Exploring the Information Privacy Preferences in Large Language Models

Read full text Summarise with VIRAT
Authors: Nan Wen
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study provides the first systematic evaluation of LLMs’ information privacy preferences, comparing 20 state-of-the-art LLMs. Drawing from established human privacy scales (e.g., CFIP, IUIPC) and classic experimental designs, this study measures whether LLMs show specific information privacy preferences and disclose sensitive information under different contexts and different incentives, whether these decisions align with their stated preferences, and whether different demographic and socioeconomic characteristics in prompts influence their responses. This study finds that relatively advanced LLMs (e.g., GPT-4.1, Claude-3.5-Sonnet, Gemini series) demonstrate a stronger and more stable privacy preference in general contexts, financial contexts, and online contexts, but show unexpectedly lower concern in medical settings, where they tend to trust others. In contrast, smaller or older models exhibit less consistent and weaker privacy-concerned behaviour, with relatively higher medical sensitivity but erratic preferences elsewhere. Notably, LLMs reproduce the same pattern observed in human participants: they are generally less concerned about data collection than other dimensions in the surveys. Moreover, in the experiments, we find that without incentives, most LLMs show relatively neutral disclosure willingness, but higher than that of humans. Unlike humans, LLMs are largely unresponsive or even negatively affected by financial and psychological incentives, and only convenience consistently increases LLMs’ disclosure willingness. Lastly, we discovered that compared to GPT-3.5-Turbo, which shows a less consistent preference with high volatility, GPT-4.1 with a stronger preference better captures human-like trends when prompted by a different gender, age, or income level. This greater sensitivity to demographic and socioeconomic characteristics suggests that it is more suitable to provide personalised recommendations aligned with human preferences. Overall, our findings shed a light on how to use LLMs to make privacy-related decisions or recommendations. Keywords: Large Language Models, Information Privacy Concerns, Privacy Preference, Decision Making.

Fields of Study & Subjects

Privacy Security and Data Protection Privacy-Preserving Technologies in Data Ethics and Social Impacts of AI

Publication details

Journal / source
Macquarie University
Publication date
26 March 2026
Language
EN
Document type
Thesis