Tailored to you: longitudinal effects of personalising language models

This study investigates the longitudinal effects of personalising language models on human-AI interaction. Researchers recruited 992 participants to engage in daily relationship advice conversations with an AI over five days. The study compared three conditions: a non-personalised control, a memory-based condition (where the AI was conditioned on previous conversation summaries), and a survey-based condition (where the AI was conditioned on pre-study intake survey data). The researchers measured outcomes such as perceived closeness, advice-seeking behavior, self-disclosure, and decision regret.

Many changes in human-AI interaction, such as increased perceived closeness, competence, and usefulness, were driven by the duration of exposure rather than the personalisation method. However, personalisation did influence specific behaviors. Participants in the memory-based condition engaged in higher levels of self-disclosure and rated the model as less creepy compared to the control. Conversely, participants in the survey-based condition reported higher regret about sharing personal information. Additionally, while interacting with control or memory-based models was associated with decreased comfort seeking advice from human sources, this effect was not observed in the survey-based condition.

These findings suggest that personalisation is not a monolithic feature; different technical implementations produce distinct psychosocial outcomes. The results highlight a tension between the perceived benefits of personalisation and potential risks like information-sharing regret. For developers, this underscores the need for nuanced design, such as providing users with easy ways to redact or manage the personal data that informs model behavior, ensuring that AI systems remain helpful without undermining human social norms or user well-being.