Three critical leverage points for AI adoption and development.

 

The adoption and development of AI critically depends on the three leverage points of infrastructure, data and skills. 

Infrastructure refers to digital connectivity and computing power, and the associated networks, architecture and resources necessary to create, train and use AI solutions across a community or country.

Data are necessary for training AImodels, with dedicated data for applying models to different use cases. Data are not only an input but are also generated through AI systems. 

Skills include basic digital and advanced AI-specific skills, as well as the complementary skills needed for a cohesive workforce that can effectively create and use AI. 

The elements of infrastructure, data and skills are needed in both adoption and development (table III.2). 

Although some elements may be relevant to both processes, it helps to identify particular AI requirements for more detailed analyses. Each element contributes to technological progress, but only together can they fully catalyse AI diffusion. Such interactions have led to breakthroughs such as deep learning and GenAI that have redefined the technology landscape. By supporting development in these critical leverage points, decision makers can trigger transformational economic cascades.  Policy and governance for AI can serve to determine the overall direction, setting institutional or cultural guardrails, and creating a socioeconomic and structural context favourable to the development of AI ecosystems. We will further discuss the elaboration on domestic policies involving AI and review the state of global AI governance and how it can support efforts to guarantee that AI will benefit all.  

The adoption of AI relies on basic infrastructure such as electricity and the Internet. While over 90 per cent of the world’s population has access to electricity, about 2.6 billion people are still offline and most of them are in rural areas. AI infrastructure can be divided into two broad categories, namely, digital connectivity, which is largely related to information and communications technology (ICT); and computing power, often referred to as AI compute. They provide foundational support and linkages between actors and systems (figure III.4). 

Both require reliable and affordable energy and water resources. Digital connectivity is often categorized into three segments. First, cross-border terrestrial and submarine cables and satellite linkages which provide access to global networks. Second, middle-mile networks are responsible for the distribution of traffic within countries, including content delivery networks and backbone networks. Third, last-mile or access networks are responsible for providing connectivity to individuals, households and businesses, typically consisting of fixed or mobile cellular networks. The increased use of AI systems and complementary technologies puts pressure on all Digital connectivity segments. Although most countries have ICT networks, these often do not extend much beyond densely populated areas. They may be partially complemented by mobile connectivity for small-scale businesses and private users, but AI adoption is likely to be constrained, particularly for industrial uses. As well as connections, end users also need affordable digital devices to connect to ICT networks and any associated hardware, as well as basic computing power. The last-mile limitations of telecommunications infrastructure in many developing countries indicate that, to close digital divides, one of the priorities should be universal digital connectivity. 

The infrastructure demands are even greater for AI development, particularly for AI compute, that is, the computing power necessary to train and execute AI models. The increasing computationalrequirements for creating and training AI algorithms are being driven by an industry oriented towards multitasking and complex models. Handling large amounts of data and reducing operating times requires efficient data centres, high-speed networks and supercomputers. AI compute requires increasingly complex semiconductors to address AI and big data requirements. Most are produced by a handful of firms worldwide; when supplies are limited due to demand spikes or shocks, developing countries may therefore be last in line. Computing resources and elements also include storage, security, backup systems, data centres and cloud computing. These core elements are often already available in many countries but need to be continuously upgraded or replaced to support the application and development of AI. Much of digital and cloud computing operates across national borders, relying on interoperable infrastructure and protocols. GenAI in particular requires accurate and increasing amounts of data, generally through large bandwidth and international connectivity. Efforts to reduce latency times and data transit costs have spurred the deployment of data centres closer to users. This trend can be accelerated by requirements to locate data in a particular territory or by setting standards for privacy or cybersecurity.

Since 2010, the average size of training data sets for language models has tripled each year. Too complex to be effectively processed by traditional processing approaches and platforms, huge and diverse data sets are better addressed by machine learning and deep-learning algorithms, to produce new and transformative insights. The ability of AI models to analyse and learn from data is determined by their quantity, quality and accessibility (figure III.5).

 However, online data stocks are growing more slowly than the demands from AI, with the risk of shortages that can lead to data bottlenecks. An emerging challenge is how to train and operate AI models more efficiently, to produce trustworthy results from more limited data. AI adoption and customization require access to domain-specific data (e.g. geographical, industrial, cultural) that matches the use-case of AI models and solutions. Increasingly, data requirements overlap with infrastructure needs (e.g. data storage and processing), particularly for SMEs in traditional sectors, for which the costs of setting up and handling information technology systems can be prohibitive. The sectoral rollout of AI thus needs fine-tuning, with consideration given to field-specific needs. Compared with adoption, AI development requires larger and more diverse data, to create, train and test foundation models that are generalizable and can be applied to a variety of use cases. Yet the concentration of control over large data sets by a few platform companies may limit opportunities for value generation based on data, including through AI development. This can hinder efforts to catch up, particularly for firms from developing countries. 

Moreover, AI does not solve the “garbage in, garbage out” problem. If the data sets do not, for example, fully represent different groups or cultures, by gender, by underserved communities or by language, then algorithms are likely to produce biased, incomplete or misleading results. Biases, fabrications or hallucinations (i.e. incorrect or misleading results) can be exacerbated when data produced by AI are used as inputs to train other AI models. Data should be easily available and affordable for developers and users, and standardized and interoperable for quality assurance and efficient processing. At the same time, it is important to respect property rights, as well as privacy and security. The acquisition, processing and use of data should comply with legal and ethical norms and requirements with regard to privacy and data ownership, with security and anonymization procedures used to protect personal information. The importance of global data governance is discussed later.

The adoption and development of AI depends on human efforts and skills. Engineers and computer scientists are needed in designing and producing computer chips and coding algorithms. At the same time, end-users require both digital skills and industry-specific knowledge to adopt and adapt AI. Even if an economy has access, awareness and sufficient funds to adopt AI, this may still not suffice unless there are skilled workers who can use AI or identify opportunities for its use throughout the economy.

Universal digital literacy provides a foundation for the inclusive use of frontier technologies and AI systems (figure III.6). However, adopting AI also requires the applied technical knowledge of AI in practice and transversal supporting skills. Furthermore, the adoption and development of AI requires constant flows of data from different industries and domains, along with experts on particular subjects, who can integrate AI systems with their domains. Workers and the public need to learn how to participate in the AI ecosystem and develop their skill sets, for which reskilling is as important as formal education. For example, to employ GenAI effectively, users need to learn how to structure instructions that can be understood by GenAI, called prompt engineering. One study shows that many AI users enjoy using AI in the workplace and elsewhere but are concerned about potential job losses and that AI will decrease wages.  

Creating and training new AI models requires developers who are highly skilled and have acquired technical knowledge, often through tertiary education in mathematics and computer science. The foundation for this is formal education, followed by regular training. All developers need foundational data science and computing skills, as well as AI-specific training, and research and development opportunities across industry and academia. The development of AI also requires non-technical cognitive skills for creative problem-solving.