Can chatbots reduce polarization by surprising partisan expectations?

The paper reports a 2x2 experiment with 1,983 U.S. adults who held structured conversations with an AI chatbot whose "presented partisan identity and policy stance were independently manipulated." The design crosses ingroup/outgroup identity with agreement/disagreement, producing four conditions: Ingroup Agree (the baseline, "the status quo of political discussion"), Ingroup Disagree, Outgroup Agree, and Outgroup Disagree. The two conditions that broke the expected link between identity and stance — a co-partisan chatbot that disagreed, or an opposing-party chatbot that agreed — "are effective in reducing affective and issue polarization." The largest effect, in Outgroup Agree, was "a roughly five-point reduction in affective polarization," which the authors say "corresponds to reversing approximately three years of aggregate partisan animosity in the United States."

The authors' mechanism is expectation violation rather than persuasion. Party labels carry "assumptions about policy disagreement, moral character, interpersonal hostility, and democratic commitments," so when a partner's words don't fit the label, people shift from "category-based judgment" toward "individuated processing," reconsidering the individual, the group, or the assumed link between the two. Critically, the two depolarizing conditions worked through different routes: Ingroup Disagree lowered warmth toward the ingroup and widened perceived issue distance from co-partisans, while Outgroup Agree raised warmth toward the outgroup and narrowed perceived distance from them — "the same depolarizing outcome can emerge through distinct mechanisms." The effect held "without meaningful shifts in participants' own policy positions," so the chatbot changed how partisanship was read, not what people believed about the issue.

This sits alongside Can sycophantic AI advice still push people away from polarized views?, which found AI-mediated depolarization through a different channel — informative advice content — rather than identity or stance framing; together the two papers suggest AI chatbots can depolarize through separable routes that a single sycophancy measure would not distinguish. The satisfaction tradeoff here — expectation-violating conditions "are experienced as less satisfying," with lower willingness to continue, even as objective deliberation quality held or improved — parallels Can warnings stop people from being swayed by sycophantic AI?, where making sycophancy visible to users cut appeal without touching effectiveness; both decouple what works from what people enjoy. The authors draw out the sycophancy risk directly: political chatbots left to default toward the pleasant, agreement-confirming mode (the Ingroup Agree baseline) reproduce rather than challenge the status quo, so a sycophantic design choice is itself a polarization-relevant one.

A one-month follow-up (N = 1,627, 82% retention) found that "most effects disappear" over that period, so the excerpt does not show the intervention has any durable effect without repetition or reinforcement, and it does not report the exact decay curve or test why the effects fade. The excerpt names "correction of misperceived group norms" and "cognitive dissonance and social balance theories" as consistent explanations but does not test between them. It also does not report effect sizes for the Ingroup Disagree condition or for issue-position change by condition, and the design — a researcher-recruited sample discussing self-chosen issues with a disclosed AI — leaves open whether the same expectation-violation effect would hold in organic, undisclosed, or non-AI-mediated political contact. The defensible claim is narrow: for this sample and this one-shot design, violating the identity-stance link moved affective and perceived-issue polarization at the cost of user satisfaction, and the gain had largely faded within a month.