Ethics in climate AI: From theory to practice

Climate science, and climate artificial intelligence (AI) in particular, cannot be disconnected from ethical societal issues, such as resource access, conservation, and public health.An apparently apolitical choice-for example, treating all data points used to train an AI model equally -can result in models that are more accurate in regions where the density and quality of data is higher; these often coincide with the northern and western areas of the world (e.g., 1,2).Inequity in the access to data and computational resources exacerbates gaps between communities in understanding climate change impacts and acting towards mitigation and adaptation, often in ways that are detrimental to those who are most affected (e.g., 3,4).While these issues are not exclusive to AI, widespread opacity in the development and functioning of AI models, presentation of AI model outcomes, and the rapid evolution of the AI field further increase the inequality in power and agency among differently resourced parties.This creates an opportunity for climate scientists to rethink the role of ethics in their approach to research.There are many ways in which climate scientists can interact with society.Here we focus on the process of scientific research, identifying some good practices for building trustworthy and responsible models and then providing some resources.In creating and training models, we encourage researchers to recognize that science cannot claim to be purely "objective", and that the choice of priors, data, and metrics all carry biases (e.g., 5).Resolving or eliminating them is not realistic, as the interpretation of a "better" model or result is highly dependent on the user's specific goal.Hence, it is crucial to be open and specific about the assumptions made, the algorithms and hyperparameters used, and the evaluation metrics and processes, and ideally to also make data and code available, following the principles of reproducible science (e.g., 6).In evaluating and presenting the performance of statistical or machine learning models, considering possible failure modes can be a useful lens through which to examine systems' behaviors.Good starting points are the taxonomy proposed by 7, which considers flaws in design, implementation, and communication, and, specifically for climate science, the list compiled by 8.Failure modes that are particularly relevant for AI models from the climate domain include robustness under distribution shifts; for example, for models trained on historical data, it is difficult to predict how they will perform in the unseen conditions brought about by climate change.Good practices induced by consideration of failure modes may include efforts to quantify expected model performance through simulations, transparency in defining the anticipated range of applicability of the model, and considering how existing physics/climate knowledge can be applied in defining an evaluation strategy and in validating