Artificial intelligence holds vast potential to help solve a number of challenging problems in Earth system science, from improving prediction of severe weather events to increasing the efficiency of climate models. But as in all AI applications, the use of machine learning and other techniques in environmental science has the potential to introduce biases that could deepen inequities.

The authors of a new paper published in the journal Environmental Data Science argue that researchers must develop ethical, responsible, and trustworthy approaches to applying AI in Earth system science to ensure that unintentional consequences do not worsen environmental and climate injustice. 

“It’s really exciting to see all the ways researchers are finding to creatively apply artificial intelligence in weather, climate, and other environmental science research,” said David John Gagne, a scientist at the National Center for Atmospheric Research (NCAR) and a paper co-author. “But we have a responsibility to ensure that we don’t cause more harm than good.”

The paper’s lead author is Amy McGovern of the University of Oklahoma. Other co-authors include Imme Ebert-Uphoff of Colorado State University and Ann Bostrom of the University of Washington.

A central bias that could be exacerbated by AI is related to where and how weather and climate data are collected. For example, hailstorms, tornadoes, and other severe weather events are more likely to be reported in areas with higher populations. Therefore, the severe weather datasets used to train machine learning models may not adequately represent the amount of severe weather that takes place in rural, sparsely populated parts of the country. The machine learning model, then, will also tend to underpredict the weather in those regions. 

These relatively low-population areas may be home to communities that are already underserved by the weather community.

The authors list a range of other issues that can arise through the use of AI for environmental science, including the use of non-trustworthy models or applying a model to inappropriate situations. 

Read the University of Oklahoma news release
 

See all News