The evidence reviewed in this section suggests that deliberate policy choices—such as fact-​­checking, algorithmic transparency, and opportunities for citizens to deliberate—can counter the default trajectory. The investments required are not prohibitive. The main challenge may be the lack of institutional capacity and political will. Policy choices will shape power imbalances among governments, corporations, and individuals Power has always been about access and control, whether to land, capital, markets, or information. What distinguishes AI is the speed and scale at which it can redistribute access, often in ways invisible to those most affected. The governance challenge may be more fundamental than simply writing better rules or investing more resources. It may require rethinking what sovereignty means when essential infrastructure is privately owned and globally distributed, what accountability means when decisions are made by systems too complex for even their designers to fully explain, and what consent means when the costs of refusing to participate are too high for most people to bear. A broader consideration is whether the current patterns shaping AI align with the needs of most of the world’s population. Most people live in developing countries, work in informal sectors, speak languages other than English, and interact with institutions limited by capacity constraints. Yet AI systems are predominantly designed by and for populations in high-income countries—a relatively small fraction of humanity. History shows that outcomes depend on choices about regulation, public investment, labor protections, and international cooperation. For developing countries, the path forward involves building governance capacity, making strategic investments in areas of comparative advantage, and coordinating with other countries to shape international norms. The technology does not determine outcomes; governance choices do. Notes 1. The AI stack, as discussed in chapter 2, consists of five interdependent layers: applications, AI models, data, infrastructure, and hardware. 2. Allen and Chan (2017). 3. Sutter (2025). 4. Baskaran (2024); Szczepański (2025). 5. Szczepański (2025). 6. Open-source models provide transparency and customization, allowing those that build AI applications to access the model weights, architecture, training data, and source code, with minimal licensing restrictions. Examples include AI Singapore’s Southeast Asian Languages in One Network (SEA-LION), Allen Institute for AI’s Olmo models, BigScience Large Open-science OpenAccess Multilingual (BLOOM) Language Model, EleutherAI’s GPT-NeoX, Google’s Bidirectional Encoder Representations from Transformers (BERT) models, and the Swiss AI Initiative’s Apertus. Open-weight models allow developers to access the model weights, but restrict access to other areas. Examples include Chile’s National Center for Artificial Intelligence’s Latam-GPT, 220 DeepSeek’s models, Google’s Gemma models, Lelapa AI’s InkubaLM, and Meta’s Llama models (although they are sometimes marketed as open source). A couple of OpenAI’s GPT models also fall into this category. 7. United States, White House (2025). 8. As of 2019, China had reportedly spent $79 billion on projects related to the Digital Silk Road and, by 2020, had signed cooperation agreements or provided related investments to at least 16 countries (Kurlantzick 2020). The actual numbers are likely to be higher because many of these agreements are unreported. 9. Liu and Yang (2025). 10. Clayton et al. (2025). 11. Schneider and Urpelainen (2013). 12. Refer to Pax Silica (dashboard), Under Secretary for Economic Affairs, US Department of State, https:// www.state.gov/pax-silica. 13. Kaushik et al. (2025); Marivate (2026). 14. Computing infrastructure refers to the hardware needed to train and run AI models, including specialized chips, networking and storage World Development Report 2026