PlaceBot: A Python tool for AI-assisted georeferencing of museum specimen locality data, with cross-vendor model evaluation

Accurate georeferencing of museum specimen locality descriptions is essential for biodiversity research, yet hundreds of millions of records remain as historical free-text labels that are difficult to standardise at scale. Large language models (LLMs) offer a promising approach, but their practical utility depends on positional accuracy, prediction stability, coordinate-return reliability and uncertainty calibration. Here we present PlaceBot, a Python workflow for georeferencing and structuring museum specimen locality strings, and evaluate 12 models from four providers on 700 UK bumblebee (Hymenoptera, Bombus ) and 900 globally distributed dragonfly (Odonata) type-specimen locality descriptions. Gemini 2.5 Pro had the lowest median distance to the manually curated reference coordinates in both datasets (1.61 km UK; 5.83 km global), although repeated-call paired comparisons did not distinguish it from every other leading model after multiplicity correction. Coordinate-return rates exceeded 95% for most model–dataset combinations, but o4-mini returned usable UK coordinates for only 383 of 700 attempts. All models showed run-to-run variation. Model-reported confidence sometimes tracked observed error, but category sizes were uneven, and reported uncertainty radii were not uniformly calibrated. Close agreement between Gemini 2.5 Pro and GPT-4.1 identified a higher-reliability subset: when their outputs were within 2 km, both were within 5 km of the reference for 95.3% of pair-complete UK records and 75.0% of pair-complete global records. This post hoc, internally evaluated signal supports prioritisation for review rather than unverified automatic acceptance. Reassessment of conspicuous model–reference disagreements also identified errors in some manually curated coordinates. LLM-derived coordinates should therefore be treated as machine-generated estimates carrying uncertainty and provenance, not as verified determinations. The model-specific results are a dated snapshot; PlaceBot’s durable contribution is a reusable harness for repeating the evaluation on current models and local records.