Who makes the map? A typology of GeoAI cartographic dysfunctions and a framework for cartographic literacies

Geospatial artificial intelligence (GeoAI) has rapidly entered cartographic production, and a growing empirical literature documents the ways in which it fails: biased outputs, factual hallucinations, fabricated satellite imagery, recursive model degradation, and opaque labor and infrastructure. Recent theoretical contributions on posthuman cartography and neopragmatist fallibilism have begun to reframe cartographic agency for the GeoAI era, but neither integrates the empirical evidence on dysfunctions, and neither operationalizes its theoretical claims into concrete competencies. This paper develops both an integrative diagnosis and a usable response. Drawing on a structured review of cartographic, GIScience and AI-ethics literature published between 2020 and 2025, we propose three contributions. First, a typology of five GeoAI cartographic dysfunctions organized along a previously unnamed axis: dysfunctions visible in the artifact (geographic bias, cartographic hallucination, deepfake geography) and dysfunctions hidden in the production system (model collapse, infrastructural opacity). Second, a dysfunction-literacy mapping that articulates each dysfunction with the cartographic competencies required to detect and remediate it across four literacy domains (semiological, geodetic, datalogical, ethical), yielding an operational matrix supported by four curricular instruments. Third, a reframing of cartographic agency as distributed across the human-AI production network, which extends posthuman cartography by making upstream actors—corpus assemblers, annotators, model developers—explicitly visible. We illustrate the framework on the Tacoma/Seattle/Beijing CycleGAN case, showing how each literacy domain contributes a distinct detection capacity. We discuss implications for publication standards, professional accreditation, and the redistribution of cartographic responsibility. We define GeoAI cartographic dysfunction formally and demarcate it from long-standing cartographic limitations, analyze the causal and cumulative relations among the five families rather than treating them as independent, link them to the GeoAI components that generate them, and pair the literacy framework with current technical mitigation strategies and with an assessment of how rapid progress in AI may attenuate several dysfunctions. The framework has not yet been empirically tested; we report a first round of expert consultation as a preliminary and partial element of validation.