Working with uncertainties.

 

If the history of past general-purpose technologies is any indication, it could take years or even decades for the full extent of the impacts of AI to materialize. It will take time to acquire a substantial stock of AI technology across a wide range of industries and in firms of different sizes. It will also take time to build complementary assets in AI infrastructure,data and skills. In addition, firms need time to discover new productive uses for AI and integrate them within production activities. The aggregate economic outcome of AI in the long term is thus highly uncertain. In advanced economies, such as Japan and the United States, optimistic projections place long-term annual productivity gains over a 10 to 20-year horizon at between 1 and 2 per cent. 

With less sectoral exposure to AI, most emerging economies are expected to experience lower levels but still substantial annual growth, at between 0.7 and 1.3 per cent. To put these numbers into perspective, in the past two decades, annual productivity growth in advanced economies has averaged at around 1 per cent and in emerging markets and developing economies, at around 4 per cent. However, these expectations may be overstated. For instance, one estimate for the United States puts the annual AI-induced productivity boost over the next 10 years at less than 0.1 per cent. This is because AI systems may find it difficult to cope with certain tasks and, while the use of AI may generate new tasks that increase revenue, it may also generate others that are more malign, such as cyberattacks. Moreover, AI may harm consumers through manipulation or addiction. The impact of AI on welfare may be lower than its effect on productivity. To shed light on the conditions needed for the use of AI to generate large and longterm aggregate benefits, three sources of uncertainty should be considered.

Part of the disagreement over the longterm aggregate effects of AI originates from uncertainties about the rate of development of the technology and how well and quickly it can be integrated into future economic production. Optimistic observers state that AI will have ever-broadening applications and will spawn adjacent innovations, leading to major productivity improvements.

Advances in AI-powered machine vision for example, have increased the potential of self-driving cars and of autonomous drones. However, the current rapid success of AI may be misleading, since it has largely been accomplished through easy tasks that can be readily learned. In the near future, AI may be faced with increasingly difficult tasks of a more complex and context-dependent nature that cannot be automated with similar efficiency. In such cases, there may be no straightforward mapping between actions and defined outcomes of success and not enough data to teach machines about hidden relationships. An example is in the diagnosis and treatment of psychiatric illnesses, which tend to have complex and historical causes that are difficult to capture in data. For such tasks, AI may be no more productive than existing technologies or human workers.

 At the same time, AI is also likely to create new “bad” tasks that can harm overall productivity and well-being. Examples are deepfakes, misinformation and AI-powered surveillance, which raises social, ethical and privacy-related concerns. It is too early to predict with any degree of confidence how AI systems will transform production in the long term, but it seems that AI technology, as in previous waves of technological innovation, may bring a welcome boost to economic growth, although it may be less impressive than some might have hoped. Moreover, maximizing the positive effects on societies depends on proper guidance and policy measures. Later we will focus on national policies, to seize the opportunities brought by AI and consider AI policies and governance from an international perspective.

Productivity gains depend on the long-term structural adjustments in the labour market, as AI can augment or displace labour. If AI is designed and used primarily as a labour substituting technology, in the long term, the declining employment share in sectors that are more AI intensive can diminish the overall economic effect of productivity gains.

 While workers displaced from AI-impacted sectors may be partially absorbed by sectors with lower productivity, this could result in job polarization and widening income inequality. Thus, although productivity can increase in AI-intensive sectors, the aggregate productivity impact could be limited by slower productivity growth in labour-intensive sectors. This outcome resembles a scenario of Baumol’s cost disease, in which aggregate productivity growth is defined less by the sectors at the forefront of technological change than by those that are slower to improve. The actual outcome depends on future interactions between AI adoption and the labour market. If AI acts as a labour complementing rather than labour-displacing technology in a sufficient number of sectors, it can raise aggregate productivity. Another mitigating factor is the extent and nature of job creation. In the past, automation technologies initially caused job losses that were offset in the long term by the appearance of new job. This reinstatement effect can be strong if AI spawns many complementary industries, particularly in areas in which humans retain a comparative advantage over machines. Yet this could take time. Due to skill mismatches and frictions in the labour market, the transition of workers into these new industries could be slow and costly, and fail to keep pace with rapid changes in AI.

The adoption of AI in many developing countries may be hindered by constraints involving the three leverage points of infrastructure, data and skills, creating uncertainty about how these countries can fully exploit the potential of AI. Developing countries have a higher proportion of occupations concentrated in primary and non-knowledge–intensive sectors and, in general, fewer opportunities for AI applications, but large countries can leverage their size and critical mass (see chapter III). More importantly, developing countries may be weaker with regard to critical digital infrastructure and complementary assets such as data and skills. The low level of penetration of reliable electricity and high-speed Internet limits the deployment of AI services, particularly in rural areas. A further impediment is the availability of relevant data. AI models need to be trained on large amounts of high-quality data, but the best data sets are often controlled by global corporations. This can significantly hinder the capacity of developing countries to tailor AI systems to local needs. In addition, with regard to skills, in developing countries in particular, only a small portion of the population has general digital literacy or specialized technical know-how, which hinders the adoption of AI.

 

The need for long-term and significant adjustments does not imply that AI is less relevant in developing countries. With careful and targeted implementation, the use of AI can generate immediate and positive changes. However, developing countries need to create the right conditions in order to seize the gains of AI and ensure that they are not left behind. In addition to boosting productivity for workers and firms, the use of AI offers distinct benefits for sustainable development. It can, for example, help decision makers optimize the distribution of scarce resources. Using advanced analytics, they can draw insights from new sources of unstructured data. GenAI systems can also offer support for individuals who would otherwise not have access to specialized knowledge, for instance in education and agriculture. To help fill the gap of systematic evidence about AI, section E showcases AI applications in developing countries that can deliver improvements in productivity and human welfare across three key sectors. The case studies also show how limitations in infrastructure, data and skills can be addressed through careful implementation and collaboration among stakeholders, to fit local contexts.