Tan, YongMa, Lijia2025-08-012025-08-012025Ma_washington_0250E_28030.pdfhttps://hdl.handle.net/1773/53432Thesis (Ph.D.)--University of Washington, 2025This thesis offers a thorough analysis of multiple scenarios of human-generative AI interaction, including large-language-model-based search engine optimizations, heterogeneous adoption of generative AI, and the impact of generative AI tools on work. First, I investigate how LLM-powered chat search, exemplified by Bing Chat, selects its information sources. By comparing thousands of Bing Chat citations with traditional search results, I show that it systematically prefers text that is highly readable, well-structured, and of low perplexity. I replicate these findings using a GPT-4 RAG API, demonstrating that these preferences arise from the language models themselves rather than from bespoke engineering. I also find that RAG-cited sites are more homogeneous than those surfaced by classic search algorithms. These insights highlight the distinctive sourcing behavior and economic implications of chat-based search engines. Second, I examine how recent generative AI advances such as ChatGPT create a digital divide. Specifically, I distinguish between a learning divide, referring to differences in how quickly users update beliefs about ChatGPT’s value, and a utility divide, referring to variation in actual per-use benefit. By estimating a Bayesian learning model on six months of clickstream data, I find that lower-educated and non-white users gain greater utility per use but update their beliefs more slowly, whereas younger, male, and IT-background users excel on both fronts. I also identify a belief trap in which persistent underuse results from underestimated utility, and I show that targeted training can reduce this outcome divide. Third, I investigate how daily engagement with large-language-model-based generative AI tools reshapes the structure and focus of work. Drawing on clickstream records, I conduct an empirical analysis of the effect of using generative AI tools on working duration and distraction levels. I find that engagement with these tools significantly increases working hours while simultaneously reducing distraction rates. These findings offer practical guidance for employers and individual employees considering the adoption of generative AI tools for work. Taken together, this thesis lays a solid foundation for future research on human-generative AI interactions and on the broader challenges of AI alignment.application/pdfen-USCC BY-NCBusiness administrationBusiness administrationThesis