Year: 2026
Type: preprint
Abstract: Wikipedia is among the world’s most important public knowledge datasets and a foundational source for search engines, maps, knowledge graphs, and AI systems, including LLMs. Geographic gaps in coverage may therefore influence what downstream systems can retrieve, infer, and represent. We analyzed 21,270,978 geotagged records across 355 Wikipedia language editions by combining coordinates extracted from article wikitext and Wikidata. We release the dataset and interactive visualization as open resources for research on geographic representation and knowledge inequality. The Global North contained 275.6 records per 1,000 km², compared with 79.0 in the Global South, a density ratio of 3.49. Although this ratio declined from 5.60 in 2010 to 3.49 in 2026, it remained substantial and relatively stable over the past decade. To examine whether this was reflected in AI performance, we asked GPT-5.4 to estimate the populations of 20,000 villages sampled from OpenStreetMap. Plausible estimates were 3.9 times more likely in the Global North than the Global South (37.9% vs. 9.8%). Villages with Wikipedia evidence were also substantially more likely to receive plausible estimates. These findings suggest that geographic inequality in foundational knowledge sources is systematically associated with geographic inequality in AI capability, further marginalizing places that are already underrepresented.
Source: Research Square
Institution: Microsoft (United States)
Language: English
Subfield: Communication
Field: Social Sciences
Domain: Social Sciences
SDG: Reduced inequalities
Open Access status: green
OpenAlex ID: W7225901097
