The Absent Data Subject: Governing Third-Party Privacy in Personalized and Agentic Language-Model Assistants

Privacy protections for language-model assistants center almost entirely on the account holder. That person accepts the terms, toggles memory, and clears history. Yet people rarely talk only about themselves. They describe a friend's diagnosis, name a coworker, or paste a document that carries a stranger's phone number. None of these third parties agreed to anything. None can see what the system kept about them, and none can ask for its removal. We argue that the consent-and-control model built for first-party data is the wrong primitive for protecting them, because consent assumes the data subject is present in the interaction, and third parties are structurally absent. Two features of modern assistants deepen the problem. Persistent memory turns a single mention into a durable, queryable record, and agentic action can send that record onward without further review, while inference can manufacture new and possibly false attributes about someone who will never see them. The paper offers five contributions: a lifecycle model of how third-party data enters and escalates inside an assistant; a contextual-integrity analysis that traces the failure of consent to the gap between the information subject and the sender; a taxonomy of the resulting harms; a governance framework that assigns concrete controls to users, providers, and policymakers under two defaults, minimization and contextual integrity; and an audit rubric for measuring deployed assistants against that framework. The aim is to shift protection for these data subjects away from adding dialogs and toward changing defaults.