Large Language Models in Strategic Environments: Behavioral Proxies, Human Delegates, and Agents

Abstract

Large language models and generative AI are moving into economic life at an unusual pace. Today, generative AI is autonomously writing and executing code, negotiating procurement contracts, and placing bids in online advertising auctions. As these AI systems diffuse into the economy, we need methods for understanding how they make decisions, how their behavior differs from human behavior, how those differences can be measured systematically, and what they may imply for markets and other institutions. These questions motivate this dissertation. This dissertation studies language models in three roles: as behavioral proxies for people, as delegates acting on behalf of humans ,and as strategic agents in their own right. The first part develops a methodology for structured experimentation with large language model (LLM) agents as decision-making proxies for people. Chapter~2 places LLMs in controlled social and economic environments, varies exogenous factors systematically, and compares the resulting outcomes and causal relations with findings from human subject experiments and empirical studies. Chapter~3 takes up the question of LLMs as proxies for human behavior in auction settings, studying how model behavior changes under different incentives and information structures and comparing this behavior with human subject experiment and theoretical benchmarks. Together, these chapters provide a general procedure for identifying where LLM behavior matches or departs from established economic findings. The second part of this dissertation studies language models not as behavioral proxies but for people but as autonomous agents and develops a framework with which to understand their cognitive constraints. Chapter~4 asks whether the theory of cognitive simplicity, originally developed for human bounded rationality, i.e. cognitive scaffolding and different kinds of reasoning patterns and failures, also describes language-model reasoning. Across four model families and two settings---auctions and two-sided matching via deferred acceptance---obviously strategy-proof implementations substantially improve strategic accuracy. Using a cognitive-scaffold methodology that intervenes along the axes of contingent reasoning, forward planning, and belief formation, the chapter shows that scaffolding contingent reasoning and explaining why truthful play is safe are productive, while prompting for forward planning or belief formation is counterproductive. From this, we see that today's LLMs are cognitively-bounded agents whose behavior can be measured in different settings and with different kinds of cognitive scaffolds rather than inferred from abstract rules alone. The third part of this dissertation studies direct human-AI interaction in strategic environments. Chapter~5 develops a multi-agent bargaining environment that supports systematic comparison across human players, LLM agents, and bespoke Bayesian algorithm. The environment admits a Pareto-optimal welfare benchmark, which makes it possible to compare agent behavior (proposing, accepting and rejecting) as well as overall performance within a simple strategic setting. The first study compares three different agent populations under identical rules: humans, LLM agents, and purpose-built Bayesian agents. Humans and LLM agents achieve comparable welfare through distinct behavioral processes, demonstrating that performance parity in regard to performance metrics can mask procedural differences and with distributional consequences. Chapter~6 reports on a second study in which human participants can choose among three modes of AI assistance: an advisor that recommends, a coach that critiques, or a delegate that acts autonomously. I show that humans with access to a LLM delegate agent achieve the highest group and individual welfare, but they strongly prefer the AI advisor for their future usage, showing a preference-performance misalignment. Taken together, this dissertation offers a methodological toolkit for studying LLMs in social and economic setting. It shows how to study the differences between LLM agent behavior and human behavior, how to relate LLM agent behavior to the structure of mechanisms and the presence of cognitive scaffolds, and how to evaluate the consequences of deploying LLM agents in markets and other strategic environments.