From Large Language Models to Agentic Systems: A Systematic Review of Generative AI in GIScience

ABSTRACT

Generative artificial intelligence (GenAI) is rapidly reshaping GIScience, yet the field has grown faster than its foundations have been systematically assessed. This review synthesizes how large language models (LLMs), vision‐language models (VLMs), diffusion models, and other generative approaches are deployed across geographic research. Screening an initial pool of 1146 records from the Web of Science, we identified 295 peer‐reviewed journal and conference papers published between November 2022 and February 2026 that demonstrate substantive engagement with geospatial analysis. Utilizing a 31‐variable extraction framework, this review combines corpus‐level descriptive statistics, thematic synthesis, and cross‐cutting evaluations of methodological trends, openness, trustworthiness, and geographic grounding. The literature reveals explosive growth, with LLMs remaining the dominant model family while multimodal pipelines and hybrid systems become increasingly prominent. However, our cross‐cutting synthesis exposes persistent structural gaps: formal uncertainty quantification remains rare; hallucination and geographic bias are unevenly addressed; full code‐and‐data openness is heavily limited; and explicit engagement with geographic theory remains incomplete, particularly within technically intensive areas like remote sensing and spatiotemporal modeling. Looking forward, we argue that disciplinary maturity depends on establishing five strategic pillars: geographically grounded uncertainty standards, spatial equity and justice frameworks, geospatially specific reproducibility norms, theory‐informed model design, and agentic GIS architectures that systematically transfer mature computer‐science mechanisms—including tool use, multi‐agent planning, reflection, persistent memory, and trajectory‐based evaluation—into spatially constrained, auditable geographic systems.