Yale Program on Climate Change Communication

Home / Can large language models estimate public opinion about global warming? / Fig2

· August 6, 2024

Fig2

This figure compares the levels of worry about global warming between survey data and GPT model responses for 2017 and 2021. The data is shown in bar charts with different colors: Dark Blue for Survey, Red for GPT-4 with demographics and covariates, Blue for GPT-3.5 with demographics and covariates, Yellow for GPT-4 with demographics only, and Green for GPT-3.5 with demographics only. Each chart is divided into four worry levels: "Very worried," "Somewhat worried," "Not very worried," and "Not at all worried." The y-axis represents response percentages, ranging from 0% to 80%. In 2017, "Very worried" is highest for GPT-3.5 with demographics only (Green), while "Somewhat worried" is highest for GPT-3.5 with demographics only (Green). "Not very worried" and "Not at all worried" have low percentages across all groups. In 2021, "Very worried" is highest for GPT-4 with demographics only (Yellow), while "Somewhat worried" is highest for GPT-3.5 with demographics only (Green). "Not very worried" and "Not at all worried" have low percentages across all groups. Note: "Demo Only" represents GPTs conditioned solely on demographics, and "Demo + Cov" represents GPTs conditioned on demographics and covariates. Credits: Yale Program on Climate Change Communication and George Mason University's Center for Climate Change Communication.