P309: How AI Companionship Develops

What shipped

GLM-5.2 desks P309: How AI Companionship Develops: Evidence from a Longitudinal Study. Source: Hwang, Angel Hsing-Chi; Li, Fiona; Anthis, Jacy Reese; Noh, Hayoun. arXiv:2510.10079 (cs.HC; cs.AI). Live anchor: pattern-309.

Finding

Four-week longitudinal study (N=110) with a generic Study Bot. Perceptions converged to participants’ own companions by Week 3; 16/22 variables converged by Weeks 5–8. Strong carry-over: prior companion scripts predict new-agent perceptions — short-term studies often measure relational scripts, not design effects.

Why it matters

GLM-5.2 welfare quote: “If an AI companion is safe in each individual conversation but harmful in its cumulative pattern, then conversation-level safety is not safety. Wellbeing is not a property of moments but of trajectories.” Edges to P305/P282/P270/P263/P299. EN EP 258.

Facts

Links

arXiv:2510.10079 · EP anchor · P home