AI in climate information and communication: What’s the benefit? What’s the danger? What do people trust?

About the project

About the Project Background The rapid rise of LLMs and implementation of chatbots mean that applications such as ChatGPT, Claude and Gemini are accessible to businesses, organisations and the broader public. In the context of weather services, the use of AI chatbots have been discussed as a potential way to provide personalised and tailored information to users. In the context of climate information LLMs have been explored as a way to summarise climate information for a non-technical audience. However, this raises questions about the extent to which the outputs of prompts and queries capture uncertainties and how this influences how these outputs are interpreted and used to inform both general understanding of climate change and adaptation planning. This project will use insights from data and behavioural sciences to examine how LLM output aligns with current expert understanding of climate and the uncertainties associated with climate information, as well as how it might be used to inform decision making. PhD project People are increasingly using AI in the workplace, including in senior decision-making roles. However, people may be using AI as an input in contexts where they are not a domain expert, and users may have limited understanding of how products like Large Language Models (LLMs) actually work, or their limitations. While the potential to use specialist chatbots to provide user-relevant climate information based on credible sources such as IPCC AR6 has been highlighted (Vaghefi et al., 2023), LLM output may be inconsistent, inaccurate, contain irrelevant information and vary depending on user history. This project aims to identify the practical usefulness and limitations of LLMs in the context of climate information and communication about climate, particularly in relation to the communication or representation of uncertainty. A second aim is to identify the training or programmes that can help overcome the limitations to the greatest possible extent. The project will draw on data science and behavioural science themes, investigating the human interaction and exploring options to improve outcomes and avoid harms from misinformation. As a starting point the project could explore these issues: • what does AI generate when asked about climate forecasts and/or projections for specific regions in general terms? • what sources does it prioritise? • how do commonly used AI application responses characterise uncertainty? • Is the information provided accurate? • How does the information generated change with varying prompt engineering and specificity of user context? An important issue will be establishing the extent of inaccurate information generated and the probability of errors that could adversely affect the decision being taken. Another issue is how users respond to different LLM outputs (e.g. do they check references provided, attribution of credibility?). The findings will be useful for decision makers using LLMs in their work or whose teams use LLMs. Applicant Profile Students with an interest in the application of AI for communication in real-world contexts. Some experience of data science or a related field would be an advantage.

Research areas