The Invisible Hand: Characterizing Generative AI Adoption and its Effects on An Online Freelancing Market

TL;DR

The paper investigates how Generative AI (GenAI), led by ChatGPT, reshapes an online freelancing market using a large-scale dataset from Freelancer.com (1.8M+ posts, 3.8M+ users). It employs a mix of Difference-in-Differences, bid-network analyses, topic modeling (BERTopic), and skill-prediction classifiers to quantify GenAI adoption, its effects on demand and bidding, and the evolving skill landscape. Key findings show GenAI-related jobs surge after ChatGPT’s release, with ChatGPT dominating GenAI postings and commanding higher budgets and bids; adopters occupy central network positions and the connecting fabric between GPT-related and non-GPT-related workers strengthens post-release. The study also demonstrates that ChatGPT use falls into two main job categories (software integration and content generation) with distinct economic signals, and that pre-release features can predict which skills will remain relevant, enabling early, data-driven decision making for job seekers and platforms. Overall, GenAI appears to catalyze skill transformation and integration within the freelancing ecosystem rather than merely displacing human labor, with implications for platform design and upskilling initiatives.

Abstract

Since the COVID-19 pandemic, freelancing platforms have experienced significant growth in both worker registrations and job postings. However, the rise of generative AI (GenAI) technologies has raised questions about how it affect the job posting in freelancer market. Despite growing discussions, there is limited empirical research on the GenAI adoption and its effect on job demand and worker engagement. We present a large-scale analysis of Freelancer.com, utilizing over 1.8 million job posts and 3.8 million users. We investigate the emergence of jobs with the adoption of GenAI and identify leading position of ChatGPT in the freelancing market. With a focus on ChatGPT related jobs, we inspect their specific skill requirements, and the tasks that workers are asked to perform. Our findings provide insights into the evolving landscape of freelancing in the age of AI, offering a comprehensive profile of GenAI's effects on employment, skills, and user behaviors in freelancing market.

Figures (10)

• Figure 1: Comparison of minimum/maximum budget, bid count, and average bid price between GenAI-related job posts and non-GenAI-related job posts. The statistical significance is reported by the Mann-Whitney U test; ****: $\bm{p<0.001}$.
• Figure 2: Temporal distribution of GenAI job post count.
• Figure 3: Time series of monthly network modularity and proportion of inter-community edges between users who have ever submitted or bid any GenAI-related job posts and those who haven't. The black dash line denotes ChatGPT's release
• Figure 4: Comparison of minimum/maximum budget, bid count, and average bid price between GPT-related job posts within diverse targets. The statistical significance is reported by the Kruskal-Wallis test with Dunn's post-hoc test; ****: $\bm{p<0.0001}$, ***: $\bm{p<0.001}$ **: $\bm{p<0.01}$, *: $\bm{p<0.05}$.
• Figure 5: Classifiers performance in defining whether job skills are relevant or irrelevant to ChatGPT market for every 30 days after ChatGPT's release.