Human-in-the-loop large language models can accelerate biodiversity data mobilisation and synthesis

Biodiversity faces significant threats from unprecedented environmental changes, and effective conservation largely depends on biodiversity data to make predictions, prioritise actions, and guide decisions. However, much of this data remain unorganised. AI-based large language models are increasingly being explored to help mobilise data; however, their acceptance within the scientific community is still under discussion. In this study, I examined how well OpenAI’s Codex, a publicly available large language model (LLM), can support biodiversity data mobilisation by identifying its performance and limitations in terms of retrieval and assembly. Using trials with Philippine terrestrial vertebrate data, I found that large language models can accurately retrieve biodiversity data from unstructured text, achieving over 90% accuracy and saving more than 90% of the time compared to human reviewers. For coordinate extraction, the accuracy was lower for exact matches but improved within a 1–5 km range. In trait retrieval, the LLMs reached up to 100% concordance with the human-retrieved dataset. However, the observed variations across taxonomic levels, with some taxa showing very low accuracy and concordance, suggest the need for caution and human oversight. By enabling species record retrieval at the human-level accuracy with reduced processing time, LLM-assisted workflows can scale up biodiversity data collection from the scientific literature. However, AI should not replace the physical presence, local community involvement, early career training, or ethical judgment of conservation biologists. Therefore, a human-in-the-loop approach is essential for advancing LLM application of LLMs in biodiversity research.