A Design Space of Notification Delivery for AI Agents' Privacy Risks

Users value their privacy but often experience privacy fatigue. AI agent sharpens this challenge, as its autonomous and invisible workflows increasingly place users out of the execution loop. While existing research has explored privacy safeguards within human-LLM interaction, there still lacks systematic investigation of privacy notification delivery in AI agents. To address this gap, we framed notification delivery as a selection problem: when should systems bring users back into the loop, and how intensive should that oversight be? Then we organized it around two dimensions: when to notify users and how intensive that notification should be. To operationalize this design space, we proposed methods to computationally identify privacy-relevant transitions and presented a case study for our design space. We finally outline challenges for evaluation in real workflows, adaptive notification delivery strategy, and privacy cases involving multiple people.