PrivacyReasoner: Can LLM Emulate a Human-like Privacy Mind?

TL;DR

PrivacyReasoner is an LLM-based agent that models human-like privacy reasoning by detecting subtle cues, reconstructing individual privacy minds from online comments, and dynamically applying relevant privacy beliefs.

Contribution

It introduces a novel architecture that combines privacy cue detection, user privacy mind reconstruction, and contextual filtering, advancing understanding of privacy reasoning in LLMs.

Findings

Outperforms baselines in predicting individual privacy concerns

Successfully generalizes across domains like AI, e-commerce, and healthcare

Uses an LLM-as-a-Judge evaluator calibrated against privacy taxonomy

Abstract

Prior work on LLM-based privacy focuses on norm judgment over synthetic vignettes, rather than how people think about a specific data practice and formulate their opinions. We address this gap by designing PrivacyReasoner, an agent architecture grounded in three key ideas: (1) LLMs can detect subtle privacy cues in natural language and role-play human characteristics; (2) a user's ``privacy mind'' can be reconstructed from their real-world online comment history, distilling experiences, personality, and cultural orientations; and (3) a contextual filter can dynamically activate relevant privacy beliefs based on the contexts in a scenario. We evaluate PrivacyReasoner on real-world privacy discussions from Hacker News, using an LLM-as-a-Judge evaluator calibrated against an established privacy concern taxonomy to quantify reasoning faithfulness. PrivacyReasoner significantly outperforms baselines in predicting individual privacy concerns and generalizes across different domains, such as AI, e-commerce, and healthcare.