Text-Based Personas for Simulating User Privacy Decisions

Abstract

The ability to simulate human privacy decisions has significant implications for aligning autonomous agents with individual intent and conducting cost-effective, large-scale privacy-centric user studies. Prior approaches prompt Large Language Models (LLMs) with natural language user statements, data-sharing histories, or demographic attributes to simulate privacy decisions. These approaches, however, fail to balance individual-level accuracy, prompt usability, token efficiency, and population-level representation. We present Narriva, an approach that generates text-based synthetic privacy personas to address these shortcomings. Narriva grounds persona generation in prior user privacy decisions, such as those from large-scale survey datasets, rather than purely relying on demographic stereotypes. It compresses this data into concise, human-readable summaries structured by established privacy theories. Through benchmarking across five diverse datasets, we analyze the characteristics of Narriva's synthetic personas in modeling both individual and population-level privacy preferences. We find that grounding personas in past privacy behaviors achieves up to 88% predictive accuracy (significantly outperforming a non-personalized LLM baseline), and yields an 80-95% reduction in prompt tokens compared to in-context learning with raw examples. Finally, we demonstrate that personas synthesized from a single survey can reproduce the aggregate privacy behaviors and statistical distributions (TVComplement up to 0.85) of entirely different studies.

Figures (15)

• Figure 1: High-level overview of Narriva, which comprises three stages: initial persona generation, light-weight optimization, and prediction using the optimized persona.
• Figure 2: Individual- and population-level performance of text-based personas using 80:20 generation/evaluation split. Error bars show 95% confidence intervals obtained via bootstrapping. We report performance of summarized persona derived from question-answer pairs (Narrative), against that of the original question-answer pairs (Raw) and the LLM's internal privacy baseline without personalization (Baseline).
• Figure 3: The fraction of individuals across the five datasets for which each specific privacy theory template yields the highest predictive accuracy.
• Figure 4: Comparison of individual accuracy and population-level $TVD$ when using the Basic text-based persona versus dynamically selecting the Best-performing theoretical template per respondent. Error bars are obtained via bootstrapping.
• Figure 5: Performance when generating from behavioral questions and evaluating on behavioral questions, across prompt templates. Error bars are obtained via bootstrapping.