The Privacy Guardian Agent: Towards Trustworthy AI Privacy Agents

Vincent Freiberg's position paper, presented at the CHI26 Workshop, tackles the broken 'notice and consent' paradigm where manipulative consent dialogues overwhelm users. The proposed Privacy Guardian Agent offers a middle ground between full automation and manual control: it uses user profiles and contextual awareness to automate routine consent decisions, but escalates unclear or high-risk cases to the user. This human-in-the-loop approach ensures transparency by making the agent's reasoning reviewable, and for problematic sites, it alerts users and suggests alternatives. The goal is to reduce consent fatigue while preserving meaningful user autonomy and trust.

The agent's design addresses key limitations of current LLM-based tools, which either require active user engagement or risk hallucinations and opaque decisions. By automating low-stakes consent choices and escalating only when necessary, the Privacy Guardian Agent aims to provide a scalable, trustworthy solution for privacy management. The paper emphasizes that even with minimal consent, the agent can alert users to problematic sites and suggest switching to alternatives, ensuring users maintain agency over their data.