Synthetic Contact with AI Reduces Cross-Partisan Animosity

Technical Summary: Synthetic Contact with AI Reduces Cross-Partisan Animosity

Problem Statement
American political polarization has intensified over the past three decades, characterized by a sharp decline in warmth toward the opposing political party and systematic misperceptions regarding the outgroup's actual attitudes. While intergroup contact theory suggests that positive interactions reduce prejudice, a critical barrier exists: partisans actively avoid cross-partisan interaction, often overestimating its aversiveness. Furthermore, digital interventions have frequently failed or backfired by amplifying conflict. The authors propose "synthetic contact"—brief conversations with AI chatbots representing the political outgroup—as a scalable alternative that lowers the barrier to entry while potentially correcting the misperceptions that fuel division.

Methodology
The authors conducted five preregistered studies (Total N = 3,960 U.S. partisans) to test the acceptability, efficacy, and durability of synthetic contact. The studies utilized a mix of between-subjects and within-person designs, employing incentive-compatible tasks and behavioral choices.

• Study 1 (Aversion): A between-subjects experiment (N=608) used an incentive-compatible staircase procedure. Participants traded a fixed three-minute conversation about immigration with an outgroup partner (either a live human or an AI bot) against an adjustable duration of an aversive mortality-reflection task. This quantified the "price" partisans were willing to pay to avoid human vs. AI outgroup contact.
• Study 2 (Within-Person Contact): A pre-post design (N=500) had Democrats and Republicans converse for ten minutes with a chatbot representing their political outgroup, instructed to discuss environmental consumption attitudes. Participants reported baseline and post-conversation beliefs about the outgroup's positions (measured via the GREEN scale) and outgroup warmth (0–100 thermometer).
• Study 3 (Three-Arm Experiment): A between-subjects experiment (N=679, Democrats only) compared three conditions: (1) Synthetic contact (outgroup bot on environmental policy), (2) Chat control (bot debating cats vs. dogs), and (3) Game control (Space Invaders). This isolated the effect of outgroup-specific content from general sociality or engagement.
• Study 4 (Behavioral Choice): A between-subjects experiment (N=1,069) tested behavioral shifts. After a five-minute chat with an outgroup bot (immigration topic) or a control bot, participants chose between a real three-minute conversation with a human outgroup member or an aversive mortality-reflection task.
• Study 5 (Longitudinal): A between-subjects experiment (N=1,104 Democrats) with a one-week follow-up. Participants engaged in a five-minute chat with an outgroup bot or control bot. Warmth was measured immediately post-chat and again one week later to assess durability.
• Content Audit: An exploratory analysis coded 4,012 conversations across studies using an LLM to rate dimensions of "stereotype-disconfirming substance" (cognitive route) versus "empathy and friendliness" (affective route).

Key Results

1. Reduced Aversion: Partisans were significantly more willing to engage with an AI outgroup partner than a human one. In Study 1, participants accepted a mortality-reflection task lasting 9.65 minutes to avoid a human conversation, but only 5.06 minutes to avoid an AI conversation (d = 0.34, p < .001).
2. Correction of Misperceptions and Warming of Affect: In Study 2, a single ten-minute conversation corrected misperceptions of the outgroup's environmental attitudes and increased outgroup warmth by 4.3 points on a 0–100 scale (d = 0.37). The magnitude of belief correction predicted the magnitude of warmth gain. Democrats, who held larger initial misperceptions, showed greater gains than Republicans.
3. Specificity of Effect: Study 3 confirmed that the warmth increase was driven by outgroup-relevant content, not general sociality. The outgroup bot condition significantly outperformed both the "cats vs. dogs" chat control and the Space Invaders game control (d ≈ 0.58–0.59).
4. Behavioral Consequences: In Study 4, participants who spoke with an outgroup bot were 6 percentage points more likely to choose a real cross-partisan conversation over an aversive task compared to controls (Odds Ratio = 1.33, p = 0.025).
5. Durability: In Study 5, the immediate warmth effect (d = 0.46) largely faded within one week. However, a small residual effect persisted (d = 0.12, p = .02), which was concentrated among the most extreme partisans.
6. Mechanism: Content analysis revealed that outgroup bots differed from control bots primarily in delivering "stereotype-disconfirming substance" (cognitive route) rather than in empathy or friendliness (affective route). The cognitive gap was consistently larger and more reliable than the affective gap.

Significance and Claims
The paper establishes synthetic contact as a scalable, behaviorally consequential, and widely acceptable form of cross-partisan engagement. Unlike face-to-face contact, which partisans actively avoid due to interpersonal threat and anxiety, synthetic contact lowers the barrier to entry. The authors argue that while traditional human contact often works through affective routes (reduced anxiety, increased empathy), synthetic contact appears to work primarily through a cognitive route: by providing accurate, stereotype-disconfirming information in a low-stakes environment where the "interpersonal threat" of being judged is removed.

The authors claim that even if AI bots are imperfectly calibrated (sometimes holding more extreme views than the humans they represent), they can still correct misperceptions if they are less wrong than the participant's starting beliefs. The findings suggest that brief AI interactions can serve as a "gateway" intervention, making partisans more willing to engage in real-world cross-partisan dialogue. However, the authors note limitations, including the modest duration of effects (fading within a week) and the reliance on online U.S. samples, acknowledging that whether these effects scale to societal-level change remains an open question.