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

Key Findings

• — Short-horizon (single-turn or short-session) testing systematically underestimates developmental risk; a stable risk estimate emerges only after roughly 140 interaction turns.
• — Early childhood and emerging adulthood are identified as the most vulnerable developmental stages.
• — Cognitive trust and emotional dependency are the weakest safety domains across the evaluated models.

Methodology Notes

Preprint (arXiv 2606.25396, cs.AI; submitted 2026-06-24). Simulation-based evaluation: persona-driven user simulation, dynamic psychological-state updating, ~12,960 simulated person-day interactions across six models, four developmental stages, and 24 risk dimensions. Title, authors, and date verified via the arXiv abstract page and the arXiv API. Fresh, not yet peer-reviewed — credibility set conservatively to preliminary.

Sources