PrivacySIM: An Evaluation Suite for LLM Simulation of User Privacy Behavior

PrivacySIM aggregates user privacy preferences from five published user studies into a unified evaluation suite for LLM simulation of user privacy behavior. Each row represents one participant's responses to a study-specific questionnaire about whether data-sharing involving an LLM is appropriate, plus persona facets (demographics, previous experience, privacy attitudes).

The five contributing studies cover distinct data-sharing contexts: LLM chatbots, LLM healthcare consultations, AI agent permissions, group chat with bots, and conversational agents. Identifiers are derived deterministically from (user_id, domain) via UUID5.

Loading

Each configuration exposes a single train split containing every row of the corresponding CSV file (the data is evaluation-only; the split name is a Hub convention, not a train/test partition).

from datasets import load_dataset

# Full unfiltered dataset (2,000 users)
ds = load_dataset("jamesflemings/PrivacySIM", "full_dataset", split="train")

# 1,000-user sample used in the paper
ds = load_dataset("jamesflemings/PrivacySIM", "sampled", split="train")

Source studies

The data is derived from five public user studies. Each row's domain column identifies the source study. Please cite the original studies in addition to this dataset if you use the corresponding rows.

domain	Source	Users in full_dataset.csv	Users in dataset.csv
LLM healthcare consultation	LLM health consultation user study	846	200
Chatbot group chat	Bot-among-us chatbot group chat study	374	200
LLM chatbot	LLM chatbot privacy norms factorial vignette study	300	200
LLM conversational agents	LLM conversational-agent privacy preferences study	277	200
AI agent permissions	AI agent permission user study	203	200
Total		2,000	1,000

Code

The full reproducibility pipeline (data download, processing, simulation, evaluation, plotting) is at github.com/james-flemings/PrivacySIM. See docs/SCHEMA.md for the column-by-column reference.

License

The combined dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0). Per-row content remains under each source study's original license:

domain	Source license
LLM healthcare consultation	MIT
Chatbot group chat	MIT
LLM chatbot	CC0 1.0 (public-domain dedication)
AI agent permissions	CC BY 4.0
LLM conversational agents	CC BY 4.0

Full attribution, license texts, copyright notices, and indication of modifications for every source study are provided in the THIRD_PARTY_LICENSES.md file shipped alongside the dataset. Anyone redistributing or building on this dataset must preserve that attribution file (or an equivalent).

Citation

If you use PrivacySIM, please cite the paper (arXiv:2605.12147) and the source studies above:

@article{flemings2026privacysim,
  title         = {{PrivacySIM}: Evaluating {LLM} Simulation of User Privacy Behavior},
  author        = {Flemings, James and Annavaram, Murali},
  journal       = {arXiv preprint arXiv:2605.12147},
  year          = {2026},
  eprint        = {2605.12147},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR}
}