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

Winners and Losers of Generative AI: Early Evidence of Shifts in Freelancer Demand

Teutloff, Einsiedler, Kässi, Braesemann, Mishkin, del Rio-Chanona

2025Journal of Economic Behavior and Organization15 citations
Observational labor marketCausal
LLM / Generative AIPlatforms / gig economyJunior / entry-levelAugmentation vs. substitutionWriting / content
DOI: 10.1016/j.jebo.2024.106845
Abstract

We examine how ChatGPT has changed the demand for freelancers in jobs where generative AI tools can act as substitutes or complements to human labor. Using BERTopic we partition job postings from a leading online freelancing platform into 116 fine-grained skill clusters and with GPT-4o we classify them as substitutable, complementary or unaffected by LLMs. Our analysis reveals that labor demand increased after the launch of ChatGPT, but only in skill clusters that were complementary to or unaffected by the AI tool. In contrast, demand for substitutable skills, such as writing and translation, decreased by 20–50% relative to the counterfactual trend, with the sharpest decline observed for short-term (1-3 week) jobs. Within complementary skill clusters, the results are mixed: demand for machine learning programming grew by 24%, and demand for AI-powered chatbot development nearly tripled, while demand for novice workers declined in general. This result suggests a shift toward more specialized expertise for freelancers rather than uniform growth across all complementary areas. • LLM-based clustering shows ∼ 10% of job postings are substitutable by GenAI. • GenAI cut demand for substitutable skill clusters by up to 50% in short-term roles. • Demand for some complementary AI clusters like machine learning increased. • Demand for novice workers in complementary roles declined. • Aggregate freelance demand did not decrease after the ChatGPT launch.

Summary

Teutloff et al. use BERTopic clustering and GPT-4o classification of 3+ million job postings from an online freelancing platform to study how ChatGPT's launch affected demand across 116 skill clusters classified as substitutable, complementary, or unaffected by LLMs, employing difference-in-differences analysis from 2021-2023.

Main Finding

Demand for substitutable skill clusters (writing, translation) decreased 24% relative to unaffected clusters after ChatGPT's launch, concentrated in short-term jobs, while complementary clusters showed no aggregate effect but mixed results with AI-powered chatbot demand tripling (+179%) and machine learning increasing 24%, and novice worker demand declining in complementary clusters.

Primary Datasets

Anonymous online freelancing platform job postings data collected via Online Labour Index project API

Secondary Datasets

None

Key Methods
Difference-in-differences using unaffected skill clusters as control group; BERTopic transformer-based topic modeling to create skill clusters from job posting skill tags; GPT-4o prompt engineering with chain-of-thought reasoning to classify clusters into substitutable/complementary/unaffected categories; cluster-level heterogeneity analysis by project duration and worker experience.
Sample Period
2021-2023
Geographic Coverage
Global online labor platform (US-based platform with workers from India, Pakistan, Philippines, Eastern Europe)
Sample Size
Over 3 million job postings across 116 skill clusters; 11,960 cluster-week observations in main analysis
Level of Analysis
Occupation, Task
Occupation Classification
None (custom skill clusters created via BERTopic)
Industry Classification
None
Notes
Journal of Economic Behavior & Organization, p. 106845 [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity. [Claude classification]: Uses GPT-4o as methodological tool for classifying skill clusters; treats ChatGPT launch (November 30, 2022) as natural experiment; BERTopic clustering uses sentence transformer all-MiniLM-L6-v2 with UMAP dimensionality reduction and HDBSCAN clustering; standard errors clustered at skill-cluster level; includes randomization inference robustness checks; platform requested anonymity.