UC Irvine

Human-in-the-loop (HITL) systems, from smart environments to conversational AI assistants, increasingly couple human preferences, behavior, and feedback into the computational loop. As these systems mediate decisions that affect people, three human-centered properties cannot be separated from performance: whether the system treats people fairly, whether it protects their privacy equitably, and whether it aligns with what individuals and groups actually want. This dissertation develops computational frameworks that build fairness, privacy, and personalization directly into HITL systems, covering cyber-physical systems and large language models (LLMs). The dissertation advances four contributions. First, it formalizes fairness in sequential, human-in-the-loop decision-making. FinA grounds fairness in the psychology of loss aversion, defining and minimizing an "adverse effect" through optimization, while FAIRO recasts the same goal as temporally abstracted sub-goals solved with the Options reinforcement-learning framework; together they improve perceived fairness by up to 67% over the state of the art across smart-home, water-allocation, and personalized-learning applications. Second, it treats privacy as an equity problem. FinP shows that privacy risk in federated learning is inequitably distributed, falling hardest on outlier clients, and enforces fairness-in-privacy through Hessian-ranked, Lipschitz-based client regularization and risk-aware server aggregation, reducing vulnerability disparities by up to 57% with negligible utility cost. Third, it studies personalization at the level of the individual. PACIFIC demonstrates that stable Big-Five personality traits are a more reliable latent signal for personalization than raw conversation history, and introduces a persona-aware retriever that raises label-free preference-prediction accuracy without trait annotations. Fourth, it studies pluralistic AI alignment across groups: PluralLLM learns group preferences via federated learning as a privacy-preserving reward model, and a systematic evaluation of federated preference aggregation shows that adaptive, history-aware weighting improves fairness across groups while preserving alignment quality. The four topics proceed in this order: fair decision making across a group over time; equitable distribution of information-leakage risk; personalization to an individual through stable personality traits; and alignment of generative models with the values of many groups. Being fair to people as a set is a different problem from being aligned to any one of them, which is why the work moves from group equity to individual personalization and then to group alignment. In all four, human experience, risk, and preference are treated as quantities to be measured, distributed, and reasoned over. The results show that fairness, privacy, and alignment can be achieved with minimal or no utility cost.