Can AI Help Reduce Prejudice?

A pre-registered experiment

2,877 U.S. adults, recruited in November 2024, are randomly assigned: 1,430 have a one-on-one conversation about transgender rights with an AI interlocutor built on GPT-4o; 1,447 don't. Randomization means any later difference between the groups is the conversation's doing.

Persuasion, personalized

The AI doesn't deliver a script. Each respondent first completes a moral foundations questionnaire, and over a brief dialogue of six exchanges the model makes its case for transgender rights in the values that person already holds — an appeal to fairness for one person; to care, loyalty, or sanctity for another. The matching matters: a perfect top-two moral match corresponds to roughly +4.8 additional thermometer points.

Attitudes move. Then they drift back.

Right after the conversation, support for transgender rights is significantly higher on every measure: the feeling thermometer rises +3.94 to +5.19 points (p < 0.001), behavioral intentions by 0.23–0.28 SD, attitudes by 0.09–0.13 SD, and a latent support index by 0.22–0.29 SD — ranges across model specifications. One week later the gains attenuate substantially, and after weighting and multiple-comparison corrections they can no longer be reliably distinguished from zero.

Immediate change is possible; durable change is harder

AI can deliver value-aligned persuasion at a scale and level of personalization no canvasser could match — and it works, at least in the moment, with effects of 0.09–0.29 SD across measures. The attenuation a week later is the equally important half of the finding: without reinforcement, the gains are transient. Both halves are in the paper.

estimates from the accepted manuscript — ranges across model specifications · *** p < 0.001 · ** p < 0.01 · wave 1 N = 2,681 · wave 2 N = 1,155 · standard errors, weights, and robustness checks in the paper