Visualizing the Impacts of Climate Change Using AI

• In Event: Politics and Tools of Climate Change in North America

Abstract

In spite of increasingly dire expert warnings regarding the catastrophic effects of runaway climate change and growing acceptance of the problem, the public in many developed countries remains relatively unconcerned with the issue and generally unmotivated to change their behaviour. Existing cross-national research suggests that climate change is generally perceived as a spatially and temporally distant threat, prompting researchers to explore the ways in which climate change can be construed by members of the public as being more proximate. In this study, we examine the potential of a deep learning algorithm to generate highly personalized images to help citizens visualize the possible effects of climate change where they live, as well as the impact of exposure to such AI-generated imagery on subsequently measured levels of concern, engagement, and policy support. Specifically, we ask: Does exposure to a personalized, AI-generated image of an individual's flooded home increase levels of concern, thus promoting greater engagement with the issue?

To answer this question, we fielded three experiments. In experiment 1, we first set out to examine whether or not AI generative images of major flood events are perceived differently than a real picture of an actual flood. We fielded this first experiment using a Lucid sample of 1,099 residents living in the province of Quebec, Canada. We randomly assigned respondents to one of three groups: (1) exposure to a real image of a flooded street in Montreal, (2) exposure to an image of a flood on the same street and location in Montreal, but flooded with an AI-generated algorithm, and (3) a control (no image). We found that the respondent's reaction to post-treatment outcome measures (e.g., risk perception, emotional response) did not differ across respondents exposed to AI-generated or real images of a flood.

Next, we designed a pre-registered population-based survey experiment administered in Canada (n=2000) and in the United States (n=3000). In this second experiment, respondents were randomly assigned to receive either no stimuli beyond the questions asked to all participants in the survey, or to a second treatment group where they were invited to self-administer a personalized AI-generated image of their home with a simulated 1m flood in real time. We found that, relative to the control group, treated respondents showed higher levels of perceived harm from climate change and floods, while also demonstrating greater behavioural intentions (e.g., contact their MP or discuss climate change with their family) and policy support (e.g., Increase taxes on fossil fuels). We also found that treatment effects among a subset of post-treatment measures persisted among the treated respondents we were able to recontact after two weeks.

Finally, we seek to address some of the issues in experiment 2 with a new experiment we plan to administer in Spring 2022. In experiment 3, we will employ a novel sampling strategy targeting households on the coast of the United States located in areas expected to be affected by sea-level-rise. This strategy will employ address-based-sampling and send invites with a unique link to our survey. Random assignment will occur at the respondent-level, such that respondents will randomly be assigned either 1) the pre-generated image of their flooded home, 2) a pre-generated image of a local landmark, or 3) no image. The targeting of locations expected to be affected by climate-induced sea-level-rise will serve to further enhance the internal validity of results obtained in experiment 2, while the inclusion of local landmarks will provide additional findings related to the types of local imagery susceptible to enhance public engagement with climate change.

Authors

• Erick Lachapelle, Universite de Montreal

• Thomas Bergeron