AI is the simulation of human intelligence by machines, including understanding natural language, making decisions, recognizing patterns, learning from experience, and solving complex problems. This includes machine learning and neutral networks, natural language processing and generative AI, and computer vision and robotics.
Increasingly pervasive products and services are already changing consumer experiences and corporate processes. Meanwhile, US equity market performance over the past year has been driven by a narrow number of “AI winners” (Chart 1).
Chart 1. US equity market gains during 2023 have been driven by a smaller number of “AI winners”
Source: Bloomberg, abrdn, August 2023.
Unlocking higher productivity paradigms?
Developed economies have been stuck in a period of low productivity growth since the global financial crisis (Chart 2). This has been attributed to exhausting the low hanging fruit of past innovations, stagnating educational attainment, fewer spillovers from globalization, a more intangible economy, demand deficiency, and a lack of investment, and mismeasurement.
Chart 2. US productivity growth has disappointed since the mid-2000s
Source: Haver, abrdn, August 2023.
Meanwhile, technological changes of the past few decades, including smartphones, e-commerce, cloud computing, the internet of things, and now AI, have had limited impact on measured productivity growth. This is not surprising and shouldn’t be extrapolated to mean a productivity boost from AI is not coming.
The Solow paradox referred to the absence of a measurable productivity boost from the computer revolution (the “Third Industrial Revolution”) of the 1970s and 80s – which then showed up in the 1990s. Before that, electrification and the internal combustion engine in the late 19th century's Second Industrial Revolution didn't show up in the productivity data until after World War I. It takes time for innovations to diffuse into widespread use.
This delayed but eventually transformative impact appears to be a hallmark of a general-purpose technology, or GPT (not to be confused with a Generative Pre-trained Transformer of the ChatGPT variety). AI also shares many of the other features of a GPT: pervasiveness, continuous improvement, and innovation spawning. There are too many potential labor-saving and augmenting applications, over too many domains, to not have an impact on long-run productivity growth. We are therefore cautious optimists on the eventual positive productivity impact from AI. While our long-term paradigms cite “back to the new normal” as the baseline, our next most likely paradigm is “productivity rebound”.
Does higher productivity mean higher wages?
Productivity growth is a necessary rather than sufficient condition for higher real wages. It is possible to have long periods of productivity growth without an increase in wages.
This famously occurred during the early First Industrial Revolution, when, from 1800 to 1820, productivity increased as a result of steam power, railways, and the telegram, but real wages stagnated (Chart 3). This period is referred to as Engels' pause, where the gains from higher productivity accrued almost exclusively to the owners of capital.
Chart 3. The Engels' pause saw real wages lag productivity growth during the First Industrial Revolution in England
Source: Bank of England, abrdn, August 2023.
It is plausible that even if AI delivers huge productivity gains akin to another industrial revolution (the “Fourth Industrial Revolution”), those gains could be extremely concentrated. The outcome will ultimately turn on the impact of AI on the labor market and the social institutions that mediate the distributional consequences of economic change.
Creating new jobs, enhancing existing ones, or replacing humans entirely?
On the labor market front, it is helpful to distinguish between three effects that technological adoption could have on employment: job destruction, productivity enhancement, and job creation. Job destruction is the effect that typically gets the most attention. In particular, there is a growing sense that generative AI will automate white collar, or those considered creative, jobs that were previously thought beyond the reach of machines.
The widespread adoption of AI will almost certainly lead to some job destruction. But in the long sweep of history, there is very little evidence of sustained technological unemployment. Swings in aggregate demand result in cyclical fluctuations in employment, but it is very hard to see any impact from technological change in the average rate of employment despite the many types of jobs destroyed by successive innovations (see Chart 4).
Chart 4. There is little evidence of technological unemployment over the long term
Source: Bank of England, abrdn, August 2023.
The combination of productivity enhancement and job creation has historically been much larger than job destruction. Having argued for a positive productivity impact from AI, the adoption of new technologies will also help create new jobs. These might be directly linked to the new technology – for example, training AI models – or in completely unrelated sectors that emerge as a result of the new spending power created by higher productivity.
Sectoral impacts will not be distributed evenly
Even in the best-case scenario, there will be some sectors severely damaged by AI, while others will benefit hugely. In the near term, we believe the winners will fall within three categories: enablers (e.g., chip manufacturers), scalers (e.g., tech-platform businesses), and early adopters (e.g., software companies).
In the long term, sectoral effects are more speculative and will depend on the way the technologies are regulated. But winners (from the perspective of capital owners) are likely to be sectors with large numbers of knowledge workers like finance; administration-heavy sectors such as education, healthcare, and law; and those with technical but repetitive tasks such as high-end manufacturing. By contrast, labor in these sectors could be under significant pressure (Chart 5). Sectors with a seemingly lower ability to increase productivity using AI may include manual and outdoor labor, the hospitality sector, and personal care.
Chart 5. Estimate of the share of industry employment exposed to automation by AI
Source: Goldman Sachs, August 2023.
Governments and regulators have a lot of catching up to do
The institutional arrangements under which the adoption of AI will take place are crucial. Norms, laws, and taxes all shape the pattern of income and wealth distribution. These social processes are in part endogenous to economic outcomes. Trade unions and friendly societies arising from the First Industrial Revolution were a response to the perceived inequality of the Engels’ pause. In turn, those institutions helped bring that period of stagnating living standards to an end.
Similarly, new institutions and regulatory frameworks will need to be developed in response to the impact of AI. However, governments may face a pacing problem whereby the rate of innovation is so rapid that policymakers struggle to keep up. AI regulation is still at a very early stage with the White House having already introduced a non-binding AI Bill of Rights. These nascent regulatory efforts are focused on human oversight of autonomous systems, responsibility for AI decision making, transparency of decisions, privacy, and bias. We anticipate a wave of AI regulation over coming years, although different countries will strike different balances between the strength of regulation and incentives to innovation.
Because transformative innovations are by their nature hard to predict, it is unlikely that childhood education systems can equip future workers for all the appropriate skills demanded by the future labor market. Instead, adult education and retraining is likely to be especially important to help workers acquire the necessary skills and help displaced workers find new work. A flexible labor market combined with a generous safety net could be the institutional arrangement that countries gravitate towards.
Higher rates of taxation on capital and even a citizen’s income are potential tools to more broadly distribute the gains to capital and ameliorate the potential harm to labor’s share of income from AI.
The geopolitical angles
Given the likely economic significance of AI, its many dual use (military and commercial) applications, and the location of leading-edge hardware production, the AI revolution is also likely to have a significant geopolitical angle.
The training and running of AI systems relies heavily on the most advanced graphics processing units, which are designed in the US and built in Taiwan by a very small number of firms. This geographical and firm-level concentration means hardware production is already being used as a tool of geopolitical influence, for example via controls by the US and its allies on exports of advanced chips to China. More speculatively, blockades or military action that threatened chip exports from Taiwan could significantly impede AI development globally.
The development of AI software may also become something of a cyber arms race. The physical location of software developers, and the location and security of intellectual property, is likely to be politically sensitive. We are starting to see this play out, with the UK’s decision to include China in its forthcoming AI summit in November criticized by Japan, the US, and the EU. But the more diffuse, mobile, and intangible nature of AI software, as opposed to the hardware on which it runs, means governments’ ability to exert control is more nebulous.
The values embedded in AI decision making (for example, around gender or ethnicity), and the uses to which AI is put (for example, surveillance, political influence, or even battlefield use) are all likely to vary between countries. There are initial efforts at international co-ordination on these issues, such as via the United Nations’ AI & global governance platform. But the fractured global economic and political system means countries are likely to pursue digital sovereignty.