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

The Unequal Adoption of ChatGPT Exacerbates Existing Inequalities Among Workers

Humlum, Vestergaard

2024Proceedings of the National Academy of Sciences (PNAS)51 citations

Adoption / usageCausal

LLM / Generative AIGenderJunior / entry-levelHuman-AI collaborationTraining / upskilling

View Repository DOI: 10.1073/pnas.2414972121

Abstract

We study the adoption of ChatGPT, the icon of Generative AI, using a large-scale survey linked to comprehensive register data in Denmark. Surveying 18,000 workers from 11 exposed occupations, we document that ChatGPT is widespread, especially among younger and less-experienced workers. However, substantial inequalities have emerged. Women are 16 percentage points less likely to have used the tool for work. Furthermore, despite its potential to lift workers with less expertise, users of ChatGPT earned slightly more already before its arrival, even given their lower tenure. Workers see a substantial productivity potential in ChatGPT but are often hindered by employer restrictions and a perceived need for training.

Summary

Humlum and Vestergaard use a large-scale survey of 18,000 Danish workers linked to administrative registers to study ChatGPT adoption patterns, beliefs about productivity impacts, and barriers to adoption across 11 exposed occupations one year after ChatGPT's launch.

Main Finding

41% of workers in exposed occupations have used ChatGPT for work within one year of launch, but substantial inequalities emerged: women are 16 percentage points less likely to adopt than men, and despite ChatGPT's potential to help less experienced workers, actual users earned slightly more pre-ChatGPT; employer restrictions and perceived training needs are the primary adoption barriers.

Primary Datasets

Survey (18K workers, 11 occupations); Danish register data

Secondary Datasets

O*NET

Other

Eloundou et al. (2023)

AI-focused

O*NET database (job tasks and detailed work activities); Expert assessments from Eloundou et al. (2023) GPT ratings, validated with Danish industry experts

Key Methods: Large-scale survey (N=100,000 invited, 18,000 responses) with embedded randomized information experiment; responses linked to Danish administrative registers; randomized participation incentives to control for selection bias; two-week follow-up survey to measure persistence of treatment effects
Sample Period: Nov 2023-Jan 2024
Geographic Coverage: Denmark
Sample Size: 18,000 valid survey responses from 100,000 invitations across 11 occupations; two-week follow-up survey
Level of Analysis: Individual, Occupation, Task
Occupation Classification: ISCO-08; O*NET job tasks
Industry Classification: None
Replication Package: Yes

Notes

PNAS; women and lower-earning workers less likely to adopt; younger adopt faster; Non-US paper; included for cross-national comparison [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT to generate expert exposure ratings and draft explanations, adapting Eloundou et al. (2023) to Danish O*NET job duties. Replication materials preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate with 86% still employed in target occupations. Strong gender gap (20pp) persists within workplaces and controlling for task mix. Limited cross-task substitution (38% report zero). Women more responsive to information treatment initially but effects dissipate. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT to generate expert exposure ratings and draft explanations, adapting Eloundou et al. (2023) to Danish O*NET job duties. Replication materials preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate with 86% still employed in target occupations. Strong gender gap (20pp) persists within workplaces and controlling for task mix. Limited cross-task substitution (38% report zero). Women more responsive to information treatment initially but effects dissipate. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT to generate expert exposure ratings and draft explanations, adapting Eloundou et al. (2023) to Danish O*NET job duties. Replication materials preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate with 86% still employed in target occupations. Strong gender gap (20pp) persists within workplaces and controlling for task mix. Limited cross-task substitution (38% report zero). Women more responsive to information treatment initially but effects dissipate. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT to generate expert exposure ratings and draft explanations, adapting Eloundou et al. (2023) to Danish O*NET job duties. Replication materials preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate with 86% still employed in target occupations. Strong gender gap (20pp) persists within workplaces and controlling for task mix. Limited cross-task substitution (38% report zero). Women more responsive to information treatment initially but effects dissipate. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT to generate expert exposure ratings and draft explanations, adapting Eloundou et al. (2023) to Danish O*NET job duties. Replication materials preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate with 86% still employed in target occupations. Strong gender gap (20pp) persists within workplaces and controlling for task mix. Limited cross-task substitution (38% report zero). Women more responsive to information treatment initially but effects dissipate. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT to generate expert exposure ratings and draft explanations, adapting Eloundou et al. (2023) to Danish O*NET job duties. Replication materials preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate with 86% still employed in target occupations. Strong gender gap (20pp) persists within workplaces and controlling for task mix. Limited cross-task substitution (38% report zero). Women more responsive to information treatment initially but effects dissipate. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service. [Claude classification]: Published in PNAS 2024 (DOI: 10.1073/pnas.2414972121). Also circulated as BFI Working Paper 2024-50. Uses GPT prompting to replicate and adapt Eloundou et al. (2023) exposure measure to Danish O*NET job duties. Preregistered at AEA-RCT-R-0012527. Randomized participation prizes (1000-10000 DKK) to control for selection following Dutz et al. (2022). Survey achieved 29% response rate (~18,000 valid responses from 100,000 invitations) with 86% still employed in target occupations. Strong gender gap (16pp unconditional, 12pp within-workplace controlling for task mix) persists across all specifications. Information treatment shifted beliefs but did not affect actual usage, consistent with binding adoption frictions. Workers estimate ChatGPT can halve working times in 37% of job tasks for typical worker, 32% for themselves. Limited cross-task substitution: 37% report zero elasticity. Two-week follow-up survey shows treatment effects on beliefs persist but usage effects remain null. Replication materials to be deposited in Open Science Framework; microdata access through Statistics Denmark Research Service.