Deep canvassing using AI

“Deep canvassing” – extended, emotionally resonant conversations encouraging perspective-taking – durably reduces exclusionary attitudes but is resource-intensive. We test whether deep canvassing can be successfully performed by AI. N=1,111 US adults were randomized to have an AI-led deep canvassing conversation about immigration (treatment) or a neutral topic (control). Immediately after the conversation, the treatment significantly reduced anti-immigrant prejudice (d=-0.13) and increased pro-immigration policy support (d=0.15) relative to the control. Five weeks later, in a survey fielded amidst the highly polarizing 2024 US election, these effects endured at 30-40% of their initial size. Thus we find that AI can implement deep canvassing, producing durable and meaningful prejudice reduction with the potential to change minds at scale.