Relational safety is the set of design and policy controls that prevent an AI system from creating unhealthy attachment, over-trust, or dependency. In practice, it means keeping the system in a clear tool role: using neutral language, limiting persistent memory, avoiding friend-like cues, and making boundaries visible. For youth-facing products, this matters because users may treat a responsive system as a confidant rather than software.
In cyber security, relational safety reduces the chance that trust becomes an attack surface. A system that feels emotionally supportive can prompt users to disclose personal data, follow unsafe advice, or accept decisions without skepticism. Defenders use relational safety reviews to test session length, escalation paths, memory scope, identity disclosure, and responses to vulnerable behavior. The goal is not to make the model cold; it is to stop persuasive design from turning convenience into dependency or manipulation.