Overview
Artificial intelligence (AI) implementation within the Global South is occurring at a rate comparable to or exceeding that in the Global North. However, AI governance frameworks have not developed commensurately, leading to a substantial disparity in oversight, particularly pronounced in the Global South. A decade of AI audit practice across Latin America, Sub-Saharan Africa, and Asia Pacific indicates specific patterns in the limited field of AI evaluation in these regions and points to reasons for its infrequent occurrence.
Research Context
The research is based on a decade of AI audit practice spanning Latin America, Sub-Saharan Africa, and Asia Pacific. This includes a fully published second-party audit of a deployed system (Robot Laura in Brazil), two completed but unreleased national audits (a child-welfare risk model and a public-employment matching algorithm), thirteen Responsible AI Assessments, and a regional landscape analysis. The observed deployment of AI, particularly in public-sector algorithms, and multibillion-dollar national AI investments in the Global South are juxtaposed against a limited number of published evaluations. Specifically, fewer than twenty published second- and third-party audits of deployed systems have been counted across the region over the past decade.
Approach
The paper documents trends in AI evaluation within the Global South by drawing on a set of completed and published audit practices. It analyzes specific cases, including one fully published second-party audit, two unreleased national audits, and thirteen Responsible AI Assessments, alongside a broader regional landscape analysis. Through this analysis, it identifies recurring patterns in how AI systems are evaluated or fail to be robustly assessed, and subsequently formulates lessons for evaluation, regulation, and funding.
Findings
Through the analysis of AI audit practice in the Global South, four cross-cutting patterns were identified:
- Proxy Targets: There is a substitution of predictability for validity in evaluation metrics.
- Performance Claims: Stated performance claims often collapse under prevalence analysis, indicating issues with real-world applicability or interpretation.
- Population Mismatch: Populations are being scored by models despite those models never having included these populations in their training datasets.
- Persistent Structural Bias: Structural bias remains present even after protected attributes have been removed from the models.
Five lessons emerged from the evaluation efforts, which are distinct from the regulatory, linguistic, and data conditions typically found in the Global North. The core issue underwriting the gap in AI evaluation is posited not as a capacity limitation but rather as a funding problem. The argument is that capacity follows funded demand, and currently, no actor is required or sufficiently funded to ensure accountability for deployed AI systems. The research suggests that development and philanthropic funders, who finance a significant portion of consequential AI in the region, are well-positioned to address this gap. Their funding conditions could mandate independent evaluation in the absence of existing regulatory requirements.
Why This Matters
The observed gap between the rapid deployment of AI in the Global South and the lack of commensurate governance and independent evaluation poses risks related to the validity and fairness of these systems for screened populations. The identified patterns of proxy targets, flawed performance claims, population mismatch in training data, and persistent structural bias indicate that deployed AI systems may not be functioning as intended or may be perpetuating unintended harm. Addressing the funding deficit for independent evaluation is crucial for ensuring the accountability and reliability of AI applications in regions where regulatory frameworks are still developing.