Global AI Bias Audit for Technical Governance

AI knowledge is geographically unequal: The study found that Llama-3 8B could only provide accurate factual responses in 11.4% of queries, with significant disparities between the Global North and South. This means policymakers and citizens in lower-income countries may receive unreliable or fabricated information when seeking guidance on AI governance.

The Global South faces a "digital barrier" in AI expertise: Technical AI governance knowledge is heavily concentrated in higher-income regions, leaving lower-income countries systematically underserved with reliable data-driven insights. This creates a concerning gap where some regions lack the information needed to make informed decisions about AI policy.

AI systems can mislead vulnerable populations: The research reveals that LLMs may provide "hallucinated facts"—plausible-sounding but false information—particularly when queried about topics relevant to lower-income countries. This poses direct risks to policymakers who rely on AI systems for governance decisions without verification mechanisms.

Current AI training reinforces existing inequalities: Rather than serving as a truly global resource, the study concludes that AI alignment and training processes actually amplify existing geoeconomic and geopolitical asymmetries, potentially widening the gap between well-resourced and under-resourced regions.

Inclusive data representation is essential for global AI safety: The paper argues that more diverse geographic and socioeconomic representation in AI training data is necessary to ensure equitable access to reliable AI-generated information and to support inclusive governance worldwide.