News

Abstract As AI-generated media becomes more common, transparency policies frequently require the disclosure of synthetic content. However, little is known about how different warning formats influence perceived accuracy, midpoint or ambivalent responses, and observable behavioral engagement across cultural contexts. We conducted a multi-country experiment (5435 participants; 43,512 video-level observations) spanning five countries (the United States, Japan, Germany, France, and Indonesia) and comparing two disclosure formats: a pre-exposure message-style warning and an embedded label-style disclosure. Message-style warnings significantly reduced perceived accuracy but did not increase perceived differentiation between true and false videos. In contrast, label-style disclosures increased midpoint responses, consistent with greater expressed uncertainty or ambivalence, but did not increase observable behavioral engagement or perceived differentiation between true and false videos. Across models, individual differences, including attitudes toward AI, conspiracy beliefs, trust in social media, and personality traits (Openness to Experience, Agreeableness, and Extraversion), were substantially stronger predictors of perceived accuracy than warning format. Cross-national differences also exceeded intervention effects. These findings suggest that AI transparency cues primarily modulate perceived accuracy and confidence expression rather than improve veracity-contingent evaluative differentiation. Overall, the disclosure effects were modest relative to individual differences, psychological orientations, and country-level variation. Therefore, AI disclosure systems should be evaluated not only by visibility but also by how they shape multiple user responses under different informational contexts.

Starting in April, new members have joined Shibuya Lab!