Answers are grounded in this paper and cite their sources.
Randomized trial evaluates large language models for improving disaster risk governance in the Arctic, suggesting enhanced policy design.
The Arctic region is experiencing rapid socio-environmental disruptions driven by compounding climate change, which increases the frequency of transboundary disasters. While Adaptive Goevrnacne (AG) offers critical mechanisms for institutional flexibility and multi-stakeholder learning, its execution in Disaster Emergency Management (DEM) remains constrained by entrenched top-down hierarchies and institutional silos. This study evaluates the capacity of four foundational Large Language Models (LLMs) – GPT-4, Gemini, Yandex GPT, and Deep Seek to systematically identify and ameliorate policy gaps within official post-disaster reports from Russia, the United States, and Canada. Using an analytical framework based on eight established AG indicators, we implement a multi-stage iterative human-verification protocol to define operational success thresholds and error rates across diverse model architectures. The empirical results demonstrate that while raw outputs exhibit notable analytical interpolation errors, a structured, verified human-in-the-loop pipeline yields high-precision text synthesis, with GPT-4 achieving a success rate of up to 93.75%. We introduce the concept of LLM-human policy co-design, positioning constrained generative AI not as an objective source of truth, but as an “epistemic mirror” capable of exposing institutional blind spots and accelerating the double-loop learning cycle essential for polycentric disaster risk reduction.
No takes yet. Share an insight, caveat, or question.
Olga Derendiaeva (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: