Quantify the AI safety infrastructure gap in low-income and lower-middle-income countries, identifying specific failure modes and evaluating evidence-based interventions for building regional safety capacity.

Survey of AI governance capacity across 60 LDCs; analysis of AI deployment patterns in emerging markets; assessment of regulatory, technical, and human capital requirements for AI safety; comparison with existing capacity-building models in cybersecurity and data privacy.

5B users in low-income and lower-middle-income countries are accessing AI systems with zero local governance infrastructure, zero technical safety capacity, and zero accountability mechanisms.

Typical scenario: multinational LLM deployed in country with no AI regulation, no local technical expertise to audit model outputs, no labor standards for data annotation (workers labeling harmful content earn $2-5/day with zero psychological support).

Failure modes: (1) Accuracy bias — models trained on Western data perform poorly on local contexts, enabling discrimination; (2) Regulatory arbitrage — harmful applications of AI prohibited in EU/US but deployed in low-resource regions; (3) Labor exploitation — low-cost labor for data annotation with zero worker protections.

Intervention evidence: Regional AI Safety Hubs modeled on cybersecurity centers of excellence show 60% improvement in incident detection and response times; capacity-building programs train 300-500 local safety engineers per region.