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

Technology and Labor Displacement: Evidence from Linking Patents with Worker-Level Data

Kogan, Papanikolaou, Schmidt, Seegmiller

2023NBER Working Paper Series10 citations
Exposure / measurementCausalTheoretical model
Automation / RobotsAI (General)Junior / entry-levelSenior / older workersAugmentation vs. substitutionRoutine task changeGeneral automationOccupational mobility
DOI: 10.3386/w31846
Abstract

We develop measures of labor-saving and labor-augmenting technology exposure using textual analysis of patents and job tasks.Using US administrative data, we show that both measures negatively predict earnings growth of individual incumbent workers.While labor-saving technologies predict earnings declines and higher likelihood of job loss for all workers, laboraugmenting technologies primarily predict losses for older or highly-paid workers.However, we find positive effects of labor-augmenting technologies on occupation-level employment and wage bills.A model featuring labor-saving and labor-augmenting technologies with vintage-specific human capital quantitatively matches these patterns.We extend our analysis to predict the effect of AI on earnings.

Summary

Kogan, Papanikolaou, Schmidt, and Seegmiller develop time-varying measures of labor-saving and labor-augmenting technology exposure using textual analysis of patents and occupation tasks, then link these to US administrative earnings data to study heterogeneous effects on incumbent vs. new workers across 1978-2016.

Main Finding

Labor-saving technologies reduce earnings of exposed workers by 2-3% over five years across both blue- and white-collar workers; labor-augmenting technologies increase total compensation and employment but reduce earnings for incumbent workers (especially older and higher-paid workers) by 1-2% while raising earnings for new entrants by 2-4%, consistent with vintage-specific human capital displacement.

Primary Datasets

USPTO Patent

Other

LEHD

Labor outcomes

SSA Detailed Earnings Records (DER) linked to Current Population Survey (CPS); USPTO patent data; Dictionary of Occupational Titles (DOT); Census Longitudinal Business Database (LBD)

Secondary Datasets

ACS

Labor outcomes

O*NET

Other

Decennial Census (1980, 1990, 2000); American Community Survey (ACS 2008-2012); BLS industry productivity data; O*NET task data; Kelly et al. (2021) breakthrough patent classifications; BEA Input-Output tables

Key Methods
Shift-share IV using knowledge spillovers from prior breakthrough patents; panel regression with occupation-year and industry-year fixed effects; individual worker fixed effects and controls for age and earnings history; textual analysis using word embeddings (GloVe) and GPT-4 for task classification
Sample Period
1978-2016
Geographic Coverage
United States
Sample Size
Approximately 2.8 million person-year observations from administrative data; 64,500 occupation-industry-year cells from Census/ACS data
Level of Analysis
Individual, Occupation, Industry, Task
Occupation Classification
SOC 6-digit (O*NET 2010); occ1990dd (modified Census)
Industry Classification
NAICS 4-digit
Notes
NBER Working Paper 31846 [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill. [Claude classification]: Uses restricted-access Census administrative data (SSA Detailed Earnings Records linked to CPS) to track individual workers 1978-2016. Constructs novel time-varying exposure measures distinguishing labor-saving vs labor-augmenting technologies using patent-task textual similarity. Shift-share IV exploits knowledge spillovers across patent technology classes. Model calibrated via GMM to match 17 moments. GPT-4 used to classify DOT tasks as routine/non-routine and low/high skill.