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
May AI Revolution Be Labour-Friendly? Some Micro Evidence from the Supply Side
G. Damioli, Vincent Van Roy, D. Vértesy, M. Vivarelli
Social Science Research Network·
2021·
5 citations
SourceDOI
Plain textPDFFigures
Searching for PDF
Unable to find PDF from source
Back
Study Design
Panel data analysis; Econometric modeling using system GMM approach
Data Sources and Sample Characteristics
- Geographic coverage: Worldwide (Asia: 61%, Europe: 31%, United States: 6%) - Time period of data collection: 2000-2016 - Total number of observations/data points: 26,137 observations from 3,510 firms - Specific data sources: PATSTAT database of the European Patent Office, ORBIS Intellectual Property database
AI Exposure Measurement Method
- Specific metrics or indicators used: Natural logarithm of the number of AI patent families, patent family size (number of countries where a patent is protected). - How AI exposure was quantified: Through the number of AI patent families and patent family size. - Any occupational or sectoral classification methods: Controls for industry-, year-, and country-specific differences in employment dynamics.
Key Employment Outcomes
- Employment growth: Positive and significant impact of AI patent families on employment, with an estimated elasticity of 3/4%. - Job numbers: Limited to service sectors and younger firms. - Hours worked: Not mentioned. - Skill composition changes: Not mentioned. - Wage impacts: Labour cost per employee inhibits labour demand with a strong negative effect (-0.52).
Primary Findings on AI-Employment Relationship
- Overall directional effect: Positive impact of AI patent families on employment in the supply industries. - Conditional effects: Limited to service sectors and younger firms; AI-specialized companies show larger impacts. - Mechanism explanations: AI technologies are labour-friendly in the supply industries, particularly in service sectors and among younger firms, with increasing returns for companies focused on AI.
This study investigates the possible job-creation impact of AI technologies, focusing on the supply side, namely the providers of the new knowledge base. The empirical analysis is based on a worldwide longitudinal dataset of 3,500 front-runner companies that patented the relevant technologies over the period 2000-2016. Obtained from GMM-SYS estimates, our results show a positive and significant impact of AI patent families on employment, supporting the labour-friendly nature of product innovation in the AI supply industries. However, this effect is small in magnitude and limited to service sectors and younger firms, which are the leading actors of the AI revolution. Finally, some evidence of increasing returns seems to emerge; indeed, the innovative companies which are more focused on AI technologies are those obtaining the larger impacts in terms of job creation
1.Introduction
The past two decades have witnessed major developments in artificial intelligence (AI) technologies. Similar to previous technological revolutions -as the diffusion of the ICTs in the last decades of the past century -AI displays a remarkable disrupting potential across firms, industries, economies and societies.
In particular, the possible adverse impact of AI diffusion and robotics on employment has generated concerns and vivid discussions in the academic debate and in the society as a whole. Indeed, the arrival of internet of things, self-driving autonomous cars and widespread robots has raised again a fear of a new wave of 'technological unemployment'. In this vein, according to Brynjolfsson and McAfee (2011 and2014), the root of the current employment problems is not the Great Recession, but rather a "Great Restructuring" having an ever-bigger impact on jobs, skills, and the whole economy. In fact, AI, self-learning algorithms and human-imitating robots can perform tasks usually requiring human beings' intelligence and dexterity (such as speech recognition, decision-making advise, disease diagnostics, complex documents translation, performance of unhealthy and dangerous tasks and so forth; see Frey and Osborne 2017). Moreover, Dobbs et al. (2015) from the McKinsey Global Institute estimate that, compared to the industrial revolution of the XIX century, automation and AI's disruption of society is happening 10 times faster and at 300 times the scale. This kind of disruptive potentiality might affect each job and every task, although 'matching tasks' that are close substitutes for capabilities of AI are the most affected group so far (by, for instance, Uber, Airbnb, Linkedin, Amazon; see Ernst et al. 2018).
However, one of the limitations of the current debate and the extant literature (see next section) is their sole focus on the demand side (that is the adoption of AI and robots as labour-saving process innovations in the downstream industries), while there is an obvious gap to be filled with regard to the supply side, that is the possible job-creation effect of AI technologies, conceived as product innovations in the upstream sectors. Indeed, if we adopt a Schumpeterian perspective (see Schumpeter, 1912; Porto et al., 2021), technological change cannot be reduced to labour-saving process innovation. The other side of the coin is the introduction of new products (both in manufacturing and services) which entail the raise of new branches of production and the creation of additional employment opportunities.
As was the case of ICT, AI and robots may be seen at the same time as labour saving process innovations in the user sectors (think, for instance, to the massive adoption of robots in the automotive industry) and as labour-friendly product innovations in the supply industries (think, for instance, to the electronic industry, where robots are produced, or to the scientific and technical services, where AI algorithms are conceived). In this framework, AI products not only entail the emergence of entire new sectors or a substantial expansion of existing ones (such as those related to robotisation), but also the creation of brand-new employment opportunities such as those related to data processing, transactional procedures, customerization, remote collaboration, etc.
The aim of this study is precisely to assess the possible job-creation impact of AI technologies, focusing on the supply side, namely the providers of the new knowledge base. In more detail, our analysis will be based on a worldwide set of 3,500 front-runner companies that patented the relevant technologies over the 2000-2016 time span. Controlling for the other main drivers of employment at the firm level -namely output, cost of labour and capital formation -our purpose is to detect a possible labour-friendly impact of AI and robots, seen as product innovations within the supply side of the new technologies. 1 he paper is organised as follows. Section 2 summarises the extant literature, emphasising its limited focus on the labour-saving impact detectable in the adoption of new technologies. Section 3 describes the microeconometric methodology used in our analysis. Section 4 discusses the data and the sample used for the empirical analysis. Section 5 presents and discusses the main results. Finally, Section 6 wraps up and puts forward some conclusions and tentative policy implications.
2.The literature
As mentioned above, the extant economic literature focuses on the possible labour-saving effect of AI and robots, conceived as process innovation in the user industries. In doing so, recent contributions belong to a long-lasting tradition of studies devoted to the controversial relationship between technology and employment (for a long-term historical analysis, see Staccioli and Virgillito, 2021; for a recent theoretical reprise of the issue, see Acemoglu and Restrepo, 2018 and 2019) .
Indeed, the relationship between innovation and employment is a 'classical' controversy, where a clash between two views can be singled out. One states that labour-saving innovations create technological unemployment, as a direct effect. The other view argues that product innovations and indirect (income and price) effects can counterbalance the direct effect of job destruction brought about by the process innovations incorporated in new machineries and equipment (for fully articulated surveys, see Pianta, 2005;Vivarelli, 2014;Calvino and Virgillito, 2018;Ugur et al., 2018;Barbieri et al., 2020).
As far as the employment consequences of the current widespread diffusion of AI and robots are concerned, the empirical literature provides both macroeconomic forecasting scenarios and some sectoral and microeconomic evidence.
As far as the macro scenarios are concerned, Frey and Osborne (2017), using a Gaussian process classifier applied to data from O*Net and US Department of Labor, predict that 47% of the occupational categories, mostly middle-and low-skilled professions, are at high risk of being substituted by AI algorithms and robots (including a wide range of service/whitecollar/cognitive tasks such as accountancy, health professions, logistics, legal works, translation and technical writing).
However, Arntz et al. (2016 and 2017) , proposing the same exercise but using also information on task-content of jobs at individual-level, conclude that only 9% of US jobs are at potential risk of automation. Their main message is that, within the same occupation, some tasks can be automatized while others cannot and therefore the associated job can be preserved.
Extending the analysis to a multi-country approach, Nedelkoska and Quintini (2018) estimate the risk of automation for individual jobs in 32 OECD countries. Their evidence shows that about 14% of jobs are highly automatable (probability of automation over 70%), while another 32% of jobs present a risk of being substituted in between 50 and 70%, pointing to the possibility of significant changes in the way these jobs will be carried out as a result of automation.
At the European level, Pouliakas (2018) -using data on tasks and skill needs collected by the European Skills and Jobs Survey (ESJS) -bundles jobs according to their estimated risk of automation. Following Frey and Osborne (2017) and Nedelkoska and Quintini (2018), the author utilises highly disaggregated job descriptions and shows that 14% of EU adult workers are found to face a very high risk of automation.
Turning our attention to the sectoral and microeconomic evidence, the extant empirical literature has particularly focused on robotisation.
For instance, Acemoglu and Restrepo (2020) investigate the employment effect of the exposure to robots, using the sectoral "International Federation of Robotics" (IFR) data (national penetration rates instrumented by European data). According to their 2SLS estimates, robotization has a significant negative impact on the change in employment and wages in each US local labour market over the period 1990 -2007. In more detail, they show that one more robot per thousand workers reduces the employment/population ratio by about 0.18/0.34%.
Following the approach adopted by Acemoglu and Restrepo, Chiacchio et al. (2018) apply it in the context of EU labour markets. They assess the impact of industrial robots on employment and wages in 116 NUTS regions of six EU countries, namely Finland, France, Germany, Italy, Spain, and Sweden, largely representative of the European automation. Their results suggest that robot introduction is negatively associated with the employment rate (one more robot per thousand workers reducing the employment/population ratio by about 0.16/0.20%). Graetz and Michaels (2018) use panel data on robot adoption (IFR and EUKLEMS data to estimate robot density) within industries in 17 countries from 1993 to 2007. Dividing employees in three skills' groups (namely high-, medium-and low-skilled workers), their estimated employment coefficients for the two higher-skilled groups result positive (but limited in magnitude and not always significant), while the coefficient for the low-skilled workers turns out to be large and negative. However, their main finding is at odds with the studies discussed above, since they conclude that robots do not significantly reduce total employment, although they do reduce the low-skilled workers' employment share.
Finally, Dauth et al. (2017) propose a local empirical exercise on Germany using IFR data over the 1994-2014 time-span, using a measure of local robot exposure for every region. They find no evidence that robots cause total job losses, although they provide evidence that robots do affect the composition of aggregate employment: while industrial robots have a negative impact on employment in the manufacturing sector, there are positive and significant spillover effects as employment in the non-manufacturing sectors increases and, overall, counterbalances the negative impact in manufacturing.
As obvious on the basis of what discussed so far, the extant literature only focuses on the possibly negative employment impact of AI and (particularly) robots conceived as process innovations, while there is no evidence so far about the possible positive employment effect of these new technologies, considered as product innovations in the supply sectors. This is unfortunate, since product innovations are theoretically considered as drivers of job creation (see Katsoulacos, 1984; Vivarelli, 1995;Edquist et al., 2001) and several empirical studies have provided evidence in support of this view (see Freeman and Soete, 1987 and1994;Bogliacino and Pianta, 2010;Bogliacino et al., 2012;Van Roy et al., 2018).
However, the labour-friendly impact of product innovations may vary according to their nature. Indeed, new products may be either brand-new entities or substitutes of obsolete ones. If revenues from new products cannibalize the sales of old ones, the net result in terms of employment expansion might be ambiguous. For instance, a new AI algorithm might be alternatively an entire novelty or an advanced substitute of a previous version of a given software package.
In other words the "welfare effect" should be compared with the "substitution effect" (using the terminology originally put forward by Katsoulacos, 1984 and 1986; see also Vivarelli, 1995 andDosi et al., 2021). Empirically, this means that the expected sign of the correlation between product innovation and employment is positive, but uncertain in magnitude.
The aim of the next sections is to fill the discussed gap in the recent literature devoted to the employment impact of automation, providing some novel evidence on the link between product innovation and employment using longitudinal data from the front-runner companies in patenting AI and robot technologies worldwide.
3.Econometric methodology
Similarly to the prior microeconometric literature investigating the employment effects of technological change using longitudinal firm-level datasets (see Van Reenen, 1997;Lachenmaier and Rottmann, 2011;Bogliacino et al., 2012;Van Roy et al., 2018), we derive our empirical specification from a stochastic version of a standard labour demand, augmented with an innovation proxy. Along these lines, the labour demand function for a panel of firms i over time t is defined as: , = , + , + , + , + + , with: i = 1, .., n; t = 1, .., T
(1) Lower case letters denote natural logarithms, l corresponds to labour (proxied by employment level), y to output (proxied by turnover), w to wages (proxied by labour cost per employee), and I to gross investments (proxied by growth in fixed capital). The labour demand function is augmented with a variable innov capturing technological change due to innovations. Lastly, is an unobserved firm-specific and time-invariant effect and the usual error term.
We subsequently move from this static expression (1) to a dynamic specification as in (2) in order to account for viscosity in labour demand (see Arellano and Bond, 1991;Van Reenen, 1997):
: i = 1, .., n; t = 1, .., T
As measure for technological change, we use AI and non-AI patent families (respectively denoted by , and , - ), as outlined in specification (3). In essence, this specification aims to proxy for the technological progress brought about by firms through the development of innovative and marketable (and hence patentable) technologies in AI and non-AI related fields:
Dynamic labour demand specifications as in (2) suffer from simultaneity and endogeneity problems, which may lead to a biased estimation of the different covariates if they are not accounted for by the empirical model. 2 To tackle both problems, we use a system GMM approach as developed by Blundell and Bond (1998, 2000) . The system GMM model employs instrumental variables to provide consistent and efficient estimates when dealing with dynamic panel data, as is the case for our data. In a system of equations (i.e. in level equations and equations in differences), lagged and differenced lagged variables are used to solve for persistency in times series and endogeneity.
Unfortunately, possible problems of endogeneity are not confined to the lagged dependent variable. Other explanatory variables of the labour demand function may also be affected as pointed out in prior literature (e.g. Bogliacino et al., 2012;Van Roy et al., 2018). It might be the case that wage and employment are simultaneously decided, while the output and investment decisions may be jointly affected by a temporary shock. Therefore, in line with previous studies, all the explanatory variables have been considered as potentially endogenous to labour demand and instrumented when needed.
In more detail, in the level equation, we used differenced values of the explanatory variables as instruments, i.e. twice-lagged or thrice-lagged differences in labour demand, AI and non-AI patent families, gross investments and cost of labour. The level equations also include a set of sector, country and year dummies. In the equations in differences we employed twice-lagged or thrice-lagged values of the above-mentioned right-hand side variables as instruments. The lag limits of the instruments were chosen both to satisfy the outcomes of the autocorrelation tests3 and to limit instrument proliferation, as highlighted in Roodman (2009a, b) .
4.1Data
Our novel dataset is based on a worldwide panel of firms patenting in AI. While traditionally considered as a proper measure of a commercially valuable innovation output (Griliches, 1990;Ernst, 1995), limitations of patents in capturing innovations are well known (e.g. Hussinger, 2006;Hall et al., 2014): for instance, the fact that companies may prefer not to patent their inventions to keep them secret. Moreover, patents better proxy product rather than process innovations, which are often embodied in machineries and can be more easily kept secret than products (Levin et al., 1987;Lissoni et al., 2013). Yet, patents have the appealing advantage to allow identifying on a global scale firms that innovate in AI. Accordingly, an increasing number of studies relies on patents to track and analyse the development and adoption of AI technologies in production processes and their economic consequences (Webb et al., 2018; WIPO, 2019; Baruffaldi et al., 2020;Van Roy et al., 2020;Damioli et al., 2021;Martinelli et al., 2021).
However, the identification of patents related to AI technologies is a challenging task. There are neither an established definition of the boundaries of the AI technological domain, nor an agreed methodology to empirically singling it out. On the one hand, conceptual definitions of AI typically insist on the ability of a system to perform human-like cognitive functions (learning, understanding, reasoning and interacting) aiming to obtain rational outcomes (Ertel, 2018;Russell and Norvig, 2016). On the other hand, albeit AI technologies focus on a core of software technologies including, inter alia, machine learning, neural networks, logic programming and speech recognition, various studies consider a broader definition of AI including a combination of software and hardware components, as well as functional applications such as robots and "big data" (European Commission, 2018; Fujii and Managi, 2018; WIPO, 2019) .
Van Roy et al. (2020), on which this study relies upon for the selection of patents, follow the latter more comprehensive view by applying a keyword-based approach consisting in the search for specific terms in the title or the abstract of patents. An analogous approach to select AI patents has been pursued in previous studies on AI and robotics technologies (Keisner et al., 2015;De Prato et al., 2019; European Commission, 2018; Cockburn et al., 2019; WIPO, 2019; Baruffaldi et al., 2020). Some of these studies applied their keywords' search to patents falling in pre-selected technological classes (Keisner et al., 2015;Cockburn et al., 2019; WIPO, 2019) , while others relied on all patents falling in specific technological classes mapping to AI technology areas (e.g. Inaba and Squicciarini, 2017;Fujii and Managi, 2018; OECD, 2017) . 4 n this study, we rather prefer unrestricting our patents to pre-determined technological classes in reason of the transversal nature of AI technologies that, as any other general-purpose technology, cut through many scientific disciplines and technological domains (Bianchini et al., 2020; WIPO, 2019) . Our list of keywords takes stock of the findings of prior relevant literature and is shown in Table A1 in Appendix A.
Text-mining searches have been conducted to retrieve the keywords in either the patent title or the abstract of the Spring 2018 edition of the PATSTAT database of the European Patent Office covering more than 90 patent authorities including all the major countries. (see Van Roy et al., 2020 for additional details). We grouped retrieved patents in patent families to avoid double counting of the same or similar inventions filed in different patent offices. We then obtained key company characteristics of AI patent applicants through Bureau van Dijk Electronic Publishing (BvD) ORBIS databases. While the issues of coverage and data availability are known limitations of ORBIS, these are by far outweighed by its advantage of offering a comprehensive cross-country micro-level dataset for scientific research purposes (e.g. Gal, 2013;Hallak and Harasztosi, 2019). We used patent application numbers to track applicant firms in the ORBIS Intellectual Property database, which we also exploited to retrieve all other non-AI patent applications of such firms, and matched location and economic information from the ORBIS Companies database. Figure A1 in Appendix A shows a synopsis chart illustrating the data collection process.
4.2Variables and sample
After excluding observations with missing values in employment, value added, fixed assets or cost of labour and outliers in both levels and growth rates 5 , our final dataset covers 3,510 firms (resulting in 26,137 observations) active in AI patenting over the years 2000-2016. It provides a worldwide coverage and includes firms belonging to manufacturing and service sectors. It comprises information on firms' patenting activities in AI and non-AI related fields, accounting information (including employment, turnover, value added, capital formation, and cost of labour), year of birth or consolidation, country location, and industrial activity (NACE sector at 2-digit level).
Our dependent variable is the natural logarithm of the number of firm employees in head counts. 6 Explanatory variables are derived from a standard labour demand function and include the natural logarithm of firm turnover, labour cost per employee and gross investments measured as the annual change in fixed assets. We expect a positive impact on labour demand of turnover and gross investments and a negative impact of labour cost. The models also control for industry-, year-and country-specific differences in employment dynamics.
The key explanatory variable of interest is the natural logarithm of the number of AI patent families, which measure the development and adoption of the AI technology as broadly defined in the previous sub-section. In addition to AI patent families, we also take into account innovative efforts in non-AI related fields through the number of non-AI patent families.
Since patents can differ both in economic and technological value, simple patent count may give a distorted measure of a firm's technological basis. Therefore, other indicators have been proposed to correct for quality or value of patents, such as forward citation-weighted patents and family size (Harhoff et al., 2003;Hall et al., 2005;Gambardella et al., 2008;Neuhäusler et al., 2011;Squicciarini et al., 2013;Van Roy et al., 2018). Weighting patents with forward citations is particularly difficult in the case of AI, as the real take-off of AI patenting activities take place relatively late (since the late 2000s), and boomed only from 2015 onwards 5 Data cleaning and outlier treatment are described in Appendix A. 6 The description of the variables used in the empirical analysis is summarised in Table B1 of Appendix B. (Cockburn et al., 2019; WIPO, 2019; Van Roy et al., 2020). We therefore resume to using patent family size, i.e. the number of countries in which an invention is protected by a patent. Given the costs of acquiring patents in multiple jurisdictions, patent family size is often used to approximate the value that applicant attribute to the invention (Harhoff et al., 2003;Lanjouw and Schankerman, 2004;Johnstone et al., 2012). To get a measure at the firm level, we took the family size of the patent application as the average size of all applications a company made in each year, and we computed it separately for AI and non-AI patents. Before log transformation, all variables were shifted positively by 1 in order to accommodate 0 values.
Table 1 reports the summary statistics of the dependent and explanatory variables used in the estimations. Firms in the sample report an average of over 4,200 employees, due to the large presence of big corporations. AI patent families are relatively low, about 1 AI-related family per year every three firms in the sample, due to a highly skewed distribution with a large number of firm-year pairs without any AI family (about 86% of all observations). By contrast, non-AI patent activity is considerable, 30 yearly patent families on average per each firm, which confirm that the sample includes highly innovative companies. As for patent family size, non-AI patents are filed in more patent offices than AI applications, even after conditioning on the firm-year pairs with at least one application: the conditioned average of the family size is 1.8 for AI patents and 2.5 for non-AI patents. Table 2 displays the distribution of firms according to their main activity, age and AI intensity, which are the subsamples of particular interest for the following analyses. 7 The majority of the firms in the sample are active in manufacturing (55%, vs. 45% of active in services), have been founded or consolidated after 1990 (67%) and are small or medium enterprises (SMEs, 66%). The largest share of firms belongs to electronics (22%) and machinery (15%) manufacturing industries, and telecommunication (20%) and scientific services (10%). We refer to Table B2 in Appendix B for a more detailed distribution of firms across sectors. We also consider intensity in AI patenting measured by the ratio of AI patents over the total number of patents in the period. In particular, we consider AI-specialized companies those with the ratio of AI patents over total patents above the revealed median in our sample (4.8%). 8 n terms of geographical distribution, the majority of firms are located in Asia (61%). This large percentage is driven by the dominating AI patenting activity of Japanese and South-Korean firms as highlighted in prior studies (WIPO, 2019; Van Roy et al., 2020). About 31% of the firms are located in Europe, with highest percentages in Germany, France and the United Kingdom. Lastly, firms in the United States constitute around 6% of the sample.9 Age is defined based on the year of foundation or consolidation of the firm. AI-specialized companies are those with the share of AI patents over total patents in the period above the median. SMEs are defined following the EC recommendation in which SMEs are denoted as firms with a number of employees below 250 and a turnover equal to or below € 50 million (European Commission, 2003) . The number of SMEs and large firms does not sum up to the total number of firms because 218 firms have changed size category over the period. Notes: N=26,137 observations. Values expressed as natural logs, apart from gross investments that is expressed as the log difference of fixed assets between time t and t-1. Industry, country and year dummies are omitted due to space limitation.
5.1Model selection
To support the chosen methodology, Table B4 in Appendix B reports the estimation coefficients for pooled ordinary least square (POLS), fixed-effects (FE) and system generalized method of moments (SYS-GMM) models. Lagged employment is highly significant in all three different estimations tested. Its magnitude ranges from 0.495 in the FE estimation, to 0.854 in the POLS estimations. While FE tends to underestimate the impact of the lagged dependent variable, POLS, by contrast, overestimates it. Solving for persistency and endogeneity, it is therefore to be expected that the SYS-GMM estimates for the lagged dependent variable fall within these two boundaries. In fact, the coefficient for lagged employment, obtained from the one-step SYS-GMM at 0.523, meets the above methodological expectation. This finding applies to the standard setup as well as to an alternative one where, for reflecting patent quality, family size is used instead of patent counts (with FE estimates for lagged employment at 0.500, SYS-GMM at 0.532 and POLS at 0.857). Unsurprisingly, labour demand is persistent and autoregressive, confirming its path dependency.
As per diagnostics for the baseline SYS-GMM model, the Wald test on the overall significance of the regressions and the LM tests on AR(1) and AR(2) autocorrelation dynamics confirm the robustness of the model. For the Hansen test on the adequate instruments, however, the null hypothesis is rejected. Blundell and Bond (2000) and Roodman (2009a) demonstrated that, for very large samples, the Hansen test tends to over-reject the null hypothesis. Following the extant literature, the model was re-estimated on random sub-samples comprising 10% of the baseline observations, where the null of the Hansen test was never rejected, suggesting the validity of the selected instruments. 11 Finally, given the observation that a high number of instruments may imply a downward bias in the standard errors for two-step SYS-GMM models (Roodman, 2009b) , we opted for the more conservative one-step methodology.
5.2Baseline results
The baseline estimation reported in Table 4 (applying the one-step SYS-GMM estimation on the full sample of 26,137 observations from 3,510 firms active in AI patenting) provides results that are in line with prior studies of similar setup (Bogliacino et al., 2012;Van Roy et al., 2018;Pellegrino et al., 2019). The coefficients of the explanatory variables are significant and have the expected sign. The large positive effect of lagged employment (0.52) confirms the persistence of labour demand. The positive effect of turnover is large (0.26) and highly significant, while the effect of gross investmentswhich can also incorporate the adoption of process innovationsis more contained (0.03), but still significant at the customary 5% level.
The labour cost per employee inhibits as expected labour demand with a strong negative effect (-0.52).
Regarding our key innovation variables, model estimates show positive and highly significant effects of AI and non-AI patent families on employment. They imply a similarly moderate elasticity of labour demand to AI and non-AI patent families equal to about 3-4%. This finding supports the employment friendly nature of product innovation in general, and provides novel evidence for AI technological development. When patent family size is used, only AI patent families provide a significant positive effect on employment, while the effect of non-AI patent does not reach customary levels of statistical significance.
5.3Sample splits
In order to enrich our understanding of the impact of AI innovation on employment, results for three sample splits are presented in Table 4 and 5, namely: sector of main economic activity, age and AI intensity as defined in Section 4.2. We note that the overlaps between the groups singled out in the three splits are sufficiently limited to allow differentiated findings. 12 In all the considered subsamples, the main control variables in our specification -i.e., lagged employment, turnover, gross investments, and labour cost per employee -are significant and have the expected sign as in the baseline model.
When distinguishing companies by sectoral belonging, the coefficients for AI and non-AI patents are only significant among service sector firms, but not for manufacturing ones (we recall that about 45% of the companies in the sample are in services, about three-quarters in what could be considered as knowledge-intensive sectors). 13 Compared with the baseline estimations, the employment impact for firms in service industries is stronger in magnitude and 11 Results are available from the authors upon request. 12 As shown in the Venn diagram in Figure B1 in the Appendix, about 83% of the companies from the full sample are included in the union of the three groups (services, established after 1990 and AI-intensive), and about 25% of the firms are in the intersection of the three. 13 The lack of significance of any of the patent variables in manufacturing companies is also observed when further distinguishing high-tech and low-tech manufacturing sectors (results available from the authors upon request).
applies for AI and non-AI patents defined according to both patent measures. In fact, it is only among services that the coefficient of both AI and non-AI patent family size is significant (0.05 and 0.03, at 1% and 5% significance levels, respectively).
Recognizing that many of today's AI and robotics innovations goes back to the booming of the ICT revolution since the 1990s, we further investigated heterogeneous effects for sub-samples comprising firms established before and after 1990. 14 The results for the focal patent variables show that AI and non-AI patents are only significant for the younger companies (around 2/3 rd of the firms in the sample). The effects observed for these firms turn out to be highly significant for both AI and non-AI patent counts, while patent family size leads to significant coefficients only for AI technologies, similarly to the baseline scenario.
Some companies stand out even among the technologically active firms constituting our sample, based on the relative importance of their AI patents. 15 The estimation results show that AI and non-AI patents are only significant for the set of AI-specialized companies (in this case, the choice of the patent measure -count or family size -does not make any difference with regard to the level of significance). This is an important result, hinting at the magnitude of potential job creation by the leading front-runner product innovators in AI; moreover, the positive and significant effect of their non-AI patents on labour demand underlines the fact that, even for AI-specialized firms, AI turns out to be complementary to other labour friendly innovations.
Finally, in all the estimations reported in Table 4 and 5 (with the only exception of patent counts in services) the labour-friendly impact of AI technologies is larger in magnitude -and sometimes more significant -than the labour-friendly effect of non-AI innovations. If we jointly take into account this outcome and the observed role of AI-intensive firms in driving the detected job-creating effects (see above), we may conclude that the emerging AI technologies are those that drive the overall positive employment impact of product innovation in the investigated companies.
14 1990 may be an arbitrary threshold; we note however that adjusting it by 5 years has limited impact on the findings (results available from the authors upon request). 15 Recall that we refer to those specializing in AI as the firms that are above the median in the relative share of AI technologies among their patent portfolios.
6.Conclusions
In contrast with a literature solely devoted to forecast and assess the possible labour-saving impact of automation in the user sectors, this paper departs from a view considering AI devices (including robots) only as process innovations and investigates the possible labour-friendly nature of these technologies, seen as product innovations in the supply industries. Our main results can be summarised as follows.
Our estimates reveal a positive and significant impact of AI patent families on employment, supporting the labour-friendly nature of product innovation in the AI supply industries. Interestingly enough, this positive employment impact is additional and larger when compared to the job creation effect of other patenting activities.
However, the labour-friendly employment effect of patenting in AI technologies is small in magnitude (the estimated elasticity being equal to 3/4%) and unlikely able to compensate the possible labour-saving impact in the downstream user industries (see the studies surveyed in Section 2). Moreover, the positive employment impact is limited to service sectors and younger firms, which are the leading actors of the AI revolution.
Finally, some evidence of increasing returns seems to emerge; indeed, the innovative companies which are more focused on AI technologies are those obtaining the larger effects in terms of job creation.
Putting together our results, we can conclude that the possible employment benefits of AI technologies are, at least so far, mainly coming from companies that are at the core of the current technological revolution.
These pieces of evidence suggest that technological leaders within the emergence of the AI paradigm can realize (modest) labour-friendly outcomes; however, heterogeneity is also detected, with manufacturing, older and less innovative companies unable to couple product innovation with job creation.
In terms of (tentative) policy implications, these findings call for a cautious aptitude in considering the actual magnitude of the job creation effect of new technologies: compared with the labour-saving impact implied by the usage of the AI technologies (massive according to some studies, see Section 2), the labour-friendly extent in the supply industries appears limited in magnitude and scope.
Nevertheless, industrial and innovation policies should consider promoting these new and emerging sectors, being sure to achieve positive complementary targets in terms of employment creation. In particular, the revealed evidence of possible increase returns might support an increase in subsidies targeted to those companies already fully engaged in AI patenting.
At the same time, safety nets and active labour market policies keep to being necessary to deal with the employment displacement due to the widespread diffusion of AI technologies in the user industries.
Obviously enough, this study is affected by some limitations that might be overcome by future research. In particular, our results are micro-based and focusing on companies active in AI patenting, as discussed in Section 3. On the one hand, this means that our findings cannot be generalized at a more aggregate and macroeconomic level; on the other hand, this study does not claim to be representative of the entire population of firms, but rather to investigate the link between product innovation and employment within the sub-population of companies actively engaged in AI technologies. Further research is needed to assess the employment impact of the new knowledge basis at the level of the entire economy, taking into account both innovators, non-innovators and user firms.
Data cleaning
Data cleaning was carried out in 3 steps. The first step was the identification of the most common clerical errors or typos in the key economic variables (number of employees, turnover, fixed assets and cost of employees) from the Orbis database (for instance, missing 000's). Moreover, we used the imputation procedure put forward by Hallak and Harasztosi (2019); in particular, data for number of employees missing in a year between two known time points (not more than 4 years apart) were imputed applying a linear interpolation. No imputation was carried out for missing gross investments and cost of employees figures due to high annual fluctuation.
In a second step, outliers in year-on-year growth rates for the key economic variables were removed (we decided to use growth rates in order to take into account firm's size: i.e. an additional increase of 10 employees may imply doubling the size of a small firm, while growing by a few percentage points for a large corporation). The thresholds applied corresponded to trimming a bit less than the bottom and top 1% of the growth distributions of the different variables.
Finally, in a third step, the top 1 percentile in levels were also trimmed with regard both the key economic variables as well as the patent variables.
Data collection
Figure A1 presents a synoptic chart illustrating the data collection process. The Tools for Innovation Monitoring (TIM), which is an analytics tool developed by the Joint Research Centre to support policy making in the field of innovation and technological development, has been used for the patent search. TIM provides access to patent documents of the PATSTAT database from the European Patent Office and allows for text-mining searches. TIM contains patents from more than 90 patent authorities including all the major patenting countries and regroups patent documents per patent family when at least one of the members of the family is in English.
To retrieve AI patents, text-mining techniques have been used to detect the keywords presented in Table A1 in either the patent title or the abstract.
Once the relevant patent families in AI have been retrieved with this text-mining technique, the patent family number has been linked with the ORBIS Intellectual Property database. Subsequently, patent applicants have been linked to their BVD ID number to obtain account information on turnover, employment, fixed capital and sector.
In a final step, the complete patent portfolio of the identified AI-active companies was retrieved to populate the dataset on non-AI patent families.
Funding
The authors are grateful to the participants of the Workshop "The Economics and Management of AI Technologies" (Bureau d'Economie Théorique et Appliquée, Université de Strasbourg) and of the GLO Virtual Seminar.Giacomo Damioli acknowledges financial support from the European Union's Horizon 2020 Framework Programme under the 'Innova Measure IV' project (Grant agreement no.857088).Marco Vivarelli acknowledges the support by the Italian Ministero dell'Istruzione, dell'Università e della Ricerca (PRIN-2017, project 201799ZJSN: "Technological change, industry evolution and employment dynamics"; principal investigator: Marco Vivarelli).The scientific output expressed here does not imply a policy position of the European Commission or the International Telecommunication Union and their Member States.Neither the European Commission nor
Failed to load PDF:
StripeM-Inner
Paper sources
Abstract screening pilot
Abstract screening results
Extraction pilot
Extraction results
Research report
Modify setup