Detecting Synthetic, Doubting Authentic: AI Attribution Bias for Political Imagery

Generative AI (GenAI) has enabled the creation of synthetic political content, raising widespread concerns about its impact on public trust during elections. In a preregistered two-by-two experiment, 1,800 participants evaluated images randomly selected from a pool of forty stimuli, sampled from images circulating on social media and balanced by partisan slant (pro-Democrat vs. pro-Republican) and image type (AI-generated vs. authentic photos). We found the AI attribution bias—a systematic tendency to suspect AI manipulation—participants labeled 58.83% of all images as AI-generated, significantly above the true base rate of 50%. This led to an imbalanced pattern of accuracy, with higher accuracy for identifying AI-generated images (81.81%) compared to authentic ones (64.14%). This highlights how synthetic media shapes political perception not only through potential deception but also by eroding confidence in authentic imagery. We also found significant demographic disparities in discernment ability, with older adults performing worse than younger adults, as expected, and women performing worse than men, which was less anticipated. Partisan congruency further exacerbated misjudgments, as participants often misclassified authentic images conflicting with their partisan views. We identified two complementary protective factors: actively open-minded thinking enhanced detection of both synthetic and authentic content, while GenAI knowledge primarily improved detection of synthetic images. These findings demonstrate that GenAI introduces novel challenges to political information integrity, emphasizing the need for comprehensive interventions that address existing digital inequalities, cognitive biases in evidence evaluation, and domain-specific knowledge deficit.