Extreme Heat in the Age of Generative AI: Assessing the Accuracy, Accessibility, Equity, and Inclusion of Theory for Public Health Communication.

Extreme heat, driven by climate change, poses significant health risks, particularly for vulnerable populations such as older adults, infants, and individuals with chronic illnesses. Generative AI may help tailor and scale climate adaptation messages, including for extreme heat, to better support those at risk. This research aims to assess the ability of ChatGPT 4o and DALL-E-3 to create accurate, tailored, theory-based public health text- and image-based messages about extreme heat in Canada. The objectives of this research include: 1) to engineer prompts to develop text- and image-based generative AI outputs about extreme heat, 2) to evaluate the outputs, and 3) to explore the ability of generative AI to tailor outputs further using theory and best practices. A variety of prompts were developed and tested to evaluate ChatGPT 4o and DALL-E 3's ability to generate health messages. 31 generated text and 17 image outputs were evaluated for accuracy, literacy, bias, and alignment with health communication models/frameworks to evaluate their effectiveness in addressing the needs of heat-vulnerable populations. While most text outputs were accurate, none met the target of a grade 8 reading level. Messages often included direct heat impacts and protective actions, but fewer framed heat within climate change. Image outputs were generally accurate but lacked diversity, and two images reinforced stigma. These findings highlight both the potential and the limitations of generative AI in health communication, emphasizing the need for careful oversight to ensure accuracy, accessibility, and inclusivity.