Interpretable by design: Language model–derived ecological rules for species distribution modelling

Predictive models in biodiversity research face a persistent trade-off between accuracy and interpretability. Complex machine learning approaches often achieve high predictive power, yet their opacity limits ecological understanding and undermines trust in conservation decision-making. Here we present a transparent framework that uses large language models (LLMs) to generate shallow, human-auditable decision-tree ensembles for predicting species distributions. Using three freshwater crayfish with contrasting ecological contexts, Austropotamobius torrentium (well-studied native), A. bihariensis (data-poor endemic), and Faxonius limosus (invasive generalist), we show that LLM-derived rule ensembles are competitive with both shallow decision trees and random forests (Macro-F1: 0.708 for A. torrentium , 0.689 for A. bihariensis , 0.927 for F. limosus ), and provide the best performance among all tested models for the data-poor endemic A. bihariensis , where random forests exhibited the highest variance across folds (SD = 0.142). On three synthetic datasets with known ground-truth ecological rules, the LLM recovers true thresholds within 1% accuracy, providing methodological validation that the generated rules reflect genuine ecological structure rather than statistical noise. Beyond performance, the framework yields human-auditable explanations for every prediction, including explicit decision paths, supporting rules, and stable ecological rationales across cross-validation folds. These results suggest that, for structured ecological problems, interpretable models built from LLM-generated rules can approach black-box accuracy, particularly under data scarcity, while offering a practical pathway toward trustworthy biodiversity modelling in conservation applications.