Explainable AI reveals the quantitative hierarchical architecture of global bird extinction risk

Identifying what makes species vulnerable to extinction requires accounting for complex biological and environmental interactions. Due to their high predictive accuracy, machine learning methods have been widely used for these assessments; however, relying on black-box models offers limited interpretability. Here, using a comprehensive dataset of anthropogenic, ecological, morphological, demographic, and biogeographical variables from 9,053 species (81% of birds worldwide), we applied Inductive Logic Programming (ILP), an explainable artificial intelligence framework, to generate explicit and quantitative IF-THEN rules with confidence scores for bird extinction risk. Our approach revealed that extinction vulnerability follows a hierarchical structure, shaped by interactions among range size, morphological traits, and human pressures. The framework recovered well-established knowledge, while also revealing previously undescribed extinction patterns. For example, consistent with prior evidence, species with geographic ranges below [~]13,500 km{superscript 2} were identified as higher risk (88% confidence). Nevertheless, this threshold shifted to [~]3,270 km{superscript 2} when human impacts were removed, revealing quantitatively how anthropogenic activities expand the pool of vulnerable species beyond those at risk due to biological and biogeographical traits alone. Beyond established patterns, species with tail length >304 mm were identified as higher risk (82% confidence), a pattern not previously documented. ILP models achieved 91% overall accuracy, slightly lower than Random Forest (93%), but notably better than Neural Networks (83%). These results show that ILP can offer high accuracy results with full interpretability, also providing quantitative transition thresholds that clarify the structural architecture of extinction risk, and translate complex ecological interactions into actionable tools for conservation.