Not yet human-reviewed

This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.

Open Veins of AI Auditing: How AI Scrutiny Lags Deployment in the Global South

Citation. Galdon Clavell, G. & Magaard, A. (2026). Open Veins of AI Auditing: How AI Scrutiny Lags Deployment in the Global South. arXiv:2607.21317.

Summary. This preprint argues that AI deployment in the Global South has outpaced the independent evaluation infrastructure needed to examine systems in use. Drawing on audit practice, assessment engagements, and a regional landscape analysis, it identifies recurrent validity, performance, coverage, and bias failures and attributes the evaluation gap principally to absent funded demand and enforceable accountability. It proposes impact-focused, locally situated evaluation and funding conditions that require independent scrutiny.

Key claims

• The authors identify fewer than twenty published second- and third-party audits or assessments of deployed AI systems across Africa, Latin America and the Caribbean, South Asia, and Southeast Asia over the prior decade, despite widespread public-sector deployment and investment.
• Across the evaluations discussed, the paper reports four recurring patterns: proxy targets replacing validity, performance claims that fail under prevalence analysis, systems scoring populations absent from training data, and structural bias persisting after protected attributes are removed.
• The paper argues that impact evaluation should examine outcomes for affected people rather than model benchmarks alone, and that funders can create accountability where regulation and local evaluation capacity are absent by requiring independent evaluation and publication.

Evidence & limitations

• The paper synthesizes a decade of the authors’ audit practice, including a published Brazilian health-system audit, two anonymized unreleased Latin American audits, thirteen Responsible AI Assessments across Sub-Saharan Africa and Asia Pacific, and a Latin America and Caribbean landscape analysis.
• It is a working draft posted to arXiv rather than peer-reviewed research, and several underlying engagements are unreleased or anonymized, limiting independent inspection and generalization beyond the documented cases.
• The public arXiv PDF exactly matches the ingested Drive artifact by SHA-256.

Feeds

• responsible-ai-deployment
• Proposed new topic: global-south-ai-evaluation-infrastructure — the paper’s core contribution is the regionally uneven infrastructure, incentives, and capacity for independent evaluation, which is broader and more specific than the existing deployment-governance page.