This site is a work in progress and has not been widely shared. Content may contain errors. Feedback is welcome.
This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.
Back to papers

How Are Patented AI, Software and Robot Technologies Related to Wage Changes in the United States?

Fossen, Samaan, Sorgner

2022Frontiers in Artificial Intelligence19 citations
Observational labor market
LLM / Generative AIAutomation / RobotsMachine Learning (pre-LLM)AI (General)Augmentation vs. substitutionGeneral automationRoutine task change
DOI: 10.3389/frai.2022.869282
Abstract

We analyze the relationships of three different types of patented technologies, namely artificial intelligence, software and industrial robots, with individual-level wage changes in the United States from 2011 to 2021. The aim of the study is to investigate if the availability of AI technologies is associated with increases or decreases in individual workers' wages and how this association compares to previous innovations related to software and industrial robots. Our analysis is based on available indicators extracted from the text of patents to measure the exposure of occupations to these three types of technologies. We combine data on individual wages for the United States with the new technology measures and regress individual annual wage changes on these measures controlling for a variety of other factors. Our results indicate that innovations in software and industrial robots are associated with wage decreases, possibly indicating a large displacement effect of these technologies on human labor. On the contrary, for innovations in AI, we find wage increases, which may indicate that productivity effects and effects coming from the creation of new human tasks are larger than displacement effects of AI. AI exposure is associated with positive wage changes in services, whereas exposure to robots is associated with negative wage changes in manufacturing. The relationship of the AI exposure measure with wage increases has become stronger in 2016–2021 in comparison to the 5 years before. JEL Classification: J24, J31, O33.

Summary

Fossen, Samaan, and Sorgner use OLS regression on CPS-ASEC panel data (2011-2021) combined with Webb's patent-based technology exposure measures to study associations between AI, software, and robot technologies and individual wage growth in the United States.

Main Finding

AI exposure is associated with wage increases of 2.68 percentage points per standard deviation (strengthening over time), while software and robot exposure are associated with wage decreases of 3.26 and 2.46 percentage points respectively, suggesting AI has transformative rather than destructive effects on labor.

Primary Datasets

CPS

Labor outcomes

Webb (2019)

AI-focused

Current Population Survey Annual Social and Economic Supplement (CPS-ASEC) 2011-2021 via IPUMS-CPS; Webb (2020) patent-based technology exposure measures

Secondary Datasets

O*NET

Other

OES/OEWS

Labor outcomes

O*NET database for task descriptions and occupation characteristics; Bureau of Labor Statistics Occupational Employment Statistics (May 2018); Blinder-Krueger offshorability scores; routine task intensity measures following Acemoglu-Autor (2011)

Key Methods
OLS regression of individual annual wage growth on occupation-level technology exposure measures from Webb (2020), controlling for individual demographics, occupation characteristics, industry fixed effects, and year fixed effects. Standard errors clustered at occupation level.
Sample Period
2011-2021
Geographic Coverage
US
Sample Size
58,394 person-year observations (main specification 2016-2021); 131,539 observations (full period 2011-2021)
Level of Analysis
Individual, Occupation
Occupation Classification
SOC codes linked to CPS occupation codes; 435 occupations merged
Industry Classification
Census industry codes (52 major industry groups)
Notes
Frontiers in Artificial Intelligence, vol. 5 [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results. [Claude classification]: Uses Webb (2020) exposure measures which employ NLP to extract verb-noun pairs from patents and link to O*NET tasks. Authors explicitly state they cannot establish causal relationships. The paper distinguishes 'transformative' digitalization (AI) from 'destructive' digitalization (software/robots). Analysis excludes pandemic years as robustness check with similar results.