From prediction to resilience: transferable and equitable AI for climate extremes in a data-unequal Anthropocene

Artificial intelligence contributes to climate-extreme prediction through at least three overlapping channels: integrating heterogeneous data, learning alternative predictive mappings, and reducing computational cost. Recent systems show substantial gains, but predictive skill alone does not establish robustness under unseen extremes, transfer across regions or climates, or equitable resilience. This invited short review uses a structured narrative synthesis of recent literature and proposes an eight-gate validation framework. The gates evaluate predictive skill, data representativeness, out-of-distribution robustness, spatial and temporal transferability, calibrated uncertainty, impact translation, local validation and actionability, and equitable resilience. The gates are sequential in claim strength rather than an automatic causal pipeline: failure at one gate limits stronger downstream claims and should trigger iteration rather than invalidate earlier scientific value. We also distinguish the technical contribution of artificial intelligence (AI) from the institutional, communication, and response capacities required for resilience. The framework provides practical questions and example diagnostics for moving from impressive forecasts toward trustworthy and equitable climate-risk value.