Essays in economic theory and AI

This thesis explores the intersection of Artificial Intelligence and Economic Theory, focusing on two complementary directions. First, I examine how insights from economics, particularly social choice theory, can inform the development of AI systems. Large language models (LLMs) are trained using reinforcement learning from human feedback (RLHF), a process designed to align them with human preferences. However, in pluralistic societies, human values are diverse and conflicting. This raises a fundamental question: what does it mean to align an AI system with heterogeneous human values? I argue that this question can be analyzed through the lens of social choice theory. Current RLHF pipelines rely on aggregation mechanisms that lack desirable theoretical properties established in the social choice literature. As an alternative, I propose multiple frameworks grounded in social choice theory and economic theory that offer more principled approaches to preference aggregation in AI alignment. Second, I address the reverse question: how can deep learning enhance econometric methods? While machine learning has revolutionized prediction tasks, its integration with causal analysis remains theoretically challenging. Standard deep learning techniques, optimized for predictive accuracy, can introduce biases when applied to causal questions. I examine several limitations of current approaches: the challenges posed by overparameterization, theoretical and experimental issues related to early stopping, the application of Double/Debiased Machine Learning (DML) methods, and the problematic presence of measurement error in learned embeddings. Through both theoretical analysis and empirical investigation, I demonstrate how these issues can compromise causal inference and propose solutions that better integrate machine learning tools. Together, these two lines of work establish a bidirectional exchange between AI and Economics. Economic theory provides rigorous analytical tools for resolving open questions in AI alignment. Conversely, deep learning contributes powerful new methodological tools to empirical economics, expanding the toolkit available for causal inference. This thesis aims to bridge both domains, offering new theoretical insights and practical solutions at their intersection. By drawing on the strengths of each field, this work contributes to both the development of more aligned AI systems and the advancement of empirical methods in economics.