UC Berkeley

Technical fixes for AI harm are usually framed through an individual‑control lens: give people notice, let them opt out, erase their row, and the problem is solved. Building on relational theories of data governance, this dissertation probes the limits of that premise across the full machine‑learning pipeline. It traces three tightly coupled ‘critical cases’ that sit at successive system layers. At the input level, a forensic audit of privacy‑branded synthetic‑data workflows shows how nominal anonymity can still magnify racial mis‑representation. At the model level, a scenario‑based study of machine unlearning reveals that participants' wish to be erased is conditioned by collective‑risk calculations, not by autonomy alone. At the evaluation level, a meta‑analysis of contextual benchmarks demonstrates that large language models which pass instruction‑following tests routinely fail when placed in norm‑conflict or long‑context settings.Across these settings the findings converge on four recurring failure modes. First, individual row‑level privacy gains often intensify group harms, exposing relational blind‑spots. Second, generative and unlearning tools rupture data lineage, undermining regulatory oversight that depends on provenance. Third, performance that looks robust in context‑free scorecards collapses under realistic framing, revealing deep contextual brittleness. Finally, the operational cost of verifying compliance with individual‑control mechanisms creates enforcement frictions that outstrip regulatory capacity.Methodologically, the work illustrates how combining dataset forensics, interpretive user research, and quantitative stress‑testing produces a multi‑level socio‑technical diagnosis unavailable to any single approach. In doing so it heeds recent calls for a richer ‘evaluation science’ in AI ethics and charts a research agenda toward lineage‑aware tooling, context‑sensitive benchmarks, and governance structures capable of addressing the relational nature of machine‑learning harm.