An AI ‐assisted workflow for generating historical species distributions from avian museum specimens

The Earth's biodiversity is changing rapidly due to human activities. Natural history collections could help us better understand how and why biodiversity is changing but are grossly under‐used in this respect. Artificial intelligence (AI) has the potential to transform this situation by enabling us to unlock data associated with museum specimens more rapidly and accurately than ever before, but this potential needs to be rigorously evaluated. Here, we develop, test and apply a workflow for extracting data from the labels associated with skin preparations of birds using large language models (LLMs) and illustrate how to generate historical species distributions from the resultant data. To do this, we used specimens from the Archbold‐Vernay Expedition that visited Madagascar from 1929 to 1931 and collected >7000 specimens as a case study. We collected images of 2576 specimen labels and then used ChatGPT‐4o to extract the data from these images. Testing against a manually transcribed set of 130 labels showed excellent levels of transcription accuracy (fuzzy‐matching scores >95). Further analyses showed that transcription accuracy varied between labels and data categories, and was significantly poorer if the location information was handwritten compared with typed. We geocoded each specimen using the location information from its label and GEOLocate, a web‐based geocoding application, and some manual processing. This process successfully geocoded 91.4% of the 4965 specimens for which we had label‐derived location information. We then used freely available software (MaxEnt) to generate historical species distributions from the geocoded specimen data for a small sample ( n = 4) of endemic species. Our workflow illustrates the potential of AI to generate valuable ecological data from natural history collections, and together with other recent work, suggests AI‐assisted workflows are likely to accelerate the availability of data from museum specimens over the next few years.