When personal AI creates a multi-agent problem

Abstract

Personal AI systems are deployed and evaluated around individual users. Yet the people using them are often connected. Advice from one assistant can change a person’s behaviour, alter what another person experiences and, in turn, influence what that person’s assistant recommends next. The systems need not communicate or even know of each other. Their users provide the connection. This paper examines human-mediated AI coupling and the problem it creates for evaluation. Two providers may each correctly assess their own system while overlooking a consequential condition of ordinary deployment: both systems operating within the same human relationship. The paper distinguishes interpersonal interference and joint intervention effects from directed and recursive AI coupling, and considers when evidence gathered at an individual AI–user boundary is sufficient for a specified outcome. The argument is conceptual. It does not claim that such coupling is necessarily harmful or establish how frequently it occurs.