This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.

Quantifying the Distribution of Machine Learning's Impact on Work

Brynjolfsson, Mitchell, Rock, Syverson

2023NBER Working Paper Series

Theoretical / conceptualTheoretical model

LLM / Generative AIMachine Learning (pre-LLM)Computer Vision / Image AIAI (General)General automationAugmentation vs. substitution

Summary

Brynjolfsson, Rock, and Syverson develop a theoretical framework arguing that AI as a general purpose technology creates a productivity paradox where rapid capability advances coexist with slow measured productivity growth due to implementation lags and unmeasured complementary intangible capital investments.

Main Finding

The paper resolves the productivity paradox by showing that AI as a general purpose technology requires substantial time and complementary investments (intangible capital) before productivity gains appear in aggregate statistics, creating a J-curve pattern where measured productivity may initially decline during rapid intangible capital accumulation before rising when those investments yield returns.

Primary Datasets

Brynjolfsson Customer Service AI

AI-focused

Proprietary customer support chat data from Fortune 500 enterprise software company; AI platform usage logs from AI software vendor

Secondary Datasets

AI system outage logs; agent attrition records

Key Methods: Theoretical framework with formal model of intangible capital accumulation; historical productivity growth analysis; descriptive statistics on TFP and labor productivity trends across OECD; back-of-envelope calculations for specific AI applications
Sample Period: 2020-2021
Geographic Coverage: US (customers); Philippines, US and other countries (agents)
Sample Size: 30 OECD countries; quarterly US data 1948-2016; decadal averages for regression analysis
Level of Analysis: Country, Firm, Occupation
Occupation Classification: None
Industry Classification: None

Notes

Forthcoming [Claude classification]: This is the first study examining generative AI deployment at scale in a real workplace. The paper includes a small initial RCT (50 agents, 7 weeks) followed by staggered rollout. Uses multiple robust DiD estimators (Sun-Abraham, Borusyak et al., Callaway-Sant'Anna, de Chaisemartin-D'Haultfoeuille). Text analysis uses all-MiniLM-L6-v2 for embeddings and SiEBERT for sentiment. Authors note they cannot observe wage effects, overall labor demand changes, or skill composition of new hires. Revised November 2023 version of April 2023 working paper. [Claude classification]: This is the first study examining generative AI deployment at scale in a real workplace. The paper includes a small initial RCT (50 agents, 7 weeks) followed by staggered rollout. Uses multiple robust DiD estimators (Sun-Abraham, Borusyak et al., Callaway-Sant'Anna, de Chaisemartin-D'Haultfoeuille). Text analysis uses all-MiniLM-L6-v2 for embeddings and SiEBERT for sentiment. Authors note they cannot observe wage effects, overall labor demand changes, or skill composition of new hires. Revised November 2023 version of April 2023 working paper. [Claude classification]: This is the first study examining generative AI deployment at scale in a real workplace. The paper includes a small initial RCT (50 agents, 7 weeks) followed by staggered rollout. Uses multiple robust DiD estimators (Sun-Abraham, Borusyak et al., Callaway-Sant'Anna, de Chaisemartin-D'Haultfoeuille). Text analysis uses all-MiniLM-L6-v2 for embeddings and SiEBERT for sentiment. Authors note they cannot observe wage effects, overall labor demand changes, or skill composition of new hires. Revised November 2023 version of April 2023 working paper. [Claude classification]: This is the first study examining generative AI deployment at scale in a real workplace. The paper includes a small initial RCT (50 agents, 7 weeks) followed by staggered rollout. Uses multiple robust DiD estimators (Sun-Abraham, Borusyak et al., Callaway-Sant'Anna, de Chaisemartin-D'Haultfoeuille). Text analysis uses all-MiniLM-L6-v2 for embeddings and SiEBERT for sentiment. Authors note they cannot observe wage effects, overall labor demand changes, or skill composition of new hires. Revised November 2023 version of April 2023 working paper. [Claude classification]: This is the first study examining generative AI deployment at scale in a real workplace. The paper includes a small initial RCT (50 agents, 7 weeks) followed by staggered rollout. Uses multiple robust DiD estimators (Sun-Abraham, Borusyak et al., Callaway-Sant'Anna, de Chaisemartin-D'Haultfoeuille). Text analysis uses all-MiniLM-L6-v2 for embeddings and SiEBERT for sentiment. Authors note they cannot observe wage effects, overall labor demand changes, or skill composition of new hires. Revised November 2023 version of April 2023 working paper. [Claude classification]: This is the first study examining generative AI deployment at scale in a real workplace. The paper includes a small initial RCT (50 agents, 7 weeks) followed by staggered rollout. Uses multiple robust DiD estimators (Sun-Abraham, Borusyak et al., Callaway-Sant'Anna, de Chaisemartin-D'Haultfoeuille). Text analysis uses all-MiniLM-L6-v2 for embeddings and SiEBERT for sentiment. Authors note they cannot observe wage effects, overall labor demand changes, or skill composition of new hires. Revised November 2023 version of April 2023 working paper. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated. [Claude classification]: This is a foundational conceptual paper on the productivity paradox, not an empirical study of AI effects on workers. The paper proposes four explanations (false hopes, mismeasurement, concentrated distribution, implementation lags) and argues implementation lags are most compelling. Includes formal model showing how unmeasured intangible capital creates J-curve mismeasurement pattern in TFP. Uses historical analogies to electricity and earlier IT adoption. The regression analysis simply shows past productivity growth does not predict future growth (R-squared values of 0.009-0.03). Contains illustrative back-of-envelope calculations for autonomous vehicles (1.7% labor productivity gain if reducing drivers from 3.5M to 1.5M) and call centers (1% productivity gain from 60% automation of 2.2M workers). The formal model (equations 1-5) demonstrates mismeasurement effect: when growth rate of unmeasured capital investment exceeds growth rate of unmeasured capital stock, TFP is underestimated; when reversed, TFP is overestimated.