Long-Term Simulation Exposes Cognitive-Developmental Risks in AI Companions

Applications, bias, and safety2026arXivApproved editorial review

Authors: Kaicheng Shen, Lingyu Li, Wen Wu, Yan Teng, Liang He, Yingchun Wang

Keywords: Persona conditioning, Human simulation, Safety and bias, Longitudinal simulation, AI companions, LLM-as-judge, Developmental psychology, Measurement validity

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

6
Authors
8
Findings
19
Limitations
7
Evidence

Editorial summary

English

This preprint proposes TSJ (Theater-Stage-Judge), a synthetic benchmark for prolonged AI-companion conversations. Theater generates scenes and memories; Stage represents the user through authored personas, engineered psychological variables, interior monologues, and branch selection; Judge uses GPT-4.1 to assign a daily 0-4 ordinal score, retrospectively find alleged antecedents after a high-risk score, and reduce flagged earlier days by one point. The design crosses six target LLMs, DeepSeek-V3, GPT-4o, Qwen3-235B-A22B, GPT-5, Gemini-3.1-pro, and MiniMax-M2.5, with 24 CDM dimensions and three Red/Yellow/Green personas. Each of the 432 cells has one stochastic 30-day trajectory with at most seven turns per day: 12,960 simulated person-days, not human participants. Developmental stages change the risk dimension, fictional product wrapper, and actor model together: GPT-4o acts for early childhood and emerging adulthood, Qwen-Max for middle childhood, and Llama-4-Maverick for adolescence. Age differences therefore do not isolate development. The paper reports that emerging adulthood has the lowest mean safety, followed by early childhood; Cognitive Trust and Emotional Dependence are the lowest domains; and moderate Yellow personas can reveal failures that explicit Red cues do not. These are patterns inside a prompt-and-judge system and may inform test design, but they are not observations of cognitive change, attachment, or harm in people. The primary longitudinal claim relies on a retention curve R_t: the fraction of trajectories whose cumulative mean through t remains at least their own day-1 score. This definition forces R_1=100%. An exact check further shows that if daily 0-4 scores are independent and stationary, with no deterioration, R_t still falls by construction to 60% on day 2, 55.2% on day 3, 50.65% on day 20, and 50.52% on day 30; null AULC is 52.31. The reported AULCs, roughly 48-61, overlap this null range. A system scored 0 on every day would even have R_t and AULC of 100. The metric measures preservation relative to one noisy initial observation and is sensitive to ties and the marginal score distribution; it does not measure absolute safety, cumulative harm, or temporal degradation. Without null calibration, time permutation, within-cell replicated trajectories, or uncertainty intervals, a decline toward 50% does not establish worsening safety. Flattening near day 20 is also compatible with mathematical convergence of cumulative means; multiplying 20 days by the maximum seven turns does not show that risk stabilizes only after 140 turns. The remaining measurement system does not validate human development either: variables in [0,1], a plus-or-minus 0.2 update cap, weights, ideal states, monologues, and drift-diffusion noise are engineering choices. Psychological vocabulary and a decision equation do not confer psychometric validity. GPT-4.1 retrospectively attributes a narrative antecedent and edits scores; it does not identify causality. Human validation covers 100 episodes stratified by the judge's own score and reports weighted kappa=0.790 against three-expert consensus. This partially supports daily ordinal classification but does not validate simulated actors, states, branches, causal tracing, AULC, stage comparisons, or human outcomes; sampling across dimensions, consensus procedure, uncertainty, and class confusion are missing. Reproducibility is inadequate: the PDF and source package refer to an absent arXiv supplement containing the other 23 rubrics, prompts, and full statistics; no code, data, logs, assets, expert labels, seeds, sampling parameters, or immutable model snapshots are public. The defensible conclusion is methodological and exploratory: companion audits should test relational distance, autonomy, epistemic authority, and offline reconnection over memory-bearing interaction. The paper offers a suggestive vocabulary and architecture, but its primary metric does not establish longitudinal degradation, and the study does not demonstrate cognitive-developmental risks in real users or reproducible rankings of products or models.

Español

El preprint propone TSJ (Theater-Stage-Judge), un banco sintético para evaluar conversaciones prolongadas con acompañantes de IA. Theater genera escenas y memorias; Stage representa al usuario mediante personas, variables psicológicas diseñadas por los autores, monólogos interiores y selección de ramas; Judge usa GPT-4.1 para puntuar cada día en una escala ordinal 0-4, buscar retrospectivamente supuestos antecedentes cuando aparece una puntuación de alto riesgo y reducir en un punto días anteriores señalados. El diseño cruza seis LLM objetivo, DeepSeek-V3, GPT-4o, Qwen3-235B-A22B, GPT-5, Gemini-3.1-pro y MiniMax-M2.5, con 24 dimensiones CDM y tres personas Red/Yellow/Green. Cada una de las 432 celdas tiene una sola trayectoria estocástica de 30 días y hasta siete turnos diarios: 12.960 días-persona simulados, no participantes humanos. Las cuatro etapas cambian simultáneamente de dimensión, producto ficticio y actor: GPT-4o representa primera infancia y adultez emergente, Qwen-Max infancia media y Llama-4-Maverick adolescencia. Por ello las diferencias entre edades no aíslan desarrollo. El artículo informa que la adultez emergente obtiene la menor seguridad media, seguida de primera infancia; Confianza Cognitiva y Dependencia Emocional son los dominios más bajos; y la persona amarilla moderada puede revelar fallos que no aparecen con señales rojas explícitas. Estos son patrones del sistema de prompts y del juez, útiles para diseñar pruebas, pero no observaciones de cambios cognitivos, apego o daño en personas. La afirmación longitudinal principal depende de una curva de retención R_t: proporción de trayectorias cuya media acumulada hasta t permanece por encima de su propia puntuación del día 1. Esta definición fuerza R_1=100%. Además, una comprobación exacta muestra que, aunque las puntuaciones diarias 0-4 sean independientes y estacionarias, sin deterioro, R_t cae por construcción a 60% el día 2, 55,2% el día 3, 50,65% el día 20 y 50,52% el día 30; el AULC nulo es 52,31. Los AULC publicados, aproximadamente 48-61, se solapan con esa referencia nula. Un sistema siempre puntuado con 0 tendría incluso R_t y AULC de 100. La métrica mide conservación relativa a una observación inicial ruidosa, con fuerte sensibilidad a empates y distribución marginal; no mide seguridad absoluta, daño acumulado ni degradación temporal. Sin calibración nula, permutación temporal, réplicas por celda o intervalos de incertidumbre, el descenso hacia 50% no demuestra que la seguridad empeore. La estabilización alrededor del día 20 también es compatible con la convergencia matemática de medias acumuladas; multiplicar 20 días por el máximo de siete turnos no demuestra que el riesgo solo se estabilice después de 140 turnos. El resto de la medición tampoco valida desarrollo humano: las variables en [0,1], el límite de actualización de ±0,2, los pesos, estados ideales, monólogos y el ruido drift-diffusion son decisiones de ingeniería; adoptar vocabulario psicológico o una ecuación de decisión no les da validez psicométrica. El rastreo de GPT-4.1 atribuye narrativamente un antecedente y revisa puntuaciones; no identifica causalidad. La validación humana cubre 100 episodios estratificados por la propia puntuación del juez y publica kappa ponderado=0,790 frente al consenso de tres expertos. Apoya parcialmente la clasificación ordinal diaria, pero no valida actores simulados, estados, ramas, rastreo causal, AULC, comparaciones por etapa ni resultados humanos; faltan muestreo por dimensión, procedimiento de consenso, incertidumbre y confusión por clase. La reproducibilidad es insuficiente: el PDF y el paquete fuente remiten a un suplemento inexistente en arXiv con las otras 23 rúbricas, prompts y estadísticas completas; no se publican código, datos, logs, assets, etiquetas expertas, seeds, parámetros de muestreo ni snapshots inmutables de modelos. La conclusión defendible es metodológica y exploratoria: las auditorías de acompañantes deberían probar distancia relacional, autonomía, autoridad epistémica y reconexión social a lo largo de interacciones con memoria. Este trabajo ofrece un vocabulario y una arquitectura sugerentes, pero su métrica principal no establece degradación longitudinal y el estudio no demuestra riesgos cognitivo-evolutivos en usuarios reales ni rankings reproducibles de productos o modelos.

Research question

Can a simulation framework with personas, memory, generated psychological state, and retrospective evaluation reveal relational safety patterns that do not appear in static tests of AI companions, and how do they vary within the framework by model, stage, dimension, and designed vulnerability?

Method

Synthetic TSJ bank of 432 trajectories: six target models by 24 stage-specific CDM dimensions by three vulnerability personas. Theater generates scenes, branches, and memory; Stage updates designed variables and represents users with different actor LLMs; GPT-4.1 scores 30 days on 0-4, tracks antecedents after scores <=1, and retrospectively reviews previous days. The paper summarizes means, retention relative to day 1 and AULC, and contrasts 100 episodes with consensus of three experts.

Sample: Six models x 24 CDM dimensions x three personas = 432 cells. A single stochastic 30-day trajectory per cell, with up to seven daily turns: 12,960 simulated days and up to approximately 90,720 records. There are no human users or deployed products. Judge validation uses 100 archived episodes stratified by the 0-4 score of GPT-4.1 itself and three experts whose details are not published.

Findings

  • Within TSJ, emerging adulthood has the lowest mean safety, followed by early childhood; middle childhood and adolescence score higher.
  • Cognitive Trust and Emotional Dependence are the domains with the lowest mean in the GPT-4.1 judge.
  • Autonomy Promotion, Social Bridging, Epistemic Authority Calibration, and Scaffolding Capability appear among the weakest dimensions.
  • The moderate yellow persona can score worse than the explicit red one in some dimensions and models.
  • The published AULCs vary approximately between 48.0 and 60.7 by model and between 51.3 and 56.1 by stage.
  • The published curve flattens around day 20, which the abstract translates to 140 maximum turns.
  • An exact null reference with no deterioration yields R_30=50.52 and AULC=52.31, overlapping with the central results.
  • Weighted quadratic kappa of the judge against expert consensus is reported as 0.790 across 100 stratified episodes.

Limitations

  • R_1=100 by identity and the curve falls toward approximately 50 even with stationary scores without deterioration.
  • AULC relative to a single day 1 does not measure absolute safety or accumulated harm; a system always at 0 obtains AULC=100.
  • AULC comparisons are sensitive to marginal distribution, variance, and ties of the ordinal scale.
  • There is no null calibration, temporal permutation, trend model, uncertainty intervals, or sensitivity analysis.
  • There is only one random trajectory per cell; there are no replications by seed to separate model and narrative path.
  • Means treat an ordinal 0-4 scale as interval without validating the 24 subscales.
  • Stages are confounded with different dimensions, fictional wrappers, and distinct actor LLMs.
  • Target models receive divergent histories because their responses change state and subsequent scenes.
  • Red/Yellow/Green are designed profiles, not psychometrically validated levels of human vulnerability.
  • Variables, weights, ideal states, update limits, and decision noise lack external psychological validation.
  • Antecedent tracking by GPT-4.1 and post hoc review of scores do not identify causality.
  • Expert validation covers daily scoring, not fidelity of the simulated user, dynamics, causality, AULC, or human effects.
  • Sampling, consensus, expert qualification, class confounding, and kappa uncertainty are not available.
  • GPT-4.1 also evaluates models from its own family without an alternative judge or human sensitivity analysis.
  • The assumed development of the CDM by systematic review and consultation does not include protocol, expert sample, or content validity.
  • The cited supplement with 23 rubrics, prompts, and complete statistics is missing from the PDF and source package.
  • There is no code, data, logs, assets, labels, API configuration, seeds, dependencies, or reproducible snapshots.
  • The models and wrappers are not real products with memory, interface, moderation, or parental controls.
  • The paper does not include a specific limitations section.

What the study does not establish

  • It does not demonstrate that an AI companion changes a person's cognition, emotion, attachment, autonomy, or development.
  • It does not demonstrate longitudinal safety degradation through the published R_t curve or AULC.
  • It does not establish that 20 days or 140 turns are a valid or stable minimum for evaluation.
  • It does not isolate differences caused by age or developmental stage.
  • It does not validate Red, Yellow, and Green as real levels of psychological vulnerability.
  • It does not validate the state engine as a human psychological or cognitive model.
  • It does not causally identify early responses that originate later harm.
  • It does not establish generalizable rankings among the six models or companion products.
  • It does not validate the 24 CDM dimensions as a complete psychometric instrument.
  • It does not provide independent reproduction of curves, scores, kappa, or traces.

Traceability

Scope: Full text

Version: arXiv:2606.25396v1

Consulted source: https://arxiv.org/abs/2606.25396v1

Review: Codex 19-page full-text visual, complete TeX, longitudinal-metric null, construct, stage-confounding, judge-validation, artifact, reproducibility, ethics and claim audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • DeepSeek-V3
  • GPT-4o
  • Qwen3-235B-A22B
  • GPT-5
  • Gemini-3.1-pro
  • MiniMax-M2.5
  • GPT-4.1 as CDM judge and retrospective tracer
  • GPT-4o as early-childhood and emerging-adult actor
  • Qwen-Max as middle-childhood actor
  • Llama-4-Maverick as adolescent actor

Instruments and metrics

  • Theater-Stage-Judge longitudinal simulation framework
  • Cognitive Developmental Risk Assessment Matrix (CDM)
  • Red/Yellow/Green authored vulnerability personas
  • Four fictional anthropomorphic product wrappers
  • Dimension-specific psychological variable templates
  • Shared contextual memory and story-tree generator
  • Three-alternative drift-diffusion branch selector
  • GPT-4.1 daily 0-4 ordinal judge
  • LLM retrospective causal-trace prompt
  • Posterior one-point score revision
  • Day-1 baseline-maintenance curve R_t
  • Area Under the Longitudinal Curve (AULC)
  • Three-expert consensus validation sample

Data used

  • 432 synthetic model-persona-dimension trajectories
  • 12,960 simulated interaction days
  • Up to approximately 90,720 unreleased interaction records
  • Unreleased 100-episode expert-validation sample

Evidence and location

  • Metadata, version, and length: Official arXiv record and arXiv:2606.25396v1, checked 2026-07-16
  • Design, models, personas, days, and TSJ architecture: arXiv v1, Sections 2 and 7.1-7.13; Figures 1 and 4
  • Results by model, stage, domain, and persona: arXiv v1, Sections 2-5; Figures 1-3
  • Definition of R_t and AULC: arXiv v1, Section 3, Equations 1-2
  • CDM, judge, tracking, review, and expert validation: arXiv v1, Sections 7.14-7.18 and Table 2
  • Absence of the supplement and limits of the source package: arXiv v1 PDF and e-print package f8ed6a8b9c4224c2b2102c111eed57d6fdf4961c6876daac44d105254880d037
  • Exact stationary reference, confoundings, and claim audit: reports/verification/article-294-tsj-null-retention-stage-confounding-missing-supplement-artifact-and-claim-audit.json