Sojung Yoon

The rapid diffusion of generative artificial intelligence (GenAI) is reshaping organizational practices and labor market dynamics, yet its implications for skill demand remain unclear. While prior studies have examined how GenAI technologies reshape employment opportunities, little guidance is given on how workers should respond to shifts in skillsets expected in operating environments characterized by these AI tools. To this end, we examine whether GenAI adoption changes firms’ demand for technical, evaluative, and human skills. We construct a firm-quarter panel of U.S. public firms from 2021 to 2025 by integrating 10-K filings and conference call transcripts. Using large language models, we classify firms as GenAI adopters (i.e., firms that adopt GenAI), Traditional AI-only adopters (i.e., firms that adopt traditional AI but not GenAI), or Non-adopters (i.e., firms that adopt neither). We combine this classification with 113 million job postings and measure firm-level skill demand along the extensive and intensive margins. Analyses show that GenAI adoption increases demand for evaluative and human skills, both by expanding the share of postings requiring these skills and by increasing the number of skills required within postings. Additional analyses show that firms increasingly bundle evaluative and human skills with technical skills, and these bundled requirements are associated with wage premiums. Moreover, these effects are more salient for entry-level positions compared to mid/senior-level positions. These findings suggest that GenAI reconfigures workforce demand toward broader categories of skills beyond technical capabilities. This study contributes to research on AI and future of work and offers implications for workforce development.

The rapid integration of artificial intelligence into the workforce, particularly in the gig economy, presents both opportunities and challenges. Algorithmic control is often used to align individual worker behaviors with organizational objectives. While algorithmic control facilitates efficient management of workers, it also leads to intrusive exertion of control, also known as the “algorithm-as-boss” phenomenon. In this study, we attempt to understand the tradeoffs and outcomes of different algorithmic control configurations for gig workers and gig platform. Partnering with a major delivery labor union, we run a randomized field experiment involving 130 gig workers who are randomly assigned to three conditions: tight algorithmic control (i.e., no option to decline the AI-curated recommendation), loose algorithmic control (i.e., a choice to decline the AI-curated recommendation one at a time but without the ability to revisit them), and no control (i.e., free to view all AI-curated recommendations and choose a task). We analyzed the impact of different algorithmic control configurations on outcomes related to the platform’s operational efficiency (i.e., hourly throughput) and workers’ compensation (i.e., hourly profit per kilometer). We also conducted a post-experiment survey to measure workers’ perceived autonomy and self-efficacy. Our study reveals that workers under loose algorithmic control earned significantly higher profits and reported greater perceived autonomy and self-efficacy compared to those under tight algorithmic control, while maintaining comparable levels of operational efficiency. Heterogeneity analyses further uncover the mechanisms behind these effects and boundary conditions. Overall, results suggest that a nuanced approach to algorithmic control is needed in managing gig workers.

While organizations rapidly adopt artificial intelligence (AI) to enhance workforce productivity, many remain at an early stage of deployment, leaving the sources of reported performance gains ambiguous. We address this ambiguity by decoupling the technology’s objective capabilities from workers' subjective interpretations. Drawing on expectancy theory, we examine whether and how AI-generated predictions, regardless of their accuracy, function as performance expectations that shape worker performance. Specifically, we analyze task sequences to determine whether workers treat algorithmic outputs as isolated informational signals or interpret them sequentially as socially meaningful performance expectations. In collaboration with a hospital, we conducted a randomized field experiment involving 28 workers completing 1,069 task sequences. Workers received randomly assigned, statistically uninformative performance predictions framed as AI-generated, indicating either favorable (early/on-time) or unfavorable (late) completion. We then examined how the initial prediction conditioned the behavioral impact of the subsequent prediction on productivity. Our findings reveal asymmetric sequential pattern: while a subsequent unfavorable prediction had no significant effect following an initial favorable prediction, consecutive unfavorable predictions reduced task completion time by 2.74 minutes. Mechanism analyses show that this improvement emerges when workers interpret the predictions as evaluative judgments of their own capabilities rather than as objective information about task conditions. This pattern is consistent with the underdog effect, whereby individuals respond to perceived underestimation by exerting greater effort to disconfirm it. Overall, our findings suggest that AI can shape worker performance not only through its technical capabilities but also through the socially meaningful expectations its outputs create across repeated interactions.