How decision-makers trust and use machine learning predictions for managing cold-stunning risks

As extreme weather events increase in frequency and magnitude, science-informed tools are increasingly needed to support effective systems-level management, particularly for high-impact, high-risk, multi-organizational hazards. Cold-stunning events (or periods when water temperatures drop to threatening levels for wildlife) pose substantial risks to coastal systems in Texas. In response, the Texas Marine Coldwater Response Collaborative (TCRC) was established to coordinate cold-stunning response across organizations with diverse priorities and interests. A machine learning (ML) water temperature prediction model has been used to guide the cold-stunning advisory process; however, little is known about how users within the TCRC interpret and act on this information in complex, time-sensitive decision environments. This study examined how different users assess, perceive, trust, and use existing ML-based predictive information to discuss the management of evolving cold-stunning event risks. Structured interviews, which incorporated an interactive activity presenting various forms of predictive information at several stages before and during a simulated cold-stunning event, were conducted with 15 TCRC members, encompassing natural resource managers, coastal economic operators, and weather forecasters. Iterative thematic analysis revealed that users prioritized different predictive signals and information sources according to their operational responsibilities, informational needs, and the evolving decision context. How and when users placed trust in the predictive information also varied across users and decision contexts, shaped by familiarity with the predictive tools, relationships with model developers, consistency with other information sources, and the stakes associated with particular decisions. Overall, these findings highlight the importance of engaging stakeholders throughout the development of predictive tools to account for diverse decision contexts and informational needs. A user-centered, iterative development process can help ensure that predictive tools are not only scientifically robust, but also trusted and applicable within complex, real-world systems-level management.