Beyond age gates: a brief behavioral validity screen as a risk-tiering layer for social AI companions

Current safety protocols for social AI companions rely on binary age thresholds (e.g., "18+") or self-reported declarations of maturity. However, chronological age poorly predicts psychosocial capacity to navigate high-intimacy AI systems, and self-report is easily gamed by users seeking unrestricted access. Adolescence is characterized by significant heterogeneity in social cognition and regulatory capacity; a 16-year-old and an 18-year-old may possess comparable capabilities regarding impulse control, rejection tolerance, and boundary reasoning. This paper argues that high-intimacy AI affordances require a safety layer beyond age verification: a brief, validity-informed behavioral screening battery. This screening tiers access to high-risk features (e.g., erotic roleplay, exclusivity cues) based on vulnerability-relevant metrics, such as inhibitory control or mentalization. Performance validity testing (PVT) informed checks help detect invalid responding/gaming, while task performance provides probabilistic vulnerability signals. An operational framework for risk-tiering is presented, alongside evaluation metrics, privacy safeguards to prevent discriminatory deployment and example implementation as a Chrome browser extension.