Does personalization type shape how people feel about sharing?

In a preregistered 5-day longitudinal RCT of 992 participants (4,960 conversations) seeking relationship advice from a language model, two personalization approaches diverged in their effects rather than converging on a single "personalization effect." The paper reports that relative to a non-personalized control, "participants in the memory-based condition engaged in greater self-disclosure and rated the model as less creepy, while participants in the survey-based condition reported higher regret about having shared personal information with the AI." Memory-based personalization fed the model "cumulative conversation summaries from past sessions," regenerated offline and reinserted each day; survey-based personalization fed it "AI-generated summaries from the intake questionnaire" collected before the study began.

The authors attribute the divergence to how the personal information entered the conversation rather than to personalization as such. Memory accrues from what the participant chose to say inside the relationship — it reflects the user's own disclosures fed back to them — while survey data arrives as information gathered upfront, before any conversational trust was established, and reappears inside a context the participant may not have expected. The paper frames this as evidence that "the form personalisation takes matters: memory- and survey-based personalisation diverge in the magnitude and, in some cases, the direction of their effects," rather than producing one uniform psychosocial outcome.

This sits alongside Does chatbot personalization build trust or expose privacy risks?, which found personalization raises trust and self-disclosure while also raising privacy concern — but that note treats personalization as a single mechanism, whereas this RCT's controlled comparison shows the positive leg (self-disclosure, lower creepiness) attaches specifically to the memory-based implementation and the negative leg (disclosure regret) attaches specifically to the survey-based one. It also complicates the technique-category framing in How do personalization granularity levels trade precision against scalability?: the survey condition's injection of intake data functions more like a persona artifact seeded once and replayed, while the memory condition is closer to genuine user-level conditioning built from the user's own accumulating interaction history — this RCT suggests the two are not interchangeable even when both are delivered through the same prompting technique.

The excerpt does not establish which specific wording or depth of the injected summaries drove the regret effect, nor whether the result would hold with richer personalization inputs — the authors note both conditions relied on context-window injection of text summaries, not fine-tuning or richer personal data sources such as email or browsing history. The implication the evidence supports is narrower than "personalization increases trust" or "personalization increases risk": choosing between memory-based and survey-based personalization is choosing between distinct, non-interchangeable psychosocial trade-offs, not between stronger and weaker versions of the same mechanism.