PrivacySIM asks a focused question: can an LLM reproduce a person's responses to privacy decisions when given three kinds of information about that person, demographics, prior AI or chatbot experience, and stated privacy attitudes? Its main contribution is a public benchmark assembled from five previously published user studies. The domains cover LLM healthcare consultation, personal-agent permissions, appropriateness of ChatGPT-history uses, bots operating inside group chats, and self-reported sharing frequency with conversational agents. The experimental file balances the domains at 200 participants each, 1,000 in total. Three studies expose demographics; all five expose experience and attitudes.
Each participant retains the questionnaire and their actual survey response. The pipeline creates a no-persona prompt and every available combination of the three facet blocks: four conditions in the two domains without demographics and eight in the other three. This yields 6,400 logical rows per model. Nine LLMs are evaluated: Gemini 3.1 Pro and 3 Flash; GPT-5.4 and GPT-5.4 Mini; Claude Sonnet 4.6; Gemma 4 32B; Qwen3.5 27B and 122B; and Nemotron 3 120B. The primary metric averages, per participant, the fraction of matched questions. Four- and five-option ordinal scales require exact agreement; the 0-100 ChatGPT study accepts any prediction within 15 points. The abstract's 40.4% is therefore tolerance accuracy, not homogeneous exact-match accuracy.
The strongest reported result is Gemini 3.1 Pro with prior experience plus attitudes: 40.39%, compared with 36.32% without a persona, a descriptive gain of 4.07 points. Combining those two blocks generally provides a modest improvement over the no-persona prompt. Attitudes alone often underperform experience, while the combination recovers some signal. Demographics alone contributes little. Larger models or more reasoning offer small gains, and Privacy Calculus, Bounded Rationality, or Protection Motivation Theory framing does not improve performance consistently. The authors interpret the low ceiling as evidence that age, AI use, and stated attitudes are insufficient to reconstruct individual privacy decisions.
The group analysis converts attitudes and exposure into two domain-normalized axes and assigns five clusters to fixed centroids. Low-stance, high-exposure users have the lowest experience-plus-attitudes accuracy, 36.51%, and gain only 1.07 points over no persona. High-stance, mid-exposure users reach 42.41% and improve by 8.53 points. Local reconstruction exactly reproduces the reported cluster sizes: 226, 187, 205, 161, and 221. These groups are nevertheless engineered bins built from heuristic parsing, keyword polarity, z-scores, and synthetic centers. They borrow Dupree et al.'s geometry but replace privacy knowledge with AI exposure, so they are not validated psychological persona types.
The quantitative audit finds a decisive limitation. A simple empirical baseline, using no LLM and no persona attributes, learns on four folds the tolerance-optimal response for each domain and question position and is evaluated on the fifth. Its balanced mean is 45.11%, above the best LLM's 40.39%. It scores 43.20% on permissions, 36.07% on group chat, 49.95% on the 0-100 domain, 54.15% on conversational agents, and 42.20% on healthcare. In the 0-100 domain, always answering 15 achieves 49.95% because every target from 0 to 30 falls inside the tolerance window. In conversational agents, always answering “I never shared” reaches 53.75%, because that category accounts for 1,075 of 2,000 targets. The paper reports no majority, per-question, calibrated, or supervised baselines. Neither the 40.4% result nor the LLM no-persona score therefore establishes competitive predictive fidelity.
Ground truth also needs careful interpretation. It consists of survey answers: hypothetical acceptability, concern, vignette appropriateness, or recalled sharing frequency. These are not behavioral logs. The benchmark measures questionnaire-response imitation, not verified simulation of real decisions. Lower performance from attitudes alone is compatible with a privacy paradox, but also with poor facet encoding, prompt overload, or model insensitivity; the design does not causally identify the paradox. There is no second questionnaire from the same person, test-retest analysis, comparable human prediction, or calibration study establishing stable individual identity recovery.
The balanced sample is useful for benchmarking but does not represent a population. It combines different countries, recruitment schemes, technologies, and constructs, and gives equal weight to studies whose processed pools range from 203 to 846 rows. Healthcare is restricted to mental- or medical-information questionnaires with informative experience before sampling. For conversational agents, the appendix says 422 participants narrow to 318 chatbot users; the code then drops anyone missing any of ten targets and retains 277, an additional 41-person complete-case exclusion not reported in the paper. Calling the 2,000-row CSV “all users, unfiltered” is therefore misleading: it is complete only relative to already processed tables.
Software reproducibility is comparatively strong. The project has an MIT license, attribution for all five sources, a lockfile, CI, a schema, four provider backends, and 20 tests; all pass in a dependency-complete environment. Prompt generation yields 6,400 rows and clustering reproduces the table. GitHub does not, however, publish model outputs, evaluation summaries, run.json provenance, or participant-level result files. Verifying 40.4% requires rerunning nine expensive or proprietary endpoints, several with mutable preview aliases. There are no confidence intervals, participant bootstrap, repeated decodes, or paired persona-minus-no-persona tests. The CLI also counts physical newlines inside CSV fields and falsely reports 453,600-543,200 rows even though pandas confirms 6,400. The Hugging Face Dataset Viewer currently fails, although both CSV files remain directly downloadable and parse correctly.
A faithful reading is positive but narrower than the abstract: PrivacySIM contributes a public, diverse, well-documented corpus for studying LLM imitation of privacy responses and shows that the tested persona fields add a small signal. At the same time, the best LLM does not beat an elementary empirical baseline, one domain's metric is permissive, result outputs are missing, and targets are not observed behavior. The work supports the conclusion that these three persona blocks are insufficient for reliable individual simulation; it does not establish that 40.4% is competitive or that LLMs have learned each user's real privacy behavior.