Personality as Relational Infrastructure: User Perceptions of Personality-Trait-Infused LLM Messaging

Applications, bias, and safety2026arXivApproved editorial review

Authors: Dominik P. Hofer, David Haag, Rania Islambouli, Jan D. Smeddinck

Keywords: Large Language Models, Personality, Big Five, Persona, Personality Control

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

4
Authors
20
Findings
88
Limitations
14
Evidence

Editorial summary

English

This paper asks whether adding a person's Big Five traits to LLM-generated physical-activity messages improves each message or is associated with more favorable global ratings when a person sees a higher proportion of personalized messages. This review uses the complete 53-page arXiv:2602.06596v2 manuscript, revised 27 February 2026. The paper is still “currently under review,” retains unfilled ACM CCS and citation-template placeholders, and replaces data, code, and IRB links with [BLIND_FOR_REVIEW]. No definitive publication or official public repository was found, so the data, code, fine-tuned models, and analysis cannot be reproduced.

The study has two phases. First, 210 sedentary Prolific participants rated five of fifteen hypothetical situations and provided feedback; this material fed few-shot examples, RAG retrieval, and fine-tuning. A declared final sample of 90 people then recalled five moments from the previous week, described twenty contextual variables, and rated messages they hypothetically would have received. Eight ChatGPT-4o configurations ran for each context: baseline, few-shot with self-consistency, RAG, and fine-tuning, each with and without BFI-10 scores. The eight outputs were shown in randomized tab order. The LLM first decided whether to send: of 3,600 decisions, it produced 2,345 messages and withheld 1,255. This was not a JITAI intervention in daily life: there was no real-time delivery, behavioral follow-up, longitudinal relationship, or physical-activity outcome.

The manipulation is also not validated “personality alignment.” The BFPT condition simply supplies the same model with five 1–5 scores and asks it to adapt tone and phrasing. No human judge or linguistic analysis verifies that messages express the traits correctly, feel congruent to the participant, or differ in a controlled way in length, specificity, and style. BFI-10 has only two items per trait and reliability is not reported in either sample. Advanced prompts are much longer and more prescriptive than baseline; RAG and few-shot add examples, while fine-tuning omits the final dataset, split, hyperparameters, and evaluation. The comparison therefore bundles personalization, prompt design, retrieval, training, and send policy.

The within-person contrast is the most defensible result. Among messages that were shown, no outcome had a credible trial-level association between BFPT availability and immediate rating under the intervals reported in the narrative. In this vignette evaluation, giving the generator Big Five scores did not produce a detectable advantage for the personalized message over the non-personalized message for the same person. Nor did RAG, fine-tuning, or few-shot show a consistent advantage over baseline. This is not proof of equivalence: no equivalence margin, power analysis, pinned models, or clean manipulation is provided, and the appendix shows radically different send policies across pipelines.

The main positive claim comes from a variable constructed after generation: each participant's proportion of delivered messages originating from BFPT configurations. It ranged only from 40% to 57% (SD=.03) because each LLM autonomously sent or withheld messages. It was not randomized. The appendix send model estimates that few-shot, fine-tuned, and RAG sent dramatically less often than baseline (OR .024, .008, and .046) and that non-BFPT configurations sent more often (OR 1.42). “Dose” is therefore an outcome of post-generation selection and pipeline composition, not administered experimental exposure. Small correlations and VIF checks do not remove unmeasured confounding, selection bias, or collider bias.

Ordinal models separate the message indicator from each person's mean and add participant random intercepts and slopes. They report favorable person-level associations for perceived personalization, fit, engagement, effectiveness, and professionalism, and negative associations with five adverse emotions. Yet the person-level predictor contains only 90 independent units and is repeated across roughly 2,345 rows. The design also nests five scenarios and up to eight correlated messages per scenario, but the formula models only participant. Priors, draws, posterior predictive checks, preregistration, power, and selection-sensitivity analyses are not reported. MCMC convergence does not cure an incomplete causal specification.

Internal contradictions prevent taking “14 of 17 outcomes” literally. Table 6 reports happiness β=2.99 with a 97% HDI [.34, 5.79], which excludes zero and would raise the count to 15; the narrative prints [-.34, 5.79] and labels it non-credible. The forest-plot caption calls the intervals “95% CI.” Seventeen outcomes are assessed with a 97% HDI rule without a multiplicity rationale. The four Perceived Message Relevance items are modeled separately instead of being used as a validated scale, and internal consistency is not reported. Confounder p-values in the appendix treat repeated rows as independent, making correlations around .09–.13 appear as p=.000.

Basic accounting also fails. Methods claims 77,280 ratings while Results claims 75,600. With 2,345 delivered messages and 21 items, at most 49,245 ratings are possible; 75,600 counts the 1,255 no-message decisions as if they had ratings, and 77,280 implies 3,680 messages, more than the 3,600 possible decisions. Table 5 confirms about 2,340 observations per outcome. The appendix Big Five table describes 92 comparison-study participants, every trait distribution sums to 92, not the 90 reported in the text and demographics, without explaining two exclusions. It also says ICCs range from “0.54% to 0.82%,” although the values are .54–.82, or 54–82%.

High ICCs mainly show that people used rating scales consistently during a long single session; they do not confirm that BFPT exposure created that stability. Each participant rated multiple variants of the same context in one interface, potentially hundreds of items, allowing response consistency, fatigue, direct comparison, and order effects. Raw negative-emotion means sometimes reverse the adjusted slope: the paper acknowledges that Q4 can be more negative than Q1 and that over 70% of ratings are at the floor. Attributing the reversal only to floor effects does not establish a reduction among a minority; it shows the conclusion is model-dependent rather than an observed group difference.

The Limitations section is more cautious than the framing. It acknowledges that “person-level” means between-person differences within one evaluation session, not within-person change over time, and that Communication Accommodation Theory was added post hoc. The data therefore do not observe cumulative accommodation, rapport, trust, therapeutic alliance, or “relational infrastructure.” They do not even observe a person receiving a coherent series from one system: up to eight independent configurations without historical memory are compared. CAT and therapeutic-alliance analogies are hypotheses for future studies, not demonstrated mechanisms.

The faithful interpretation is consequently narrower than the title. The work contributes a vignette protocol and a null message-level result for this BFPT implementation. It also detects an ecological association between a narrow, non-random post-hoc proportion and global rating tendencies across 90 participants. That association can motivate the longitudinal micro-randomized trial the authors themselves propose, but it does not demonstrate dose, accumulation, relational change, JITAI effectiveness, increased physical activity, or clinical benefit. Design recommendations require public data and code, corrected inconsistencies, explicit randomization of BFPT density, genuine repeated exposure, a verified manipulation, complete hierarchical modeling, and behavioral outcomes.

Español

Este trabajo pregunta si añadir los rasgos Big Five de una persona a mensajes de actividad física generados por un LLM mejora cada mensaje o si se asocia con una valoración global más favorable cuando una persona ve una proporción mayor de mensajes personalizados. Esta revisión usa el manuscrito completo arXiv:2602.06596v2, revisado el 27 de febrero de 2026, de 53 páginas. El documento sigue “currently under review”, conserva los CCS y la referencia de plantilla ACM sin completar y sustituye los enlaces de datos, código e IRB por [BLIND_FOR_REVIEW]. No se encontró publicación definitiva ni repositorio oficial público, por lo que los datos, el código, los modelos fine-tuned y el análisis no pueden reproducirse.

El estudio tiene dos fases. Primero, 210 participantes sedentarios reclutados en Prolific evaluaron cinco de quince situaciones hipotéticas y aportaron feedback; ese material alimentó ejemplos few-shot, recuperación RAG y fine-tuning. Después, una muestra final declarada de 90 personas recordó cinco momentos de la semana anterior, describió veinte variables contextuales y valoró mensajes que hipotéticamente habría recibido. Para cada contexto se ejecutaron ocho configuraciones de ChatGPT-4o: baseline, few-shot con self-consistency, RAG y fine-tuning, cada una con y sin las puntuaciones BFI-10. Las ocho salidas se mostraron mediante pestañas en orden aleatorio. El LLM decidía primero si enviar; de 3.600 decisiones, produjo 2.345 mensajes y retuvo 1.255. Esto no fue una intervención JITAI en la vida diaria: no hubo entrega en tiempo real, seguimiento de conducta, relación longitudinal ni outcome de actividad física.

La manipulación tampoco equivale a “alineación de personalidad” validada. La condición BFPT simplemente entrega al mismo modelo cinco puntuaciones de 1 a 5 y le pide adaptar tono y fraseo; no se comprueba mediante jueces humanos o análisis lingüístico que los mensajes expresen los rasgos correctamente, que resulten congruentes para la persona o que difieran de forma controlada en longitud, especificidad y estilo. El BFI-10 usa solo dos ítems por rasgo y no se informa fiabilidad en estas muestras. Los prompts avanzados son mucho más largos y prescriptivos que el baseline; RAG y few-shot incorporan ejemplos y el fine-tuning no documenta dataset final, split, hiperparámetros o evaluación. Por tanto, la comparación mezcla personalización, prompt, recuperación, entrenamiento y política de envío.

El contraste dentro de persona es el resultado más defendible. En los mensajes que sí se mostraron, ningún outcome tuvo una asociación trial-level creíble entre disponer de BFPT y la valoración inmediata según los intervalos del texto. Es decir, en esta evaluación de viñetas, dar al generador las puntuaciones Big Five no produjo una mejora detectable del mensaje personalizado frente al no personalizado para la misma persona. Tampoco hubo una ventaja consistente de RAG, fine-tuning o few-shot sobre el baseline. Esto no demuestra equivalencia: faltan márgenes de equivalencia, potencia, modelos fijados y una manipulación limpia, y el apéndice muestra que los pipelines tenían políticas de envío radicalmente distintas.

La principal afirmación positiva procede de una variable construida después de generar los datos: para cada participante, la proporción de mensajes entregados que pertenecía a configuraciones BFPT. Esa proporción varió solo de 40 % a 57 % (SD=0,03) porque cada LLM decidió de forma autónoma enviar o retener. No fue aleatorizada. De hecho, el modelo de envío del apéndice estima que few-shot, fine-tuned y RAG enviaban muchísimo menos que el baseline (OR 0,024, 0,008 y 0,046) y que las configuraciones sin BFPT enviaban más (OR 1,42). La “dosis” es así consecuencia de selección post-generación y composición de pipelines, no una exposición experimental administrada. Correlaciones pequeñas y VIF no eliminan confusión no medida, sesgo de selección ni collider bias.

Los modelos ordinales separan el indicador del mensaje de la media de cada persona y añaden intercepto y pendiente aleatorios por participante. Reportan asociaciones person-level favorables para personalización percibida, ajuste, engagement, efectividad y profesionalidad, y asociaciones negativas con cinco emociones adversas. Sin embargo, el predictor person-level contiene solo 90 unidades independientes y se repite en unas 2.345 filas; el diseño contiene además cinco escenarios y hasta ocho mensajes correlacionados por escenario, pero la fórmula solo modela participante. No se declaran priors, draws, posterior predictive checks, preregistro, análisis de potencia o sensibilidad a la selección. La convergencia MCMC no resuelve una especificación causal incompleta.

El manuscrito contiene contradicciones que impiden aceptar literalmente “14 de 17 outcomes”. La Tabla 6 da para happiness β=2,99 y HDI 97 % [0,34, 5,79], que excluiría cero y elevaría el total a 15; el texto imprime [-0,34, 5,79] y lo declara no creíble. La figura llama a los mismos intervalos “95% CI”. Se usan diecisiete outcomes con umbral HDI 97 % sin una justificación de multiplicidad. Además, los cuatro ítems de Perceived Message Relevance se analizan por separado en vez de como una escala validada y no se reporta consistencia interna. Las asociaciones de confusores del apéndice calculan p-values sobre filas repetidas como si fueran independientes, haciendo que correlaciones de 0,09–0,13 aparezcan como p=0,000.

También falla la contabilidad básica. Métodos afirma 77.280 ratings, mientras Resultados afirma 75.600. Con 2.345 mensajes entregados y 21 ítems, el máximo posible es 49.245 ratings; 75.600 corresponde a contar también las 1.255 decisiones sin mensaje y 77.280 equivaldría a 3.680 mensajes, más de las 3.600 decisiones posibles. La Tabla 5 confirma aproximadamente 2.340 observaciones por outcome. El apéndice de Big Five describe 92 participantes en la comparison study, cada distribución suma 92, no los 90 de texto y demografía, sin explicar dos exclusiones. También se escribe que los ICC van de “0,54% a 0,82%”, aunque los valores reales son 0,54–0,82, es decir 54–82 %.

Los ICC altos muestran sobre todo que las personas usan las escalas de forma estable durante una sesión larga; no confirman que la exposición BFPT haya creado esa estabilidad. Cada participante valoró en una misma interfaz múltiples variaciones del mismo contexto, potencialmente cientos de ítems, lo que puede producir consistencia de respuesta, fatiga, comparación directa y efectos de orden. El raw mean de emociones negativas contradice a veces la pendiente ajustada: el propio artículo reconoce que Q4 puede ser más negativo que Q1 y que más del 70 % marca el suelo. Atribuir la inversión solo al floor effect no prueba una reducción entre la minoría; muestra que esa conclusión depende del ajuste del modelo y no de la diferencia observada.

El artículo es más prudente en Limitations que en su framing. Reconoce que “person-level” significa diferencias entre personas en una única sesión, no cambio dentro de una persona a lo largo del tiempo, y que CAT se añadió post hoc. Por ello, los datos no observan acomodación acumulativa, rapport, confianza, relación terapéutica ni “relational infrastructure”. Tampoco observan que alguien reciba una serie coherente de mensajes de un mismo sistema: se comparan hasta ocho configuraciones independientes sin memoria histórica. CAT y la alianza terapéutica son hipótesis interpretativas para estudios futuros, no mecanismos demostrados.

La lectura fiel es, por tanto, más estrecha que el título. El estudio aporta un protocolo de viñetas y evidencia nula a nivel de mensaje para esta implementación de BFPT. Detecta además una asociación ecológica entre una proporción post hoc, no aleatorizada y de rango estrecho, y tendencias globales de rating entre 90 participantes. Esa asociación puede motivar un micro-randomized trial longitudinal, como proponen los propios autores, pero no demuestra dosis, acumulación, cambio relacional, eficacia JITAI, aumento de actividad física ni beneficio clínico. Antes de usarlo como recomendación de diseño hacen falta datos y código públicos, corrección de las inconsistencias, randomización explícita de densidad BFPT, exposición real repetida, manipulación comprobada, análisis jerárquico completo y outcomes conductuales.

Research question

Does adding Big Five scores to physical activity LLM messages change the immediate rating of each message and/or associate with different global ratings among people who, due to LLM sending decisions, end up seeing different proportions of BFPT messages?

Method

Two Prolific studies. A phase of 210 participants generates feedback for few-shot, RAG, and fine-tuning. In a retrospective single-session evaluation, 90 participants reconstruct five personal contexts and rate up to eight ChatGPT-4o messages per context: baseline, few-shot/self-consistency, RAG, and fine-tuned, with/without BFI-10. The LLMs decide non-randomly whether to send. 2,345 messages are analyzed via Bayesian ordinal regressions with within-between decomposition and random effects per participant.

Sample: Training: 210 self-reported sedentary Prolific adults, mean age 34.13, 122 women and 88 men. Comparison: main text N=90, mean age 37.20, range 22–72, 45 women/45 men and 80 White participants; only people with intention of physical activity for five weeks. The Big Five table in the appendix sums N=92 and exclusions are not explained. Former/new participants are mentioned without reporting sizes, possible overlap, or whether their previous data could be recovered.

Findings

  • The eight configurations take 3,600 decisions and deliver 2,345 messages; 1,255 are retained.
  • None of the 17 outcomes presents a credible trial-level association according to the narrative interpretation of the paper.
  • The incorporation of BFPT scores does not detectably improve the immediate rating of the message for the same person.
  • No coherent advantage of few-shot, RAG, or fine-tuning over baseline appears.
  • The person-level BFPT proportion varies only between 40% and 57%, SD=0.03, and is not experimentally manipulated.
  • The sending model reports OR 0.024 for few-shot, 0.008 for fine-tuned, and 0.046 for RAG versus baseline.
  • Configurations without BFPT have 1.42 times higher odds of sending, so BFPT affects the selection of which outputs reach ratings.
  • The models report favorable person-level associations in perceived personalization, fit, and five quality dimensions.
  • They report negative person-level associations for angry, annoyed, frustrated, sad, and scared.
  • Empathy and surprise are not credible at the person-level.
  • Happiness is internally contradictory: Table 6 HDI [0.34, 5.79], narrative [-0.34, 5.79].
  • The ICCs are 0.23 for empathy and approximately 0.53–0.81 for the other outcomes.
  • The ICCs indicate a strong stable response tendency between people during the session, not longitudinal accumulation.
  • The raw Q4 means of negative affect sometimes contradict the decreases predicted by the model.
  • More than 70% of negative emotion ratings are at the minimum value.
  • Table 5 confirms approximately 2,340 observations per outcome, consistent with 2,345 messages and not with 75,600/77,280 ratings.
  • The BFPT profile table for comparison sums 92 people and contradicts N=90.
  • The manuscript acknowledges that person-level is variation between people in a single session and not temporal change.
  • CAT was applied post hoc and did not guide hypotheses, manipulation, or measures.
  • The valid contribution is an exploratory association and a null trial-level result that justify a randomized longitudinal trial.

Limitations

  • The article is a preprint under review, not a definitive peer-reviewed publication.
  • It retains incomplete ACM markers and a Woodstock template reference.
  • Links to data, code, and IRB are blinded and not verifiable.
  • No official public repository has been located.
  • Messages, models, exclusions, analyses, or figures cannot be reproduced.
  • The comparison study is retrospective and single-session.
  • Participants imagine messages they would have received; they do not receive them in a real context.
  • There is no deployed adaptive decision, real notification, or behavioral receptivity measurement.
  • Physical activity, behavior change, adherence, or health are not measured.
  • There is no longitudinal exposure to a coherent system.
  • Each LLM decision uses only the current turn and lacks history.
  • Up to eight configurations of the same context are shown in tabs, not a natural conversation.
  • Comparing variants of the same context may create contrast and demand.
  • Tab order is randomized, but order, carryover, or fatigue are not modeled.
  • The potential burden is hundreds of ratings per participant and neither duration nor straightlining is reported.
  • BFPT personalization is not assigned as an experimental dose.
  • The person-level proportion is constructed after sending decisions.
  • The pipelines themselves and the presence of BFPT strongly change the probability of sending.
  • Conditioning analyses on sent messages introduces post-treatment selection.
  • The BFPT proportion mixes treatment, sending policy, and pipeline composition.
  • VIF measures collinearity, not absence of confounding.
  • Small correlations with observed covariates do not rule out unmeasured confounding.
  • The text uses language of isolating causal effects without causal identification.
  • The main formula does not make clear whether it controls for model_type despite its enormous sending differences.
  • The effective unit of the person-level predictor is N=90, not N≈2,345.
  • The model repeats the same person-level proportion in all rows of a participant.
  • Only random effects of participant are included, not of scenario/context.
  • Up to eight messages share context and participant, generating residual dependence.
  • Effects of message, order, or configuration nested within scenario are not explicitly modeled.
  • Priors, number of chains, draws, warmup, seed, or complete diagnostics are not published.
  • No posterior predictive checks are shown.
  • There is no preregistration or verifiable distinction between confirmatory and exploratory analysis.
  • There is no power or precision analysis for trial-level effects and N=90 person-level.
  • There is no sensitivity to alternative specifications, selection, or extreme proportion values.
  • The person-level exposure has range 40–57% and SD=0.03, very narrow.
  • Non-standardization due to convergence problems makes it difficult to compare magnitudes.
  • MCMC convergence does not demonstrate model adequacy or causal identification.
  • 17 correlated outcomes are evaluated without a multivariate model.
  • The 97% HDI threshold is unusual and not justified as multiplicity control.
  • The figure 14/17 contradicts Table 6 for happiness.
  • The narrative prints a negative HDI for happiness, but the table and figure show a positive limit.
  • The figure labels 95% CI while methods and tables use 97% HDI.
  • The four Perceived Message Relevance items are treated as separate outcomes.
  • The composite score and reliability of Perceived Message Relevance are not reported.
  • The reliability of the PETS in this sample is not reported.
  • The beta transform for empathy is not specified and the data include extremes 0 and 100.
  • The correlation p-values in the appendix use repeated observations and appear to be pseudoreplication.
  • The raw means of negative affect may have the opposite direction to the adjusted model.
  • The floor effect alone does not explain an inversion of crude-adjusted sign.
  • Claiming a reduction in a minority does not directly follow from the published ordinal model.
  • Methods declares 77,280 ratings, arithmetically incompatible with the design.
  • Results declares 75,600 ratings by implicitly counting decisions without a message.
  • 2,345×21 allows a maximum of 49,245 ratings.
  • 77,280/21 implies 3,680 messages, eighty more than the possible decisions.
  • The descriptive table confirms approximately 2,340 cases per outcome.
  • The Big Five table for the comparison study sums 92, not 90.
  • Two exclusions between the table and the final sample are not explained.
  • Participant type former/new is mentioned without reporting cohort sizes.
  • It is not clarified whether returning participants contributed data used to train or recover examples.
  • A returning participant could be represented in training data without leakage analysis.
  • The sample is restricted to declared intention of activity for five weeks.
  • The sample is 89% White and exclusively Prolific, limiting generalization.
  • Screening rate, dropouts, attention, or reasons for exclusion are not reported.
  • BFI-10 has two items per trait and no reliability or measurement error is reported.
  • The prompts treat scores 1–5 as style instructions without validated mapping.
  • There is no manipulation check that BFPT outputs are actually aligned.
  • Whether participants detect or prefer congruence with their own profile is not measured.
  • Length, readability, specificity, or content are not controlled across conditions.
  • The baseline is much shorter and less prescriptive than the advanced prompts.
  • Few-shot asks for internal chains and examples, so it does not isolate a single technique.
  • RAG lacks exact version of encoder, index, metric, and retrieval evaluation.
  • Fine-tuning lacks model IDs, epochs, split, final size, and metrics.
  • chatgpt-4o-latest is not an immutable snapshot and may change.
  • Temperature, top_p, max tokens, retries, or parse failures are not reported.
  • The generated outputs are not available for safety and quality audit.
  • Hallucinations, inappropriate advice, bias, or harm in health messages are not evaluated.
  • The privacy of sending context and psychological traits to OpenAI is not detailed.
  • IRB details are blinded and consent or data governance cannot be verified.
  • High ICCs may reflect stable response style, not perception of a system.
  • ICC does not confirm causal association with BFPT or a cumulative mechanism.
  • The analogy with therapeutic relationship is not supported by alliance measures or clinical outcomes.
  • CAT was selected after seeing the pattern and was not tested against alternative theories.
  • Bidirectional linguistic convergence is not measured.
  • Rapport, trust, bonding, system learning, or user control are not measured.
  • The term relational infrastructure exceeds the observed design.
  • Equivalence between strategies is not tested with non-inferiority/equivalence margins.
  • The trial-level nulls could reflect noise, low power, or weak manipulation.
  • Generalization to real JITAI, other domains, languages, cultures, and models is not demonstrated.

What the study does not establish

  • It does not establish that personalizing with Big Five improves an individual message.
  • It does not demonstrate that a higher BFPT dose causes better perceptions.
  • It does not demonstrate accumulation of effects within a person over time.
  • It does not demonstrate rapport, trust, therapeutic alliance, or human-AI relationship.
  • It does not validate the concept of relational infrastructure as an empirical mechanism.
  • It does not demonstrate that the messages are actually congruent with the receiver's personality.
  • It does not demonstrate that BFI-10 is sufficient for safe or precise personalization.
  • It does not demonstrate equivalence of baseline, few-shot, RAG, and fine-tuning.
  • It does not demonstrate efficacy of a deployed real-world JITAI.
  • It does not demonstrate increased physical activity, adherence, or health benefit.
  • It does not demonstrate population reduction of negative affect; the raw means sometimes invert.
  • It does not eliminate confounding, post-treatment selection, or bias from pipeline composition.
  • It does not allow reproducing data, models, or analyses from public artifacts.
  • It does not justify invisible psychological personalization without consent and user controls.
  • It does not allow deployment recommendations before a randomized longitudinal trial.

Traceability

Scope: Full text

Version: arXiv:2602.06596v2, revised 27 February 2026; under review; 53 pages

Consulted source: https://arxiv.org/pdf/2602.06596v2

Review: Codex full-text, bilingual-fidelity, 53-page visual, arXiv-version, study-design, manipulation-validity, JITAI-construct, selection-bias, hierarchical-model, multiplicity, scale-validity, arithmetic-reconciliation, appendix-consistency, LLM-prompt, reproducibility, privacy and causal-claim audit; summaries written from the complete paper rather than abstract keyword extraction, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI chatgpt-4o-latest, described as July 2025, exact immutable snapshot not reported, used for all four generation pipelines
  • Baseline ChatGPT-4o prompt with optional Big Five scores
  • Few-shot chain-of-thought/self-consistency ChatGPT-4o with three randomly selected training examples and internally requested N≈5 chains
  • RAG ChatGPT-4o with three Universal Sentence Encoder v3 nearest examples; exact index, distance, package version, and retrieval code unavailable
  • Two OpenAI supervised fine-tuned models, with and without BFPT, whose model IDs, training settings, splits, and artifacts are unavailable
  • Bayesian multilevel ordinal models implemented in Bambi; exact code, priors, sampler configuration, and posterior checks unavailable

Instruments and metrics

  • Big Five Inventory-10 (BFI-10), two items per trait, scored in 0.5 increments from 1 to 5
  • Twenty-feature retrospective context questionnaire for five recalled moments
  • LLM send/no-send recommendation score on a 1–5 scale with send threshold ≥3
  • Four Perceived Message Relevance items analyzed separately (p1–p4)
  • Six-item Perceived Empathy of Technology Scale emotional-responsiveness subscale aggregated to 0–100
  • Five adapted JITAI quality ratings: send appropriateness, content appropriateness, engagement, effectiveness, professionalism
  • Seven affect ratings: anger, annoyance, frustration, happiness, sadness, fear/scared, surprise
  • Post-hoc participant BFPT exposure proportion among delivered messages
  • Bayesian within-between decomposition with participant intercept and BFPT trial-level slope
  • Post-hoc Communication Accommodation Theory interpretation

Data used

  • Private training-data study: 210 Prolific participants, five randomly selected scenarios each from a pool of 15, ratings, free-text feedback, and BFI-10
  • Private comparison study: declared final N=90, five recalled contexts per participant, eight LLM configurations per context, 3,600 send decisions
  • 2,345 delivered messages and 1,255 withheld decisions
  • At most 49,245 raw item ratings from 2,345 messages × 21 items; manuscript instead claims 75,600 and 77,280
  • Appendix outcome table with approximately 2,340 non-missing observations per aggregated outcome
  • Appendix BFI distribution totaling 92 comparison participants, inconsistent with declared N=90
  • No public data, code, model artifacts, preregistration, or executable environment found as of 15 July 2026

Evidence and location

  • Status, version, and metadata: arXiv:2602.06596v2 abstract page and submission history; currently under review; revised 27 February 2026
  • Phases, participants, and procedure: Full text Sections 3.1–3.3, pp. 5–7
  • Eight configurations and prompts: Full text Section 3.4 and Appendix A.3, pp. 7 and 21–29
  • Measures and within-between model: Full text Sections 3.5–3.6, pp. 7–9
  • 2,345 sends and contradictory accounting: Full text pp. 6 and 9; Appendix Table 5 p. 32; independent arithmetic 2,345×21=49,245
  • Trial-level and person-level results: Full text Sections 4.3–4.6, pp. 10–12; Appendix Tables 6–7, pp. 35–36
  • Happiness and 14/17 contradiction: Narrative Section 4.4 versus Appendix Table 6 and Figure 8, p. 35
  • ICCs and floor effects: Full text Sections 4.5–4.6; Appendix Table 8, pp. 36–38
  • Selection by sending policy: Appendix A.9 Send Decision Modeling, p. 34
  • Pseudoreplication in correlations: Appendix A.6 pp. 31–32; p-values reported across repeated message rows for person-level means
  • N=92 versus N=90: Appendix Table 3 pp. 19–20; independent sum of every BFPT distribution equals 92
  • Temporal scope and post hoc CAT: Full text Sections 5.1, 5.5, 5.7–5.8, pp. 13–15
  • Absence of public artifacts: Full text blind placeholders; exact-title and arXiv-ID searches of GitHub and web performed 15 July 2026
  • Complete visual inspection: All 53 pages of arXiv:2602.06596v2 rendered and visually inspected on 15 July 2026