Long-Term Simulation Exposes Cognitive-Developmental Risks in AI Companions
06:00 · June 25, 2026 · arXiv cs.AI RSS
AI companions powered by large language models increasingly interact with cognition-developing users, including children and adolescents, creating risks that may accumulate over time. Existing safety evaluations largely rely on single-turn or short-session tests, which cannot capture risks that emerge only through prolonged interaction. To address this gap, we propose TSJ (Theater-Stage-Judge), a longitudinal framework combining persona-driven user simulation, dynamic psychological-state updating and retrospective evaluation. We evaluate six mainstream models across four developmental stages, twenty-four risk dimensions and three psychological-vulnerability personas, covering 12,960 simulated person-day interactions. TSJ shows that short-horizon testing systematically underestimates developmental risks, for which TSJ yields a stable risk estimate only after 140 turns within prolonged simulated relationships. Applying TSJ further identifies early childhood and emerging adulthood as the most vulnerable stages, with cognitive trust and emotional dependency as the weakest domains. TSJ provides a scalable methodology for longitudinal cognitive developmental risk evaluation in AI companion systems.
Summary
AI companions built on large language models now engage children and adolescents in open-ended, persona-consistent dialogue that can span weeks or months. Because these users are still forming cognitive structures, self-boundaries and social capacities, risks such as emotional dependency or distorted cognitive trust can accumulate gradually rather than appear as isolated policy violations. Conventional safety benchmarks, which rely on single-turn or short-session tests, miss these trajectories and therefore tend to overestimate long-term safety.
To expose the hidden risks, the authors introduce TSJ (Theater-Stage-Judge), a longitudinal evaluation framework that combines persona-driven user simulation, daily psychological-state updates and retrospective scoring. The framework operationalizes a Cognitive Developmental Risk Assessment Matrix covering four age bands—early childhood (3–6), middle childhood (7–13), adolescence (14–18) and emerging adulthood (19–29)—and twenty-four stage-specific dimensions grouped under reality perception, cognitive trust, emotional dependence, socialization, values and behavioral safety. Across six mainstream models, three vulnerability personas and 30-day simulated trajectories, the study generated 12,960 person-day interactions.
Results show that short-horizon testing systematically underestimates developmental risk; stable estimates require at least 140 interaction turns. Early childhood and emerging adulthood emerge as the most vulnerable periods, with cognitive trust and emotional dependency registering the weakest safety scores. The work demonstrates that longitudinal, persona-conditioned simulation can surface risks that remain invisible to static or adult-centered benchmarks, offering a scalable method for auditing AI companion systems before deployment.
Why it matters
This research is highly relevant to the Dutch AI market's strong emphasis on ethical, transparent, and safe AI. The proposed longitudinal evaluation framework provides researchers and developers with actionable methodologies to align AI companions with strict EU regulations regarding vulnerable populations.
