Synthetic Personality Review

Text-Based Personas for Simulating User Privacy Decisions

Applications, bias, and safety2026arXivApproved editorial review

Authors: Kassem Fawaz, Ren Yi, Octavian Suciu, Rishabh Khandelwal, Hamza Harkous, Nina Taft, Marco Gruteser

Keywords: Persona conditioning, Human simulation

Source: Open primary source (opens in a new tab)

7
Authors
11
Findings
37
Limitations
4
Evidence

Editorial summary

English

Fawaz and colleagues introduce Narriva, a procedure that compresses a person's prior privacy-survey answers into a structured text profile and uses it to anticipate answers to other questions. Gemini 3.0 Flash generates five candidate personas per iteration and receives feedback on errors over the generation questions; Gemini 2.5 Flash-Lite answers evaluation questions. The main experiment splits answered questions within each respondent: 80% build and optimise the persona and 20% test it, with up to three iterations. Five datasets are used: SPA, Pew PP1, W49, W127, and CAuthN. Relative to a non-personalised LLM, both literal history and Narriva improve performance. Narrative-persona individual accuracy is 84.6% on SPA, 67.7% on PP1, 64.3% on W49, 66.1% on W127, and 45.8% on CAuthN; literal history obtains 85.2%, 69.4%, 65.8%, 66.6%, and 50.2%. For per-question marginal-distribution similarity, measured as 1-TVD, Narriva reaches 0.94, 0.84, 0.80, 0.81, and 0.73. Personas reduce prompt length by 82-95%, from 1,738-8,920 to 296-453 tokens. Prompts inspired by Privacy Calculus, Bounded Rationality, and Protection Motivation Theory perform similarly in aggregate to the basic persona, although post-hoc selection of the best template for each respondent improves some results. The strongest finding is engineering-oriented: within the same questionnaire, a much shorter textual representation preserves much of the raw history's predictive value. It requires precise boundaries. The near-87% maximum comes from SPA, where roughly 144 questions repeat highly similar combinations of data type, recipient, and purpose; this is generalisation to held-out items from the same instrument, not arbitrary new situations. Neutral responses are removed from SPA and part of CAuthN, empty responses and refusals are excluded, and a single random split has no published seed or sensitivity analysis. Exact accuracy also cannot be compared directly between a binary outcome and 5-, 7-, or up-to-100-value scales. At population level, 1-TVD separately compares each question's marginal distribution: it can be high despite many individual errors and does not preserve joint relationships, subgroups, or fairness. Cross-study transfer applies each source-study persona to every unique target question and compares weighted aggregates; no individuals are linked. Thus PP1-to-W127 at 0.808 or W49-to-W127 at 0.782 show transport of marginals among US privacy surveys, not longitudinal portability of one person's personality or transfer across wholly different populations. Temporal drift is a plausible interpretation, but year, panel, sample, response scale, instrument, and topic all change together. The paper claims auditability and improved privacy through compression without user studies or reconstruction, inference, linkage, or re-identification attacks; a compact persona can itself be a sensitive psychographic artifact. It also does not evaluate real autonomous decisions, calibrated confidence, abstention, user correction, safety, or governance of portable persona files. The bootstrap does not identify its resampling unit, and significance claims lack specified tests and paired contrasts. Code, processed matrices, splits, personas, predictions, and analyses are not released. Narriva is a promising proposal for within-survey response simulation and aggregate survey transport, not yet evidence of an interpretable, private, fair, or deployment-ready privacy assistant.

Español

Fawaz y colaboradores presentan Narriva, un procedimiento para condensar las respuestas previas de una persona a encuestas de privacidad en un perfil textual estructurado y usarlo para anticipar cómo responderá a otras preguntas. Gemini 3.0 Flash genera cinco perfiles candidatos por iteración y recibe feedback sobre los errores en las preguntas de generación; Gemini 2.5 Flash-Lite contesta las preguntas de evaluación. El experimento principal divide dentro de cada encuestado sus preguntas contestadas: 80% para construir y optimizar el perfil, 20% para probarlo, con hasta tres iteraciones. Se usan cinco datasets: SPA, Pew PP1, W49, W127 y CAuthN. Frente a un LLM sin personalización, tanto el historial literal como Narriva mejoran. La exactitud individual del perfil narrativo es 84,6% en SPA, 67,7% en PP1, 64,3% en W49, 66,1% en W127 y 45,8% en CAuthN; el historial literal obtiene 85,2%, 69,4%, 65,8%, 66,6% y 50,2%. En similitud de la distribución marginal por pregunta, medida como 1-TVD, Narriva alcanza 0,94, 0,84, 0,80, 0,81 y 0,73. Los perfiles reducen el prompt entre 82% y 95%, de 1.738-8.920 a 296-453 tokens. Los prompts inspirados en Privacy Calculus, Bounded Rationality y Protection Motivation Theory rinden de forma agregada parecida al perfil básico, aunque seleccionar post hoc la plantilla que mejor funciona para cada encuestado mejora algunos resultados. El hallazgo más sólido es ingenieril: dentro del mismo cuestionario, una representación textual mucho más corta conserva gran parte de la capacidad predictiva del historial. Debe leerse con límites precisos. El máximo cercano al 87% procede de SPA, donde unas 144 preguntas repiten combinaciones muy similares de dato, receptor y propósito; es generalización a ítems retenidos del mismo instrumento, no a cualquier situación nueva. Se eliminan respuestas neutrales en SPA y parte de CAuthN, se excluyen vacíos y rechazos y una única partición aleatoria no publica seed ni análisis de sensibilidad. La exactitud exacta tampoco es comparable sin más entre una variable binaria y escalas de 5, 7 o hasta 100 valores. En población, 1-TVD compara por separado la distribución marginal de cada pregunta: puede ser alto aunque muchas predicciones individuales sean erróneas y no conserva relaciones conjuntas, subgrupos ni equidad. La transferencia entre estudios aplica cada perfil del estudio origen a todas las preguntas únicas del destino y compara agregados ponderados; no hay personas enlazadas. Por tanto, PP1→W127 0,808 o W49→W127 0,782 muestran transporte de márgenes entre encuestas estadounidenses de privacidad, no portabilidad longitudinal de la personalidad de un individuo ni transferencia entre poblaciones totalmente distintas. Atribuir las diferencias a deriva temporal es plausible, pero año, panel, muestra, escala, instrumento y tema cambian a la vez. El artículo afirma auditabilidad y mejora de privacidad por compresión sin estudios con usuarios ni ataques de reconstrucción, inferencia, enlace o reidentificación; un perfil compacto puede ser por sí mismo un artefacto psicográfico sensible. Tampoco se evalúan decisiones autónomas reales, confianza calibrada, abstención, corrección por el usuario, seguridad o gobernanza del archivo portátil. Las bootstrap no especifican la unidad de remuestreo y las afirmaciones de significación carecen de tests y contrastes pareados. No se publican código, matrices procesadas, splits, perfiles, predicciones ni análisis. Narriva es una propuesta prometedora para simulación de respuestas dentro y entre encuestas a nivel agregado, no evidencia aún de un asistente de privacidad interpretable, privado, justo o listo para despliegue.

Research question

Can a textual persona, compressed from historical responses and structured with privacy theories, retain individual predictive capacity, reproduce population distributions, and transfer to questions from other surveys with fewer tokens than the literal history?

Method

Narriva generates profiles with Gemini 3.0 Flash (temperature 1.5), predicts with Gemini 2.5 Flash-Lite (temperature 0), and, in the main experiment, produces B=5 candidates for up to I=3 iterations. Within each respondent, it divides the answered questions 80:20 for generation/evaluation. It compares a normative LLM without personalization, literal history, and narrative persona via individual accuracy and 1-TVD marginal per question; it adds MEE and Wasserstein, four theoretical templates, calibration-based selection, aggregate transfer across three Pew surveys, demographic profiles, and a bootstrap of 1,000 resamples. The qualitative component selects 60 profiles with the highest and 60 with the lowest accuracy per dataset, proposes codes with Gemini 3.1 Pro, and manually reviews 100.

Sample: Five studies with 607-5,101 respondents, all adults and predominantly from the United States; SPA has 1,737 Prolific participants whose country, according to the appendix itself, is not reported in the original study. The individual unit is the respondent, but the holdout splits questions from that same person. In transfer, source and target respondents are not linked: each source profile answers all unique target questions to compare weighted distributions. The qualitative component takes 600 extreme profiles and manually reviews 100.

Findings

  • Narriva achieves individual accuracy of 84.6% in SPA, 67.7% in PP1, 64.3% in W49, 66.1% in W127, and 45.8% in CAuthN.
  • Literal history obtains 85.2%, 69.4%, 65.8%, 66.6%, and 50.2%; Narriva retains much of it, but does not always match its individual accuracy.
  • The population similarity 1-TVD of Narriva is 0.94, 0.84, 0.80, 0.81, and 0.73, higher than the normative LLM without personalization.
  • The profiles use 296-453 tokens compared to 1,738-8,920 for the history, a reduction of 82-95%.
  • The Basic, Calculus, Bounded, and PMT templates have similar aggregate results; no theory dominates globally.
  • Selecting post hoc the best template per person improves some results, especially attitude→behavior, although small calibrations limit stability.
  • Behavior questions predict other behaviors better than general attitudes do, especially in SPA with repetitive items.
  • In attitude→behavior, PP1→W127 obtains 1-TVD 0.808 versus in-study baseline 0.744; W49→W127 obtains 0.729 versus 0.744.
  • In behavior→behavior, W49→W127 obtains 0.782 versus 0.780; PP1→W127 drops to 0.676 versus 0.780.
  • One iteration improves synthetic demographic profiles from 0.632 to 0.739 in W49 and from 0.585 to 0.716 in W127, but only the population margin is evaluated.
  • The qualitative component associates greater predictability with consistent limits, clear institutional views, and a sense of control, but analyzes generated explanations, not testimonies or validated causes.

Limitations

  • The split is by questions within the same person and survey; it does not test generalization to new individuals or contexts outside the instrument.
  • SPA contains around 144 scenarios with repeated structure, which facilitates the maximum close to 87%.
  • No seed, split repetition, or sensitivity analysis to the random split is published.
  • The optimization selects five candidates and corrects errors on the generation set; the holdout avoids direct leakage, but semantic similarity favors extrapolation.
  • Neutral responses in SPA and part of CAuthN are removed, and blanks and rejections are excluded, without quantifying selection or coverage.
  • Exact accuracy is not comparable across binary results, Likert scales, and 1-100 responses.
  • An empirical baseline per question is missing, mode or observed distribution, that would separate personalization from simple fitting to the majority margin.
  • 1-TVD measures marginal distributions per question, not joint structure, correlations, trajectories, or individual fidelity.
  • A good aggregate distribution can coexist with many individual errors; CAuthN combines 45.8% accuracy with 0.73 1-TVD.
  • There is no subgroup analysis, worst group, intersections, disparities, or calibration by protected class.
  • Transfer crosses surveys, not individuals; it does not prove that the portable profile predicts its owner on another platform.
  • The three transfer surveys are American and about privacy; W49 and W127 share the American Trends Panel.
  • Year, panel, sample, topic, question, and scale change together; the difference cannot be causally attributed to temporal drift.
  • Source and target weights correct each survey, but there is no explicit adjustment for transport, overlap, or covariate alignment.
  • The bootstrap does not specify whether it resamples persons, questions, or rows, despite the cross-dependence person×item.
  • Significance claims do not publish tests, p-values, paired contrasts, difference intervals, or multiple correction.
  • Template selection splits calibration and test again; in several datasets only about 2.0-3.9 behavioral evaluation questions per person remain.
  • It is not explained how frequent ties between templates with discrete accuracies are resolved.
  • A single Gemini family predominates; there is no replication of Narriva across providers or immutable snapshots.
  • Reducing tokens and adding sections does not measure comprehension, editability, audit time, correction, or human trust.
  • The theoretical templates are instructions, not validated measurements of the real psychological theory of each respondent.
  • Profiles may infer biases, confidence, expertise, or causal logic not unequivocally contained in the responses.
  • The qualitative component uses codes proposed by an LLM, reviews only 100 of 600 extremes, and does not report double coding, reliability, saturation, or negative cases.
  • The qualitative quotes are text generated by the model, not words from participants.
  • Low trust and user consultation are mentioned, but calibrated trust, abstention, or selective scaling are not measured.
  • No reconstruction, membership, attribute inference, linkage, reidentification, or prompt extraction attacks are performed.
  • A compact profile can concentrate sensitive psychographic data; compression does not equate to anonymization or differential privacy.
  • Original consent, usage licenses for LLM profiling, and participant expectations per dataset are not audited.
  • Access, encryption, purpose, retention, revocation, deletion, correction, provenance, or freshness of the portable profile are not evaluated.
  • Predicting past behavior can automate resignation, confusion, or obsolete preferences; there is no evaluation of contestability or override.
  • The token reduction is real, but total generation consumed tens of billions and the break-even depends on unsensitized assumptions.
  • Periodic updating due to drift is used in the cost simulation, but is not validated longitudinally.
  • Immutable versions, execution dates, top-p, seed, safety, retries, parsing failures, or refusals of the models are not published.
  • Preprocessing, recoding, attitude/behavior labels, splits, executable prompts, profiles, predictions, bootstrap, and code are not published.
  • Page 2 swaps the bibliographic provenance: SPA is CHI 2021 and CAuthN PoPETs 2024, not the other way around.
  • Table 1 labels SPA as UK, but the appendix says the original study does not report the country.
  • It is an arXiv v2 preprint; the record does not certify peer review or acceptance at a conference.

What the study does not establish

  • It does not demonstrate reliable prediction of arbitrary privacy decisions outside the originating questionnaire.
  • It does not demonstrate that a profile retains the personality or preferences of its owner across platforms or years.
  • It does not demonstrate individual transfer: comparisons across studies only reproduce population margins.
  • It does not demonstrate that high 1-TVD preserves joint structure, diversity, minorities, equity, or individuals.
  • It does not demonstrate that Privacy Calculus, Bounded Rationality, or PMT describe the real psychology of the respondent.
  • It does not demonstrate auditability, comprehension, editability, or utility for human users.
  • It does not demonstrate that compression improves privacy, anonymity, or resistance to attacks.
  • It does not validate a safe autonomous assistant that knows how to abstain, request clarification, or protect the profile.
  • It does not causally identify temporal drift of attitudes or behaviors between 2014 and 2023.
  • It does not allow reproduction of results due to lack of code, processed data, splits, outputs, and complete configuration.

Traceability

Scope: Full text

Version: arXiv:2603.19791v2, submitted 2026-03-20, revised 2026-05-07, CC BY 4.0

Consulted source: https://arxiv.org/abs/2603.19791

Review: Codex 22-page visual full-text, within-survey validity, semantic-overlap, population-marginal, cross-study transport, statistical, privacy, governance, cost, metadata and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Gemini 3.0 Flash, persona generation, temperature 1.5
  • Gemini 2.5 Flash-Lite, response prediction, temperature 0
  • Gemini 3.1 Pro, initial qualitative coding

Instruments and metrics

  • Narriva persona-generation and correction prompts
  • Basic persona template
  • Privacy Calculus persona template
  • Bounded Rationality persona template
  • Protection Motivation Theory persona template
  • Basic and theoretical prediction prompts
  • Within-respondent 80:20 question split
  • Individual exact-match accuracy
  • Total Variation complement, macro-averaged by question
  • Mean Estimation Error
  • Wasserstein distance
  • Survey-weighted cross-study population comparison
  • 1,000-resample bootstrap confidence intervals
  • LLM-assisted qualitative thematic analysis

Data used

  • SPA: 1,737 Prolific respondents, pre-2021; smart-assistant data-sharing scenarios, IUIPC, and SA-6
  • Pew PP1: 607 US adults, January 2014; privacy, surveillance, control, concern, and sensitivity
  • Pew W49: 4,272 US adults, June 2019; privacy trade-offs, data control, policies, and sharing scenarios
  • Pew W127: 5,101 US adults, May 2023; collection, control, privacy behavior, AI, cameras, and messaging
  • CAuthN: 830 US Internet users, September-October 2022; continuous biometric authentication, risk, trust, IUIPC, and SA-6
  • Generated persona candidates, feedback traces, predictions, splits, and processed matrices: not released

Evidence and location

  • Method, prompts, datasets, results, metrics, transfer, qualitative, discussion, ethics, and appendices: arXiv:2603.19791v2, 22/22 pages rendered and individually inspected
  • Version v2, dates, DOI, CC BY 4.0 license, and absence of official links to artifacts: Official arXiv abstract and Atom records inspected 2026-07-17
  • No current location of a study repository and discarding of homonyms: Exact-title, arXiv-ID, author, web and GitHub repository searches; candidate repository metadata and README inspected 2026-07-17
  • Audit of validity, population marginals, transfer, statistics, privacy, governance, cost, metadata, and reproducibility: reports/verification/article-388-narriva-within-survey-semantic-overlap-population-marginals-privacy-claims-cross-study-transfer-statistics-data-governance-and-reproducibility-audit.json