UC San Diego

Artificial intelligence (AI) and machine learning (ML) promise to reduce human oversight across a wide range of tasks–from forecasting and recommendation to text summarization and multimodal content generation. Yet this reliance on data-driven learning introduces a key vulnerability: the implicit reproduction of biases in training data, which can lead to misalignments between system behavior and its intended goals. As AI takes on greater responsibility in high- impact domains, assessing the alignment between system functionality and aspirational values becomes increasingly urgent. This dissertation investigates these alignment challenges through an interdisciplinary lens, combining philosophical inquiry into the goals of AI systems (e.g., personalization, fairness) with sociotechnical analyses of the mechanisms used to achieve them, such as collaborative filtering and data practices. Across these studies, user agency emerges as a central theme – both as a normative ideal and a practical measure of system success. The work culminates in policy and measurement recommendations aimed at ensuring alignment between the intended purposes of AI systems and their real-world societal impacts so that deployment practices genuinely support equitable, meaningful outcomes for all users.