People are gradually forming relationships with AI systems. Emergent cognitive and behavioural risks are left out of the European AI Act, while China has just enacted a new pioneering regulation on anthropomorphic AI. Today, AIOLIA published a policy brief on AI companions that summarizes our previous empirical findings. We have four recommendations: 1) dynamic identification of vulnerable users, 2) longitudinal oversight mechanisms, 3) engineering friction by design, and 4) extending transparency beyond simple disclosure requirements.
