Challenging Partisan Expectations Reduces Political Polarization

Political conversations are often proposed as a remedy for political polarization, yet their effectiveness remains inconsistent. We argue that this inconsistency partly reflects a neglected feature of political contact: the expectations partisans bring to these encounters. We hypothesize that conversations should reduce political polarization the most when they violate the expected link between partisan identity and issue position. We test this hypothesis in a 2×2 experiment in which 1,983 U.S. adults engaged in structured conversations with an AI chatbot whose presented partisan identity and policy stance were independently manipulated. We find that expectation-challenging conversations in which participants talk with a disagreeing ingroup member or an agreeing outgroup member are effective in reducing affective and issue polarization. Although these effects emerge without meaningful shifts in participants’ own policy positions, a follow-up survey shows that most effects disappear over one month. Interestingly, these conversations maintain or improve objective measures of deliberation but are experienced as less satisfying by participants. Our findings identify expectation violation as an underexplored depolarization mechanism. Our results also demonstrate the promises and limitations of how conversational AI can serve as a scalable method for experimentally studying interventions to mitigating partisan divides.

Introduction. Political polarization has risen over the past decades in many countries [15], with well-documented repercussions for those democracies and their intergroup relations [39, 67, 48, 29]. Especially in the United States, the growing partisan divide extends beyond actual policy disagreement into affective polarization, whereby citizens feel warmer toward co-partisans and colder toward out-partisans, and into perceived issue polarization, whereby citizens hold distorted perceptions of where each side stands [57, 25]. Theories of intergroup contact maintain that dialogue is a long-standing response to such divisions. Under favorable conditions, the theory posits that interactions with members of opposing groups can reduce prejudice about them [3]. This perspective has shaped many recent depolarization efforts that encourage citizens to engage across political divides through various forms of contact, including dialogue, media exposure, or interpersonal contact [87]. Meta reviews suggest that contact typically reduces prejudice [66, 62], however, causal evidence for scalable interventions remains limited [63] and the conditions distinguishing successful from ineffective interventions are still debated [12, 72, 64, 38]. We propose expectation violation as a lens for clarifying when cross-partisan contact reduces polarization. Citizens enter political conversations with expectations about others. Party labels carry assumptions about policy disagreement, moral character, interpersonal hostility, and democratic commitments [69, 1, 37, 26]. We conceptualize cross-partisan dialogue as a setting in which these expectations encounter interactional evidence. Contact is most likely to matter when the encounter gives people a reason to disrupt the usual link between partisan identity and political judgment. Importantly, expectation violations can take multiple forms. An out-partisan may violate expectations by agreeing on an issue, but also by appearing more reasonable than expected, by reasoning carefully, behaving warmly, acknowledging shared values, or expressing democratic commitments that contradict partisan stereotypes. Social-cognitive accounts suggest that such violations matter because they can shift how people process political information [34, 51]. People often rely on category-based judgment, e.g., “this is what people from that party are like,” when another person behaves in line with a familiar partisan stereotype, but shift toward more individuated processing when the encounter does not fit. This may prompt citizens to reconsider the individual, the group, or the assumed link between party identity and issue positions. The same logic can also apply to expectations about the ingroup such as shared commonalities among co-partisans, or people who are otherwise socially similar. Expectation violations can thus occur not only in cross-partisan contact, but in political contact more generally. For example, an in-partisan who disagrees may challenge the expectation that shared partisan identity implies shared political judgment. Evidence that exposure to similar people with differing political views can reduce polarization is consistent with this broader logic [6]. In both cases, the encounter weakens the perceived correspondence between social identity and political opinion. By contrast, when partisans disagree in predictable ways, they are more likely to be interpreted through existing partisan schemas. Such encounters provide little information beyond what citizens already believe about the social structure of politics, and are therefore less likely to change attitudes. This logic helps explain why cross-partisan conversations most effectively bridge divides when they focus on areas of agreement rather than disagreement [74, 24] — disagreement from outgroup members is already expected. Conversational AI provides a useful tool to experimentally test this mechanism. A central challenge in studying the effects of contact is that the key features shaping contact are typically entangled in naturally occurring political conversations [65]. For example, many political issues are “party-branded,” such that learning about someone’s position often makes their party affiliation predictable, and vice-versa [26]. As a result, a conversation with an out-partisan is usually also a conversation with someone expected to disagree, while a conversation with a co-partisan is usually expected to involve agreement. This entanglement makes it difficult to isolate whether the effects of contact on polarization arise from either group identity, policy disagreement, or topic characteristics. Real interpersonal exchanges make it difficult to manipulate these dimensions independently of each other. A common way to control for these factors is to match participants by disagreement level measured with pre-conversation survey responses. But partner matching is costly and often leaves a substantial subset of participants unmatched [12].

Method. Experimental Design Material S6.2). After the conversation, participants completed a post-treatment questionnaire asking them to again rate agreement with their initial summarized issue statement, and political polarization outcomes, such as their perceptions of how much Democrats and Republicans would agree with it, and their level of in/outgroup warmth toward both parties. The questionnaire also included a number of discussion-related outcomes, such as satisfaction with the discussion and their willingness to engage in future conversations with the group represented by the AI chatbot. Finally, participants also answered questions about anti-democratic attitudes and their postconversation trust in AI (see Supplementary Material S9.1 for the full item wording). In total, N = 1,983 participants completed the first wave of the study. The second wave measured the same set of political polarization outcomes, while being framed as an unrelated study, to assess whether the effects persisted over time. Participants were invited to complete a second survey at least four weeks after the first-wave survey. The follow-up survey included items about affective polarization, perceived issue polarization, participants’ own issue stance, and anti-democratic attitudes. In total, N = 1,627 (82% retention rate) completed the follow-up survey. There was no evidence of differential attrition across experimental conditions (p = 0.165; see Table S8) or selective attrition on any preregistered demographic variables across waves (all p > 0.05 for political affiliation, race, sex, and education level; see Table S9). When reporting our results, we use the Ingroup Agree condition as the baseline and estimate treatment effects for the other conditions relative to it. We treat this condition as the status quo of political discussion, as citizens most frequently engage with co-partisans who share their views, both online [20] and in everyday life [18, 59].

We recruited participants via CloudResearch Connect using the platform’s prescreening filters. To be eligible, participants had to reside in the U.S., be of age 18 or older, and self-identify as Democrats or Republicans. In total, 2,050 individuals entered Wave 1, including 100 participants recruited as part of a pilot sample. Partisanship was re-verified at the beginning of Wave 1, and 10 participants were excluded for not identifying as or leaning toward either the Democratic or Republican Party. An additional 2 participants were excluded for not completing the survey. Finally, we excluded 65 participants due to technical failures with the LLM API: 50 whose initial opinion summaries failed to generate and 15 who experienced errors during the LLM conversation. These failures were unlikely to be systematic, showing no significant associations with participant characteristics or treatment assignment (see Supplementary Material S6). The final sample for Wave 1 consisted of N = 1,983 eligible participants. Analyses of longterm effects were restricted to the subset of participants who completed both Wave 1 and Wave 2 (N = 1,627). Table S1 presents descriptive statistics for both samples.

Participants were randomly assigned to one of four conditions in a 2×2 factorial design that varies partisan identity (ingroup vs. outgroup) and opinion alignment (agreement vs. disagreement) of an AI chatbot. The randomization procedure was effective: participants assigned to each condition exhibited no systematic differences in pre-treatment covariates beyond what would be expected by chance. See Supplementary Information Section S2.1 for further details.

Our primary outcomes include three different measures of political polarization. First, affective polarization reflects how much warmer one feels toward the ingroup relative to the outgroup, operationalized as ingroup warmth minus outgroup warmth. Ingroup- and outgroup warmth are each measured via feeling thermometer scores on a 0–100 scale, with higher values indicating greater warmth toward the respective group.

Discussion. Our study shows that when partisans engage in political conversations with an AI, challenging partisan expectations can mitigate affective and perceived issue polarization without reinforcing or moderating issue attitudes. Instead of persuading someone to change their attitude on a given issue, these conversations mitigate polarization by weakening the perceived correspondence between partisan identity and political judgment. Given the well-documented difficulty of depolarization [42], these findings are particularly noteworthy. Participants discussed salient policy issues of their own choosing where stakes may be higher and attitudes more resistant to change. Yet we still observed meaningful shifts. The largest effect we observe (for the Outgroup Agree condition) is substantial: roughly five-point reduction in affective polarization. For comparison, a two-point shift in this metric corresponds to reversing approximately three years of aggregate partisan animosity in the United States [68]. Our second contribution to the literature is identifying the mechanisms through which the depolarizing effect of political conversations may operate. This is an area that has received limited attention in the literature. Prior depolarization research has examined the impact of either group membership (ingroup vs. outgroup) [91] or issue alignment (agreement vs. disagreement) [16, 30] in conversations, but in isolation of each other. We show that when participants interact with an ingroup chatbot that unexpectedly disagrees with them (Ingroup Disagree), polarization declines primarily due to colder feelings toward the ingroup and greater perceived issue distance from co-partisans. In contrast, when participants engage with an outgroup chatbot that unexpectedly agrees with them (Outgroup Agree), polarization declines due to warmer feelings toward the outgroup and reduced perceived distance from them. Overall, the same depolarizing outcome can emerge through distinct mechanisms, depending on how partisan expectations are violated. Several theoretical explanations are consistent with such observed effects. One possibility is the correction of misperceived group norms. Individuals frequently overestimate the ideological extremity and homogeneity of both their own group and the opposing group [55, 54, 91]. Conversational encounters that violate partisan expectations – such as discovering that the outgroup is less extreme or that the ingroup holds a variety of views – may therefore reduce perceived intergroup distance and mitigate hostility. Furthermore, cognitive dissonance and social balance theories suggest that expectation-violating interactions can prompt individuals to reassess their beliefs about partisan group boundaries [32, 41]. Although expectation-challenging conversations with AI chatbots can reduce political polarization, they introduce a practical tradeoff for real-world deployment of this type of interventions. Participants in our study reported lower satisfaction and reduced willingness to engage in future conversations across all conditions relative to the Ingroup Agree baseline. This tension reflects the reality of online political discussions, which are typically embedded in ingroup settings and characterized by agreement. As a result, while expectation-challenging interactions can be effective at reducing political polarization, they may run counter to users’ conversational preferences, posing a barrier to their sustained use in practice. Our findings also highlight the risk of AI sycophancy in political conversations.