Tailored to you: longitudinal effects of personalising language models

Key Findings

• — 992 participants completed all five daily sessions (4,960 conversations); 1,381 took the intake survey
• — Several changes in human-AI interaction over the five days were driven primarily by repeated exposure rather than by personalisation itself
• — Both personalised conditions showed significantly lower advice uptake than control at the first session (memory β=-0.34, p=0.003; survey β=-0.39, p<0.001); uptake rose across sessions in all conditions and the gap narrowed by the final session
• — In v2 (27 October 2026), memory-based personalisation showed only marginally higher self-disclosure counts than control (p_adjusted=0.052, not significant after correction), and neither personalised condition differed from control in perceived creepiness or in regret about sharing information; v1 (17 September 2026) had reported more self-disclosure and less perceived creepiness under memory-based personalisation and more regret under survey-based personalisation
• — Interacting with the control and memory-conditioned models was associated with decreased post-study comfort seeking advice from human sources; the survey-conditioned model did not show a similar decrease
• — Persona summaries for the survey condition were generated with Gemini 3.1 Pro; model instructions were revised after internal pilots to avoid emotional amplification

Methodology Notes

Between-subjects longitudinal design with three conditions and five consecutive daily advice-seeking sessions on assigned topics; Prolific sample paid $20 per hour; preregistration status not stated on the abstract page. Outcomes are self-report scales plus conversation-derived self-disclosure measures; effects on interpersonal relationships outside the interaction are framed as motivation rather than measured. All nine authors are affiliated with Google DeepMind, the developer of the models studied. Date is the arXiv v1 submission date (17 September 2026); announced 21 September 2026; CC BY. Revised 27 October 2026 (arXiv v2): v2 removes the v1 abstract sentence on self-disclosure, creepiness and regret, and its results report those effects as non-significant after correction; the sample, the repeated-exposure finding and the advice-uptake result are unchanged. Key findings follow v2, read from the arXiv HTML on 2026-10-29.

Sources