AI, Privacy, and the Architecture of Inferential Harm
AI, Privacy, and the Hidden Architecture of Harm from Inference
AI summary
A series of student essays from Harvard's Berkman Klein Center argues that privacy regulation should go beyond governing data to governing the inferential capabilities that models build on top of data. The essays cite 2024 research: prompting OpenAI's GPT-4 with a single Reddit post mentioning a hook turn correctly inferred that the user was in Melbourne, Australia; they also cite an FTC 2021 warning that "neutral" AI systems can produce discriminatory outcomes. The authors propose constraining inferential capabilities by expanding the definition of covered data, capability-based governance, and enforceable impact assessments.