Does personalization actually build closeness with chatbots?

In the same preregistered 5-day RCT (N=992, 4,960 conversations), many of the changes observed over the course of repeated relationship-advice conversations were driven by time in the study rather than by which personalization condition a participant was assigned to. The authors report that "in all conditions, participants perceived the chatbot as increasingly competent and useful over successive sessions, and reported growing closeness to it" — a pattern that held across the non-personalized control, the memory-based condition, and the survey-based condition alike. They state plainly that "several changes in human-AI interaction over time are driven primarily by repeated exposure rather than personalisation itself."

Critically, personalization produced null effects on several outcomes the authors expected it to move: "relative to participants in the control condition, participants in the memory and survey conditions did not experience more closeness to the model over time; they did not accept more incorrect AI suggestions [in a post-study overreliance task]; nor did they self-report greater depth or breadth in their self-disclosures." The authors read this as evidence that closeness and perceived competence are largely functions of session count and familiarity with the interaction format itself, rather than of any personal information the model held about the user.

This runs parallel to Do chatbot relationships lose their appeal as novelty wears off?, which found relationship-formation processes with Mitsuku wore off as novelty faded — both studies treat session number as doing real explanatory work that one-shot designs miss, though in opposite directions: Mitsuku's relational signals declined while this RCT's closeness and competence ratings rose. It also qualifies Does chatbot personalization build trust or expose privacy risks?, which attributes rising trust and closeness over time to personalization's "expectation ratchet" — this RCT's control condition shows the same closeness gains with no personalization at all, suggesting repeated exposure alone, not the ratchet mechanism specifically, may account for a substantial share of that trajectory. See also Does personalization type shape how people feel about sharing? for where personalization's effects, in this same study, were not null.

The excerpt does not report effect sizes or statistical tests for the session-number trends, and the authors explicitly caution that "our study does not support causal claims about the relationship between" perceived closeness and disclosure, proposing cross-lagged panel designs as a future test. The implication the evidence supports is that researchers and designers attributing longitudinal gains in user trust or closeness to personalization specifically should rule out a simpler explanation first: mere repeated exposure to the interaction, independent of what the system remembers or infers about the user.