Closing the Diffusion Deficit: Trustworthiness Verdicts for the AI That Actually Reaches the Global South
Abstract
Technical AI governance is dominated by a frontier-lab framing, yet the systems that reach most of humanity are distilled, fine-tuned, and platform-embedded rather than frontier. We argue that the decisive object of trustworthy-AI-for-good governance is diffusion capacity—whether an AI system is affordable, adaptable, and institutionally governable where it is actually deployed—not innovation capacity. We decompose diffusion capacity into three components (accessibility, adaptability, institutional compatibility) and ground in them an Integrated Developmental Model (IDM) whose three pillars—inclusive algorithms, sustainable computing, equitable data—each inherit an established literature, together forming three sequential binding constraints on diffusion. Safety, accountability, and misuse enter not as a fourth pillar but as a governance threshold that moderates whether diffusion produces development, plus veto caps on pillar grades. We develop the model from primary fieldwork across ten Chinese enterprises and four Southeast-Asian corridors (Singapore, Bangkok, Hanoi, Jakarta), and specify a falsification design against cost-arbitrage, geopolitical, state-capital, and institutional-attraction rivals whose one out-of-sample test—the US-compute-path African build-out—is live now. We propose two externally auditable, within-island metrics—Joules per Task (JpT, bound to sustainable computing) and a PPP-adjusted Accessibility Index (PAI, bound to inclusive algorithms)—and collate divergent benchmark values across recent inference-energy and affordability evidence. Because a trustworthiness framework is credible only if it can return adverse verdicts, we exercise the threshold-and-veto machinery against our own focal cases. The contribution is a diffusion-centric evidence standard for trustworthy AI in the Global South: an auditable framework that produces trustworthiness verdicts—including adverse ones—on the deployments that actually reach most of humanity.