Advancing AI for Incentive-Aligned Systems

Abstract

Artificial intelligence increasingly shapes economic interactions and information mediation. It serves both as a computational tool for designing markets and mechanisms, and as an intermediary in deployed systems such as recommender platforms and large language models that aggregate, filter, and present information. In each of these settings, performance alone is not enough: the surrounding system must also respect incentives to be efficient, effective and safe. This dissertation studies how to advance AI for incentive-aligned systems, developing methods at the intersection of deep learning, mechanism design, and large language models. The first part of the dissertation develops computational tools for incentive-aware market and mechanism design, spanning from cornerstone problems in economic theory—such as optimal multi-item, multi-bidder auction design—to emerging challenges in new economies, such as data market design. These are foundational questions in the modern economy, yet they remain analytically intractable, highlighting a gap between what matters in both theory and practice and what current methods can solve. This dissertation shows how deep learning can serve as a structured tool for bridging this gap: neural network parameterizations, menu-based representations, and optimization-based post-processing enable the search for expressive mechanism designs while preserving dominant-strategy incentive compatibility and feasibility properties. Chapter 4 contributes a dual perspective on learned auction design, pushing the frontier of what can be said about optimality. By producing revenue upper bounds, this work certifies how close learned mechanisms are to optimal, shifting the agenda from learned mechanisms to learned and certified mechanisms. The second part of the dissertation turns from AI for mechanism design to mechanism design for AI systems. It studies multi-source LLM summarization and bid-aware generative recommendation: settings in which AI systems function as economic intermediaries, aggregating testimony and generating or ranking outcomes for strategic agents. This dissertation shows that these systems should be designed with explicit attention to incentives, strategic behavior, and welfare, rather than treated as prediction or generation modules in isolation. Taken together, this dissertation advances AI for incentive-aligned systems along two complementary directions. Machine learning expands the range of economic environments in which principled design is computationally feasible, while mechanism design provides the structure needed to make AI systems strategically sound. By developing methods along both directions, the dissertation contributes toward a broader agenda of building AI systems that are both capable and incentive-aligned in the strategic environments they inhabit.