Detection and spill-over effects of AI-generated images in political messages: Evidence from two pre-registered experiments

AI-generated images may not only be used by malicious actors in political communication, but also by (photo)journalists or NGOs who cannot or—for ethical reasons—do not wish to rely on photographs to document political events. Yet, using synthetic visuals can potentially erode message and source trust, and may mislead audiences, especially when disclosures are missing. Addressing these issues, this article reports findings from two pre-registered between-subjects experiments ( N Total = 890) among German-speaking individuals on detection and potential spill-over effects of AI-generated images, in which respondents were exposed to posts by fictional NGOs featuring real photographs or (un)labeled AI-generated images. The mostly young and highly educated sample in Study 1 showed strong detection skills, but Study 2—using a quota-based sample—revealed that the average person struggles to identify AI-generated images without disclaimers. Though labeling can significantly improve detection for the average person, it is not always bound to be successful. Moreover, labeling can reduce message and source credibility among people who are distant from the political center, highlighting potential drawbacks associated with using and labeling AI-generated images in political communication. Meanwhile, the level of images’ probative value (i.e., their apparent evidentiary power) did not affect reactions to AI-generated visuals. Perceptions of these images may thus not depend on the degree of documentation that they purport to provide. Considering the practical relevance of these findings, this article highlights not only conceptual contributions of this work, but also implications for political actors, policymakers and media education. • Study explores identification and effects of AI-based images in political messages. • Average person struggles to identify AI-generated images without labels. • Labeling improves synthetic image detection, but may reduce source/message trust. • Spill-over effects may not vary by images’ probative value.