AI-Generated Climate Disaster Images

This project examined how AI-generated climate disaster images influence audience responses and support for climate action. 

Study 1 employed a one-factor between-subjects design with three conditions: (1) a text-only control condition featuring a simple call for climate action; (2) a low visual realism condition combining the same call with a low-realism AI-generated climate disaster image; and (3) a high visual realism condition combining the same call with a high-realism AI-generated climate disaster image. Participants were randomly assigned to one of the three conditions. Within each condition, participants were further randomly assigned to one of three message versions, resulting in a total of nine stimuli. These message versions represented different versions of the same experimental conditions. Following consent, participants completed a survey assessing sociodemographic characteristics (e.g., age, gender, ethnicity, education, political orientation) and prior disaster experience. After viewing one of the nine stimuli, participants reported their reactions (i.e., perceived threat to freedom, emotions and reactance, and willingness to make personal sacrifices for climate action). All response variables were measured using 7-point Likert scales. Missing data were minimal (<1.5% per item) and showed no systematic patterns. Missing values were replaced using item-level mean imputation within each of the nine stimulus conditions prior to scale aggregation. Data were collected via the online research platform Prolific, and all participants completed the study online without researcher presence.

Study 2 utilized the same design and procedure as Study 1.

Study 3 employed a one-factor between-subjects design with three conditions: (1) a text-only message containing a call to action; (2) the same message accompanied by a highly realistic AI-generated climate disaster image; and (3) the same message and image accompanied by an additional AI-origin label stating that the image was “created by artificial intelligence.” Participants were randomly assigned to one of the three conditions. Within each condition, participants were further randomly assigned to one of three message versions (flood, wildfire, or hurricane), resulting in a total of nine stimuli. These message versions represented different versions of the same experimental conditions. The procedure, including informed consent, was identical to that in Study 1. In addition to the measures used in Study 1, we assessed perceived trustworthiness of the message sender. Participants in the image conditions (Conditions 2 and 3) were also asked whether they believed the image was AI-generated, a photograph, or if they were unsure. Missing data were minimal (<1.5% per item) and showed no systematic patterns. Missing values were replaced using item-level mean imputation within each of the nine stimulus conditions prior to scale aggregation. Data were collected via the online research platform Prolific, and all participants completed the study online without researcher presence.