Mapping with Words: Integrating Large Language Models into Geospatial Practice

Abstract. For decades, the focus of geospatial artificial intelligence (AI) has been on imagery and data, with linguistic interfaces remaining unexplored. This study presents a systematic review of research applying Large Language Models (LLMs) to Cartography and GIScience. We analysed 54 peer-reviewed articles published between 2023 and 2025, mapping the use of LLMs in data acquisition, semantic enrichment, spatial analysis, cartographic design, and user interaction, among other topics. The corpus reveals four dominant application clusters: (1) semantic creation/alignment of Geo-knowledge graphs; (2) text- and vision-based data acquisition; (3) language-driven spatial analysis and Geometric Question Answering (GeoQA); and (4) automated map styling and symbolisation. GPT-3.5/4 underpins two-thirds of the studies, while open-weight models, such as LLaMA-2, FLAN-T5, Gemini, and DeepSeek, are gaining traction. LLM work aligns most strongly with the challenges of openness and reproducibility, as well as cartographic design automation, but is noticeably weaker in areas such as provenance ethics, causal inference, participatory mapping, and mobile multimodal interaction. We outline three priorities for future research: (i) open benchmark datasets for spatial reasoning and map quality; (ii) ethics checklists that surface bias, privacy and hallucination risks; and (iii) investment in multilingual, low-resource Geo- LLMs to broaden global participation. By mapping current advances against long-standing research gaps, the review provides an actionable agenda for guiding large language models (LLMs) toward equitable and trustworthy cartographic practice.