Interest in personalized language models is rising rapidly. While personalization is often seen as a way to meet diverse user needs, the long‑term effects of sustained interaction with personalized models on users' perception and behavior toward AI remain poorly understood. In particular, downstream consequences beyond the immediate human‑AI loop, such as impacts on self‑perception and interpersonal relationships, have been largely overlooked.
In this study, 992 participants completed a daily advice‑seeking conversation over five consecutive days. Three conditions were compared: a non‑personalized baseline, a memory‑based personalization that conditioned on prior conversational history, and a survey‑based personalization that conditioned on information gathered from a pre‑study intake questionnaire. After each interaction participants reported their attitudes and behavioral changes.
The findings reveal that most changes in human‑AI interaction over time are driven by repeated exposure rather than personalization itself. Nevertheless, personalized models produced distinct effects. Participants interacting with the memory‑based model disclosed more personal information and rated the model as less creepy, whereas those in the survey‑based condition expressed higher regret about having shared personal data with the AI.
These results highlight that different personalization approaches yield nuanced outcomes, underscoring the need for responsible design and deployment of personalized AI systems that account for privacy concerns, information‑sharing willingness, and long‑term user experience.
Review: This large‑scale longitudinal study offers valuable empirical insight, reminding developers that the pursuit of personalization must be balanced with careful consideration of its psychological and social ramifications.