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

Labor Demand in the Age of Generative AI: Early Evidence from the U.S. Job Posting Data

Liu, Wang, Yu

2025World Bank Policy Research Working Paper Series

Observational labor marketCausal

LLM / Generative AIJunior / entry-levelAugmentation vs. substitutionGeneral automation

View Repository DOI: 10.1596/1813-9450-11263

Summary

Liu, Wang, and Yu use difference-in-differences and event study designs with 285 million U.S. job postings (2018Q1-2025Q2) to estimate the causal impact of ChatGPT's November 2022 release on labor demand across occupations with varying AI substitution vulnerability, conditional on similar GenAI exposure levels.

Main Finding

Job postings for occupations with above-median AI substitution scores declined by an average of 12% relative to those with below-median scores following ChatGPT's launch, with effects intensifying from 6% in the first year to 18% by the third year; entry-level positions requiring neither advanced degrees (18%) nor extensive experience (20%) experienced the steepest declines, as did administrative support (40%) and professional services (30%) sectors.

Primary Datasets

Lightcast (formerly Burning Glass)

Labor outcomes

Lightcast online job postings (285 million postings, 2018Q1-2025Q2)

Secondary Datasets

None

Key Methods: Difference-in-differences and event study designs exploiting ChatGPT's November 2022 release as exogenous shock; compares occupations with above- vs. below-median AI substitution scores conditional on similar GenAI exposure levels; state-industry-occupation-quarter panel data with multiple fixed effects
Sample Period: 2018Q1-2025Q2
Geographic Coverage: US
Sample Size: 285 million job postings aggregated into 6.8 million state-industry-occupation-quarter cells
Level of Analysis: Occupation, Industry, Region
Occupation Classification: None
Industry Classification: None
Replication Package: Yes

Notes

World Bank Policy Research Working Paper 11263. Finds 12% average decline in postings for high-AI-substitution occupations; 18% by third year. Entry-level positions hit hardest (18-20%). Administrative support (-40%) and professional services (-30%) most affected. [Claude classification]: Paper uses GPT-4o as a methodological tool to annotate empathy (llmPromptingAnnotation=true), but the paper itself studies LLM chatbots (GPT-3.5 and GPT-4) as the AI technology of interest. Experiment involves random assignment to human or chatbot conversational partner. Uses multiple empathy measurement approaches: self-reports, LLM annotations, trained prediction models, and pre-trained empathy models. Sample includes 155 conversations total (96 human-chatbot, 59 human-human). [Claude classification]: Paper uses GPT-4o as a methodological tool to annotate empathy (llmPromptingAnnotation=true), but the paper itself studies LLM chatbots (GPT-3.5 and GPT-4) as the AI technology of interest. Experiment involves random assignment to human or chatbot conversational partner. Uses multiple empathy measurement approaches: self-reports, LLM annotations, trained prediction models, and pre-trained empathy models. Sample includes 155 conversations total (96 human-chatbot, 59 human-human). [Claude classification]: Paper uses GPT-4o as a methodological tool to annotate empathy (llmPromptingAnnotation=true), but the paper itself studies LLM chatbots (GPT-3.5 and GPT-4) as the AI technology of interest. Experiment involves random assignment to human or chatbot conversational partner. Uses multiple empathy measurement approaches: self-reports, LLM annotations, trained prediction models, and pre-trained empathy models. Sample includes 155 conversations total (96 human-chatbot, 59 human-human). [Claude classification]: Paper uses GPT-4o as a methodological tool to annotate empathy (llmPromptingAnnotation=true), but the paper itself studies LLM chatbots (GPT-3.5 and GPT-4) as the AI technology of interest. Experiment involves random assignment to human or chatbot conversational partner. Uses multiple empathy measurement approaches: self-reports, LLM annotations, trained prediction models, and pre-trained empathy models. Sample includes 155 conversations total (96 human-chatbot, 59 human-human). [Claude classification]: Paper uses GPT-4o as a methodological tool to annotate empathy (llmPromptingAnnotation=true), but the paper itself studies LLM chatbots (GPT-3.5 and GPT-4) as the AI technology of interest. Experiment involves random assignment to human or chatbot conversational partner. Uses multiple empathy measurement approaches: self-reports, LLM annotations, trained prediction models, and pre-trained empathy models. Sample includes 155 conversations total (96 human-chatbot, 59 human-human). [Claude classification]: Paper uses GPT-4o as a methodological tool to annotate empathy (llmPromptingAnnotation=true), but the paper itself studies LLM chatbots (GPT-3.5 and GPT-4) as the AI technology of interest. Experiment involves random assignment to human or chatbot conversational partner. Uses multiple empathy measurement approaches: self-reports, LLM annotations, trained prediction models, and pre-trained empathy models. Sample includes 155 conversations total (96 human-chatbot, 59 human-human). [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly. [Claude classification]: World Bank Policy Research Working Paper 11263. The paper uses ChatGPT's November 2022 release as an exogenous shock (natural experiment) and employs difference-in-differences methodology. Key innovation is using two-dimensional framework: GenAI exposure (technical applicability) and AI-substitution vulnerability (practical displacement likelihood based on occupational characteristics). Uses inverse hyperbolic sine transformation for job posting counts. Conducts 400 placebo tests for robustness. Data aggregated to state-industry-occupation-quarter level to create balanced panel of 6.8 million cells. Standard errors clustered at occupation-quarter level. Verified reproducibility package available. The paper studies GenAI's impact on job postings (a forward-looking labor demand indicator), not employment levels directly.