2026-05-11 中国科学院（CAS）
＜関連情報＞
- https://english.cas.cn/newsroom/research-news/202605/t20260511_1159121.shtml
- https://conbio.onlinelibrary.wiley.com/doi/10.1111/cobi.70287
生態学および保全科学における大規模言語モデルの新たな応用 Emerging applications of large language models in ecology and conservation science
Christos Mammides, Hao Gu, Thilina S. Nimalrathna, Naufal Rahman Avicena, Harris Papadopoulos, Ahimsa Campos-Arceiz
Conservation Biology Published: 13 April 2026
DOI:https://doi.org/10.1111/cobi.70287
Abstract
Large language models (LLMs) mark a major development in artificial intelligence, with potentially transformative implications for ecology and conservation science. Built on advanced deep-learning architectures, these models can support a wide range of tasks. We reviewed emerging applications of LLMs, drawing on the wider scientific literature and practical use cases. We found that LLMs can streamline ecological workflows and accelerate evidence-based conservation by supporting the extraction of ecological information from unstructured sources, enabling natural-language interaction with structured databases and facilitating large-scale literature syntheses. They can also be used to leverage publicly available data for ecological insights, for example, through automated monitoring of news reports. They can enhance biodiversity monitoring through integration with edge devices, such as camera traps, and can assist with analytical tasks, such as code generation, and improve scientific communication and support outreach, for example, through custom models trained on domain-specific information. Other potential applications include policy analysis and decision support, such as simulating interactions among stakeholders with multiagent systems. However, the rapid adoption of LLMs also raises technical and ethical challenges, including inaccurate or biased outputs caused by hallucinations and imbalances in training data. Such limitations can also contribute to poor out-of-distribution performance and the underrepresentation of minority viewpoints. Additional concerns include limited transparency and reproducibility due to their black-box nature, high technical complexity, and computational demands, which may exacerbate access inequalities, the risk of deskilling, and environmental impacts. To mitigate these challenges, we recommend a set of best practices, including careful model selection, effective prompt engineering, retrieval-augmented generation to improve factual accuracy and representation, human-in-the-loop validation, and broader efforts to promote inclusive development, capacity building, and appropriate governance. When applied thoughtfully, LLMs can serve as a valuable addition to the ecologists’ toolkit, enhancing scientific capacity and supporting efforts toward achieving global biodiversity goals.
