor market.
9Conclusions
Recent improvements in the …eld of AI have triggered much hype. The ongoing debate highlights the fact that AI is a ‡exible technology with the potential of turning dreamlike scenarios or nightmares into reality. Nobody can predict the direction that future innovations and applications will take. However, to inform policy decision, it is important to understand the consequences of these technologies so far. The goal of this paper has been to study the e¤ect of AI adoption on labor demand as measured by changes in employment. Since the deployment of AI can potentially increase productivity but also automate work, its impact on employment is a still unanswered empirical question. 15 sing data across US CZs over the period 2000-2020, a novel measure of AI adoption based on the growth of AI-related jobs and a shift-share empirical strategy to identify causal e¤ects, we were able to estimate robust negative e¤ects of AI exposure on employment. We also found that AI's impact is di¤erent from other forms of capital and technologies, such as robots or ICT, and that it works through services more than manufacturing. Moreover, the employment e¤ect is especially negative for low-skill and production workers, while it turns positive for workers at the top of the wage distribution. Overall, these results are consistent with the view that AI, so far, has contributed to the automation of jobs and to widen inequality. Finally, while the focus of this paper has been on employment so as to best capture displacement e¤ects, AI adoption is likely to have a¤ected wages as well. In the 10) and ( 12), respectively.
annex
The sample consists of 722 CZs observed over two decades, 2000-2010 and 2010-2020. Each observation is a CZ x decade pair. The subscripts c , d and t denote CZs, Census Divisions and decades, respectively. The dependent variable is the change in the employment-to-population ratio in each CZ over each decade. AIado cdt and AIexp cdt are the measures of AI adoption and AI exposure defined in eq. ( 10) and (12), respectively. CapInt cdt , SoftInt cdt , CompInt cdt and CommInt cdt are Bartik measures of the change in, respectively, the capital-to-labor ratio, the software capital-to-labor ratio, the computer capital-to-labor ratio and the communication equipment-to-labor ratio, across industries. VA cdt is a Bartik measure of the change in log industry value added. Offsh cdt is the initial employment share of offshorable occupations in each CZ. All specifications include Census Division fixed effects, decade fixed effects and the control variables used in Table 3 of capital deepening being di¤erent from that of AI.
Next, we study the implications of ICT. In columns (2)-(4), we add three Bartik measures of changes in (i) software, (ii) computer and (iii) communication equipment intensities, respectively. In all speci…cations, the e¤ect of AI adoption remains very close to the baseline estimate. This reassures against the concern that AIado cdt might capture the e¤ect of computers, software and other high-tech capital. Also in this case, the coe¢ cients on the new controls are small and, when they are precisely estimated, their sign is positive. This pattern suggests that AI adoption has di¤erent labor demand consequences compared to ICT. The sample consists of 722 CZs observed over two decades, 2000-2010 and 2010-2020. Each observation is a CZ x decade pair. The subscripts c , d and t denote CZs, Census Divisions and decades, respectively. In columns ( 1)-( 2), the dependent variable is the change in the employment-to-population ratio in each CZ over each decade. In columns ( 3)-( 6), the dependent variables are the changes in primary sector employment, secondary sector employment, manufacturing sector employment and tertiary sector employment, respectively, as a share of population in each CZ over each decade. In columns ( 7)-( 8), the dependent variables are the changes in the number of unemployed workers and in the number of individuals out of the labor force, respectively, relative to population in each CZ over each decade. AIado cdt and AIexp cdt are the measures of AI adoption and AI exposure defined in eq. ( 10) and ( 12), respectively. In columns ( 1)-( 2), these variables are constructed on the subsets of manufacturing and non-manufacturing industries, respectively. All specifications include Census Division fixed effects, decade fixed effects and the control variables used in the economy, relative to population, as the dependent variable. The results are reported in columns ( 3)-( 6). The largest e¤ect is found in the tertiary sector, which accounts for roughly 60% of the impact of AI adoption on overall employment. In addition, negative e¤ects are evident also in the secondary sector, primarily in manufacturing, which accounts for 45% of the aggregate impact. Finally, the results show a small positive e¤ect on employment in the primary sector, which probably absorbs a small fraction of workers laid o¤ from the other branches.
The last results raise the question of what happens to displaced workers that are not re-employed. To answer this question, in the last two columns, we study the response of unemployment and non-participation rates. Speci…cally, we estimate (11) using two di¤erent dependent variables: unemployment as a share of population (column 7) and the share of population out of the labor force (column 8). The results indicate that AI adoption raises both unemployment and non-participation. However, non-participation absorbs a much larger share of the overall reduction in employment compared to unemployment (75% vs. 25%).
This fraction includes both workers who temporarily leave the labor force to update their skills and discouraged workers who drop out of the labor force. In the next section, we study heterogeneity in the e¤ect of AI adoption, with the aim of shedding light on the possible winners and losers from this new technology.
interest of space, we left a detailed analysis of this possibility to future research.
We conclude by discussing some policy implications of our …ndings. The existing literature has already pointed out that automation, in the form of industrial robots, is partly to blame for falling manufacturing employment (e.g., Acemoglu and Restrepo 2020, Dauth et al. 2021, and Bon…glioli et al. 2022a,b) . Our alarming result is that service employment, which constitutes the lion's share in developed countries, may not be immune to automation through AI. It is therefore crucial that governments put in place measures aimed at alleviating adverse labor-market consequences. Since the negative e¤ects are concentrated among low-skill workers, a …rst remedy could be to help employees acquire new skills. Facilitating job-to-job transitions and improving the ‡exibility of the labor market can also ease reallocation costs.
Since top earners seem to actually bene…t from these new technologies, appropriate transfer schemes could also be used to ensure that the gains are more broadly shared. Finally, incentive schemes could be designed to redirect innovation towards applications aimed at improving human capabilities rather than labor savings. To this end, collecting better data and developing methodologies to identify the complementarities between AI applications and jobs seem an important step for future research.
Appendix A The Conley, Hansen and Rossi (2012) approach
In this section, we illustrate the main idea behind the approach of Conley, Hansen and Rossi (2012) using our set-up. Consider the following version of ( 11):
where is a parameter measuring the size of a violation of the exclusion restriction. The baseline results presented in the text are based on the standard IV assumption that = 0.
However, if the exclusion restriction was not satis…ed, i.e., if 6 = 0, inference on could still be performed, using alternative priors about and conditional on this parameter. This can be done by estimating the following speci…cation We set to be a function of a parameter , which we progressively raise (by intervals of 0.01 starting from 0) to generate increasingly larger violations of the exclusion restriction.
Speci…cally, we set 1:594 11 , where 1:594 is the baseline 2SLS estimate of and the interquartile range of AIado cdt is approximately 11 times that of AIexp cdt . For each value of , we estimate the con…dence interval of for both the lower and the upper end of the support [ ; ], and compute the …nal con…dence interval as the union of the two con…dence intervals. 16 Appendix B Data Appendix 16 Besides this "union of con…dence intervals" approach, Conley, Hansen and Rossi (2012) discuss other strategies that use more prior information about . By imposing additional parametric restrictions, these alternative approaches tend to yield narrower con…dence intervals around the treatment parameter, and thus may be less conservative.
Failed to load PDF:
StripeM-Inner
Paper sources
Abstract screening pilot
Abstract screening results
Extraction pilot
Extraction results
Research report
Modify setup