AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-23
When Language Models Misinterpret Uncertainty in Earth System Mechanism Attribution
This study addresses uncertainty quantification and communication in Earth system mechanism attribution. It builds probabilistic fingerprints from ensemble estimates and varies the information given to LLMs through an evidence ladder. In an idealized system, numerical separation raised GPT-4o's agreement with intervention labels from 0.80 to 0.98 but did not change GPT-4o-mini's, and in a coupled El Niño model, providing full fingerprints with verbal interpretations let both LLMs identify seven of eight ambiguous events, cutting GPT-4o's false ambiguity reports from 10 of 82 to 1 of 82.