Simulating Climate Change Discussion with Large Language Models: Considerations for Science Communication at Scale

Large language models (LLMs) have shown promise in simulating public opinions on social issues. These models can be leveraged in educational simulations that allow students to acquire information and feedback from multiple perspectives. In this research, we investigate the potential of using LLMs (specifically GPT-4) to generate open-ended responses about climate change within a science communication simulation. We prompt GPT-4 to role-play as different personas with various demographics (race/ethnicity, gender, age, income, political affiliations, and ability status) and levels of concern about climate change. We find that GPT-4 is capable of representing multifaceted perspectives around climate change's impact and solutions. However, the model may exaggerate narratives for certain personas based on political affiliations, gender, and concern levels. Such exaggeration may lead to homogeneous narratives that do not fully represent the simulated personas. Our findings highlight the affordances and challenges of applying LLMs to simulating public opinions and enriching educational experiences.