Can AI chatbots reduce partisan misperceptions and warm cross-party feelings?

Across five preregistered studies (total N = 3,960 U.S. partisans), brief conversations with AI chatbots prompted to represent the political outgroup — what the authors call "synthetic contact" — lowered the barrier to cross-partisan engagement and then corrected the misperceptions that sustain it. In an incentive-compatible choice task (Study 1, N = 608), partisans "would endure almost twice as long contemplating their own mortality to avoid a human outgroup partner as an AI one," showing AI partners are far more acceptable than real ones. In a within-person study (Study 2, N = 500: 248 Democrats, 252 Republicans) using the GREEN environmental-attitudes scale, Democrats had misjudged Republicans by more than a standard deviation (d = 1.49) — "enough to flip the average Republican from supportive to opposed" — and a single ten-minute chat with an outgroup-representing bot corrected that misperception and raised outgroup warmth; the size of belief correction predicted the size of the warmth gain.

The authors isolate the mechanism in two ways. A three-arm experiment compared the outgroup bot against an equally polite apolitical chatbot and a non-social game, "isolating the causal contribution of outgroup-specific content" and ruling out pure engagement or sociality as drivers. Coding every conversation on four dimensions — stereotype-disconfirming substance and informational specificity (a cognitive route) versus empathy and friendliness (an affective route) — found that "information, more than friendliness, distinguishes outgroup bots from control chatbots." The bots were not perfectly calibrated: the Republican-representing bot scored 3.084 against a Republican mean of 3.728 (d = 0.61), and the Democrat-representing bot scored 4.577 against 4.149 (d = 0.56); but both landed closer to the outgroup's real position than partisans' own guess did, which is what the authors argue makes an imperfect guide still useful. The effect also reached a costlier test on immigration, a more affectively charged issue: participants who talked with an outgroup bot were six percentage points more likely than controls to choose a real cross-partisan conversation afterward.

This sits apart from the vault's persuasion literature. Where does AI's persuasive power actually come from? finds persuasiveness and factual accuracy trade off against each other in political messaging; synthetic contact's effect here instead runs through correcting a factual misperception, since the bot's stated position, imperfect as it was, still sat closer to the outgroup's actual attitudes than partisans' prior belief did, rather than through a persuasion technique. It also differs in kind from Can a simple warning reduce how much LLMs persuade people?, which is about defending against LLM-driven belief change; here the "synthetic contact" is not a persuasion attempt to be warned against but a stand-in for contact partisans normally avoid, and the belief change it produces moves partisans' estimates toward the outgroup's actual position, not toward the bot's own exaggerated stance. It also contrasts with Do AI writing tools improve online discussion or degrade it?, where AI tools traded engagement for perceived discussion quality; here engagement with an AI stand-in instead raised willingness to have a real, presumably higher-quality, cross-partisan conversation.

The excerpt does not establish durability: "most of the warmth effect fades within a week," with only "a small residual concentrated among the most extreme partisans" — the group the intervention most needs to reach. Nor does it establish generalization beyond online U.S. partisans and the two issues tested (environment, immigration), or beyond outgroup warmth on a feeling thermometer to social distance, trait attributions, or support for anti-democratic action. The belief-accuracy results come from a within-person design and so warrant more caution than the experimentally isolated warmth and behavioral effects. The warranted implication is narrow: brief, repeatable synthetic contact can nudge people toward a real cross-partisan conversation right afterward, not that it durably resolves affective polarization on its own.