Labeling Deepfake Videos Reduces Exposure But Not Persuasiveness

Abstract: Increasingly realistic AI generated deepfake videos are raising alarm about their potential use in disinformation campaigns. Viewers may be misled and persuaded on policy-relevant beliefs. A commonly proposed solution is to require disclosure. Across four large scale experiments (N = 7,107), we show that the concern about persuasion is justified, but disclosure is only partially effective. Specifically, participants who watched deepfakes regarding AI regulation were persuaded regardless of warning labels. Viewers rated labeled deepfakes as less deceptive, but otherwise evaluated them and their senders no differently. However, disclosure is effective at the extensive margin: people were less likely to watch a labeled video or share it. Our findings suggest labels are effective when users opt-in to viewing or sharing content—but not in contexts like news feeds that don’t offer the opportunity to selectively view content.

Keywords: Deepfakes, Misinformation, Disclosure, Persuasion