An analysis of nearly 4,000 public datasets finds that over 90% of AI training datasets came from Europe and North America, and fewer than 4% came from Africa
that's reshaping the infrastructures of our world in ways that reflect the interests of those big corporations.” Niall Firth / @niallfirth : New findings show how the sources of data are concentrating power in the hands of the most powerful tech companies - new today from @melissahei.bsky.social www.technologyreview.com/2024/12/18/ 1... @hypervisible : Data Provenance Initiative “audited nearly 4,000 public data sets spanning over 600 languages, 67 countries, and three decades. The data came from 800 unique sources and nearly 700 organizations.” Melissa Heikkilä / @melissahei : New research reveals a worrying trend: AI's data practices risk concentrating power overwhelmingly in the hands of dominant technology companies. With analysis from — @shaynelongpre.bsky.social @sarahooker.bsky.social @smw.bsky.social @giadapistilli.com www.technologyreview.com/2024/12/18/ 1...
Context & Ripple Effects
The audit adds a data-provenance dimension to earlier warnings of industrial capture in AI, which documented corporate dominance among the largest models. It identifies where the public inputs to that dominance are sourced, not merely who owns the resulting systems.
It also lands as AI developers face pressure from the depletion of conventional web training data. That makes the geographic and institutional composition of remaining usable data a more consequential competitive constraint.
First-order effects
- The findings make the geographic skew in public AI training inputs measurable: Europe and North America supply the overwhelming share, while African-origin datasets are scarcely represented.
- Developers and dataset curators face a clearer provenance and coverage gap when claiming that broadly deployed models reflect global languages, institutions, or contexts.
Second-order effects
- As general web data becomes less available, demand may shift toward regionally specific and better-documented datasets, raising the strategic value of organizations able to collect and govern them.
- The imbalance gives incumbent technology companies an advantage when they can finance data acquisition, curation, and licensing, reinforcing the concentration previously observed at the model layer.
Third-order effects
- If the pattern persists, control of AI will be shaped not only by models and compute but by access to legitimate, high-quality data from underrepresented regions—a durable AI sovereignty issue.
- The concentration of training inputs can compound other infrastructure disparities, including the later-documented concentration of AI data-center capacity, creating linked digital chokepoints rather than independent gaps.
The trend: AI competition is increasingly becoming a contest over geographically concentrated data, compute, and governance infrastructure, with representation and access becoming strategic assets.