Privacy in Personalized AI Is a System Property, Not Just a Model Property
Privacy in personalized AI must be viewed as a system-level property, not just a model property. Personalized systems expand privacy risks because they accumulate and reuse user information across interactions and components over time. The paper identifies four interconnected leakage channels—data access, inference, behavior, and composition—and proposes four audit requirements to evaluate these complex risks comprehensively.