Flickr Africa: Examining Geo-Diversity in Large-Scale, Human-Centric Visual Data

Significant data imbalance exists between Africa and Europe: The study found that African nations have substantially less geotagged human-centric image data on Flickr compared to population-matched European nations, meaning computer vision models trained on standard internet datasets will perform worse when applied to African contexts.

Within-country wealth disparities shape data representation: Image data distribution within African nations correlates with fine-grained wealth estimates, suggesting that poorer regions are systematically underrepresented in datasets—potentially exacerbating algorithmic bias against lower-income populations.

"Othering" distorts African representation: A substantial portion of images tagged as being from Africa were taken by non-local (foreign) photographers, meaning the visual data reflects outsider perspectives rather than how African people represent themselves, which can perpetuate stereotypes in AI systems trained on this data.

Data gaps are persistent and growing: Temporal analysis across two-year intervals revealed that these representation problems are not self-correcting—emerging trends suggest the disparities continue rather than naturally resolve over time.

Current data collection methods fail low- and middle-income countries: The research demonstrates that relying on standard internet data collection methods systematically excludes populations in developing regions, limiting the global applicability of computer vision models and raising concerns about whose needs AI systems are actually designed to serve.