Abstract
Human information processing often exhibits negatively biased credulity: people are more likely to attend to, believe, and share information about hazards or threats than about benefits or positive outcomes (Fessler et al., 2014). Advances in generative AI have enabled the production of hyper-realistic deepfake images and videos that frequently evade lay detection (Groh et al., 2021; Waseem, 2023). Collectively, these cognitive and technological dynamics represent significant risks to the integrity of information ecosystems. When deepfakes depict or reference hazards or threats, they can exploit underlying psychological vulnerabilities, foster greater skepticism toward credible news and scientific sources, and increase the likelihood that individuals will accept and disseminate inaccurate or harmful narratives. Deepfakes thereby contribute to the rapid, large-scale diffusion of misinformation, as malicious actors can weaponize them across political, public health, academic, and other information-dependent domains, propagating and amplifying misinformation while undermining public trust (Vaccari & Chadwick, 2020; Sharma et al., 2023; Ahmed, 2021). Psychological mechanisms such as uncertainty and motivated reasoning further bias audiences toward accepting hazard-framed content, suggesting that deepfakes that depict hazards or harms may spread more readily than benign or corrective media (Hughes et al., 2021). The present study directly investigates this proposition by examining whether negatively biased credulity extends to AI-generated content. By leveraging the rapid proliferation of synthetic media, this work aims to empirically assess the extent to which psychological vulnerabilities interact with emerging technologies to shape patterns of belief, engagement, and misinformation attenuation and diffusion. This study directly investigates this proposition by leveraging the rapid proliferation of AI-generated imagery. Utilizing photo-realistic AI image generators, we produced 60 paired stimuli depicting threatening scenarios (e.g., a snarling pit bull) alongside matched non-threatening equivalents (e.g., a tail-wagging golden retriever). This design ensures a balance in the salience and number of AI-generated cues across conditions. Participants will complete a timed online Qualtrics survey evaluating these image pairs, with time constraints to be determined through pilot testing. The primary outcome measure is whether participants are more prone to mistakenly classifying AI-generated hazard images as real photographs compared to non-threatening ones. To enhance validity and minimize prior exposure effects, a set of real photographs (dirty vs. clean toilet) will also be incorporated at the conclusion of the stimulus set. Additionally, participants will complete a series of scales to assess political orientation, disgust and threat sensitivity, AI experience and attitudes, social principles, and their belief in the threatening nature of the world.