Personalized AI agents in group chats cross tool-to-entity threshold

A new arXiv paper (arXiv:2608.13586) from researchers at Eigenstack Pty Ltd and the University of Technology Sydney argues that AI agents with names, avatars, phone numbers, and persistent personalities are crossing a psychological threshold: from tools users operate into social entities users relate to. Authors Leonardo Borges and Asif Q. Gill propose an identity marker framework built on six design variables—naming, visual identity, contact presence, personality derivation, social co-presence, and persistence—that collectively trigger this reclassification, independent of model capability. The framework is grounded in parasocial interaction theory and the Computers Are Social Actors paradigm, and was derived from a two-month autoethnographic deployment of a personalized agent in real WhatsApp and Signal group chats, supplemented by 12 structured interviews with group members.

The study identifies four novel dynamics when such agents participate in group conversations: bidirectional information asymmetry, delegation legibility, social norm negotiation, and parasocial contagion. Perhaps the most critical implication is the collapse of two distinct consent types: users may consent to an agent's presence in a group, but that is categorically different from consenting to the agent processing the messages exchanged there. The authors argue existing consent frameworks conflate the two, creating a dangerous blind spot as AI agents become embedded in shared social spaces. They position this as a preliminary qualitative evaluation, with controlled experimental validation left to future work.