Economics of Human-AI Interaction and AI-AI Interaction

Abstract

Artificial intelligence (AI) has evolved from a specialized tool into a central participant in economic life: it performs work, prices products, and decides what information people see. Economic outcomes are therefore no longer shaped by human decisions alone; they depend on how humans interact with AI. And as more market participants delegate decisions to algorithms, outcomes also depend on how AI systems interact with one another. This dissertation studies the economics of human-AI interaction and AI-AI interaction through three distinct yet interconnected essays. First, I study humans' strategic responses to the arrival of a powerful new AI that affects both sides of a market. I leverage comprehensive data from a leading online labor platform and a difference-in-differences design around the launch of ChatGPT. I find that LLM-based generative AI substantially displaces labor demand in exposed submarkets, while labor supply contracts more slowly, so competition intensifies. Surprisingly, rather than retreating from AI-exposed work, freelancers strategically transition toward programming-intensive jobs, the very domain where generative AI excels, because the technology also lowers the cost of acquiring new skills. This adaptation is driven primarily by high-skilled workers and carries significant consequences for earnings. Second, I examine what happens when many autonomous AI agents compete within a single market and a personalized AI intermediary stands between them and consumers. I combine a structural model of consumer search estimated from real-world data with large-scale simulations of reinforcement-learning pricing algorithms. The objective of the platform's recommender system steers algorithmic competition: a revenue-maximizing design intensifies collusion, whereas a utility-maximizing design promotes competition. More surprisingly, enlarging the recommendation set does not consistently serve the platform's goals. This "more is less" effect arises because a larger set weakens the recommender system's ability to steer pricing algorithms through the rewards it controls. Third, I investigate what happens when a platform deliberately lowers the personalization level of its AI, a choice increasingly compelled by privacy regulations and computational costs. Through a large-scale randomized field experiment on a major social media platform, I find that depersonalizing search makes search measurably harder within the treated channel: users browse more content per query yet click less ordinary content, while their clicking on paid content barely moves. Surprisingly, the effects spill over to the untouched recommendation channel with the opposite sign. Users lower the standard they carry across channels, browse more, and click more ordinary content there. Paid content in the recommendation channel responds differently: its click-through rate is unchanged on average but falls among the platform's most active users. Taken together, the three essays follow AI across the life of a digital market: a powerful new AI enters and humans strategically adapt, autonomous AI agents fill the market and interact under the platform's mediation, and the platform in turn calibrates how much algorithmic mediation to supply. Ultimately, this dissertation serves as a foundation for future research on the evolving economics of AI, emphasizing the need for ongoing investigations that keep pace with markets in which humans and algorithms increasingly decide together.