Deep Canvassing using AI reduces prejudice toward undocumented immigrants

“Deep canvassing”—extended, emotionally resonant conversations encouraging perspective taking — can durably reduce exclusionary attitudes, but it is resource-intensive. We test whether a large language model can deliver a deep-canvassing conversation about undocumented immigration. In a preregistered experiment (N = 1,108 U.S. adults), participants were randomized to an AI-led conversation about immigration (treatment) or a structurally matched conversation about smartphone preferences (control). Immediately after the conversation, treatment reduced anti-immigrant prejudice (d = −0.12) and increased support for pro-immigration policies (d = 0.14). Five weeks later, in a survey fielded amidst the highly polarizing 2024 US election, prejudice reductions remained detectable (b = −2.56 on a 0–100 scale, p = .005). Transcript analysis using an LLM-assisted coding pipeline revealed that narrative exchange sequences — in which the AI elicited the participant’s immigration experience, the participant acknowledged lacking such experience, and the AI shared a concrete immigrant story—were the strongest predictors of attitude change across both outcomes, whereas purely informational exchanges were not. These findings show that, with appropriate prompting, an LLM can simulate key elements of deep canvassing and produce measurable attitude change.