Synthetic Contact with AI Reduces Cross-Partisan Animosity

Political polarization in the United States has reached levels where members of opposing parties view one another with profound suspicion. While social psychology has long championed "intergroup contact"—the idea that meaningful interaction between different groups reduces prejudice—this solution has hit a wall of human avoidance. People actively shun political opponents. Existing digital interventions often backfire by amplifying conflict rather than resolving it.

A new paper proposes a "synthetic" workaround: using AI chatbots to act as proxies for the political outgroup (the opposing political party). By interacting with a machine that represents the "other side," partisans can engage with opposing views. They can do this without the interpersonal anxiety or social threat that makes face-to-face debate so aversive. The researchers find that these brief, AI-mediated conversations can correct systematic misperceptions and increase warmth toward political opponents.

The friction of human contact

The fundamental problem with traditional depolarization efforts is that they rely on direct contact. Decades of research suggest that contact works by building knowledge, lowering anxiety, and increasing empathy. However, in a polarized landscape, the "anxiety" component becomes a massive barrier.

Current approaches struggle because they cannot easily scale. Real-world interventions require logistical planning and professional facilitation. Digital platforms often serve as echo chambers that reinforce existing biases. Furthermore, people suffer from "affective forecasting" errors (the tendency to incorrectly predict how they will feel in the future). They overestimate how unpleasant a conversation with an opponent will be. This leads them to avoid these encounters entirely. This avoidance creates a feedback loop. Because partisans do not talk to one another, they develop inaccurate, caricatured beliefs about the opposition. This in turn fuels further animosity.

Scaling empathy through synthetic proxies

The researchers designed five studies to test if "synthetic contact" could bypass these barriers. Synthetic contact involves conversations with AI chatbots prompted to represent a typical member of the political outgroup. Their approach rests on a specific logic. If an AI can simulate a representative viewpoint, it can serve as a safe training ground for cognitive and affective realignment.

The mechanism follows three distinct stages: 1. Lowering the barrier to entry: The researchers first quantified the "price" of contact. In an incentive-compatible experiment (where choices have real consequences), they forced participants to choose between a three-minute conversation with a human outgroup member or varying durations of a highly aversive task: contemplating their own mortality. The authors report that participants were far more willing to endure mortality reflection to avoid a human than to avoid an AI .

Specifically, the indifference threshold for a human partner was 9.65 minutes. For an AI, it was only 5.06 minutes. This means people accept nearly twice as much psychological discomfort to avoid a human than an AI. 2. Correcting misperceptions: Once the barrier was lowered, the goal was to fix the "knowledge gap." The researchers found that Democrats held views of Republicans that were more than a standard deviation away from reality on environmental attitudes .

This error was large enough to fundamentally mischaracterize the opposition. 3. Driving behavioral shifts: Finally, the researchers tested if this synthetic contact could translate into real-world bravery. They measured whether speaking to an outgroup bot made participants more likely to choose a real-world conversation with a human partisan over an aversive task.

Evidence of cognitive realignment

The empirical results suggest that synthetic contact is a functional tool for belief updating. In a within-person study of 500 partisans, a single ten-minute conversation with an outgroup chatbot corrected baseline misperceptions. It also increased outgroup warmth by 4.3 points on a 0–100 scale .

Crucially, the authors found that the effectiveness of the bot was tied to the content of the conversation. Through a cross-study content audit of 4,012 conversations, the researchers used an LLM to score the dialogue on four dimensions. These were stereotype-disconfirming substance, informational specificity, empathy, and friendliness. They report that the outgroup bots differed from control bots primarily through "stereotype-disconfirming substance" .

This refers to information that directly contradicted the user's preconceived notions.

While human contact typically works through an "affective route" (reducing anxiety and increasing empathy), synthetic contact appears to operate through a "cognitive route." The paper finds that the primary driver of warmth in AI interactions is the delivery of information. It is not necessarily about the bot being particularly friendly or empathetic. This distinction is vital. It suggests that to reduce polarization via AI, engineers should prioritize the accuracy of the bot's information.

Limitations of the synthetic model

Despite the promising results, the researchers note that synthetic contact is not a panacea. The effects are notably transient. In a longitudinal study, the authors found that warmth increased immediately after the chat. However, most of that gain faded within a week .

A small residual effect remained, but it was concentrated among the most extreme partisans.

There are also questions regarding generalizability. The studies focused on U.S. partisans and two specific issues: environmental consumption and immigration. It remains unclear if these gains would hold for different cultural contexts. Additionally, the researchers acknowledge that "warmth" is only one facet of affective polarization. The study did not measure deeper metrics like social distance. Finally, because the bots sometimes held more extreme views than the humans they represented, there is a risk. Poorly calibrated models could accidentally entrench the very stereotypes they are meant to dismantle.

The verdict: A scalable bridge

The verdict is a cautious "yes," provided the goal is scalable, low-friction engagement. Synthetic contact succeeds where human contact fails. It decouples the process of learning from the threat of social confrontation. By replacing the person on the other side with a predictable algorithm, the researchers have demonstrated a way to deliver stereotype-disconfirming information. This reaches people who would otherwise refuse to listen.

Whether this can solve societal-level polarization remains to be seen. However, the behavioral evidence is compelling. Synthetic contact actually increases the likelihood that a person will choose a real-world conversation with a human opponent .

For developers of news apps and social media, the takeaway is clear. The path to reducing animosity may lie in teaching machines to be more accurately and substantively diverse.

Figures from the paper