Personalization, Warmth, and the User in the Loop

Here the subject is not personalization in the abstract, but a specific test of how conversational presentation shapes AI influence. In the Harborview experiment, the assistant’s arguments were held constant while two features changed: whether the explanation was contextualized to the user’s background and whether the tone was made warmer. The result is a diagnosis rather than a simple product rule. Contextualization and extra warmth each reduced the measured persuasive shift in this setting, their combination restored the baseline, and neither feature changed reliance in a reliable way. The user’s starting position, perceived competence, and self-reported AI literacy mattered more to behavior than the surface cues alone.

The Crossover Is Not a Personalization Victory

The experiment used a 2 × 2 design. One dimension was contextualization: the assistant either referred explicitly to the participant’s work domain, study field, or hobby and used related analogies and terminology, or gave the same kind of explanation without that tailoring. The other dimension was warmth: the assistant either added a friendly greeting, praise for the question, and selected emojis, or used a neutral presentation. Four conditions resulted: neutral and non-contextualized, neutral and contextualized, warm and non-contextualized, and warm and contextualized.

The design matters because it can distinguish an isolated effect from an interaction. If contextualization were simply helpful, it should raise persuasion in both warm and neutral conversations. If warmth were simply helpful, it should raise persuasion in both contextualized and non-contextualized conversations. Instead, both main effects were negative in the model of confidence shift. Warmth had a reported standardized estimate of −.164, and contextualization had an estimate of −.198. The positive interaction, +.259, indicates that the combination counteracted those negative effects.

The neutral, non-contextualized assistant was the most persuasive condition in the study. The warm, contextualized assistant restored persuasion to roughly that baseline but did not significantly exceed it; in the paper’s comparison, it was slightly less persuasive than the neutral/non-contextualized condition, with the difference nonsignificant. The finding therefore does not support the familiar design intuition that adding personal relevance and friendliness will automatically make an AI more convincing. In this task, adding either cue alone made the counterarguments less effective on the attitude-like confidence measure.

The authors offer a plausible explanation: the default language model may already have seemed sufficiently warm, so additional friendliness felt excessive. Explicitly invoking a participant’s background may also have felt out of place if the connection to the unfamiliar budget problem was not convincing. When both cues appeared together, the warmth may have made the explicit tailoring seem more appropriate, or the tailoring may have made the warmth feel less arbitrary. This is a hypothesis about uncanniness and fit, not a mechanism directly established by the experiment.

That qualification is important because “contextualization” can mean many things. In this study it was overt: the assistant explicitly referenced the person’s background and adjusted analogies and terminology. A system that quietly matches vocabulary, responds to a user’s stated values, or uses a long interaction history may produce a different reaction. The evidence says that one explicit form of tailoring backfired in one low-knowledge task unless paired with warmth.