If, as we believe, AI should be considered a general-purpose technology (GPT), then the implications for the macroeconomy could be huge.
There are generally considered to be three classic examples of GPTs, all of which had profound economic effects. The first took place during the Industrial Revolution with the development of the steam engine in the UK during the 1700s and 1800s. The second occurred with the introduction of electricity in the US in the early 1900s. And the most recent GPT wave took place during the late 1900s with the ICT digital revolution. Although it may not fit the definition on its own, the invention of the internal combustion engine in the late 1800s was, if not a fourth GPT, then at the very least a major technological innovation.
GPTs tend to affect economies in three phases. In the first phase, while the technology is still new and not widely used, the benefits to productivity are small. We discuss the reasons for the slow adoption of new technologies in more detail later in this section. An extreme example of this was during the Industrial Revolution in the UK, in a 50-year period known as “Engel’s Pause”, which was categorised as a time of strong innovation and capital accumulation but low productivity growth.
During the second phase, the technology is improved upon, the cost of implementation tends to fall and the technology becomes more widespread, eventually leading to significant productivity gains. Finally, in the third phase, the law of diminishing marginal returns takes effect. At this point the pace of improvements and rollouts slows, causing productivity gains to taper off.
What this means is that, historically, it has often taken decades for productivity gains from GPTs to materialise. However, this delay has been shortening. This is illustrated by the steady decline in “adoption lags”, namely the amount of time for countries to implement a new technology after it is invented. (See Chart 1.)
Chart 1: Technology Adoption Lags (Years)
Source: Comin and Mestieri (2018)
Chart 2 shows productivity gains from past major technological breakthroughs in those countries driving the progress. The boost to annual productivity growth in the UK and US from steam and electricity averaged a modest 0.2% or 0.3%, in part because, especially in the case of steam, the gains came through over such a long period. But the effect of the ICT revolution was bigger, in part because shorter impact lags meant that the gains were more compressed. The US, which saw far by the biggest gains from ICT developments, is estimated to have enjoyed an average boost to annual productivity growth between 1995 and 2005 of 1.5 ppts.
Chart 2: Contributions to Labour Productivity Growth (% Points Per Annum)
Sources: Crafts (2020), Capital Economics
We know that GPTs can have major implications for economic development, especially for those countries at the forefront of innovation. The fact that the US was at the forefront of major technological inventions such as electricity and the internal combustion engine in the early 1900s – and reaped the ensuing productivity gains – helped it to surpass the UK as the world’s biggest economy.
The invention of a GPT does not necessarily lead to strong productivity gains, though. Historically, the largest productivity gains have taken place in instances when technological advancements coincided with political and/or social changes which allowed the technologies to be exploited to their full potential. For example, it has been argued that Great Britain was able to lead the Industrial Revolution because it was better able to exploit its resources, including labour, thanks to the social and political reforms during the Enlightenment.
The impact of previous GPTs on labour markets has generally been positive. Major technological advances have caused significant short-term friction in the labour market. But workers have ultimately been able to transition to new higher paying jobs. This is because the GPT-related increase in productivity has either increased demand for certain goods or services or has created demand for new sectors altogether. A recent MIT report estimated that 60% of workers in the US are now employed in occupations that did not exist in 1940.
Unsurprisingly, the effects have not been homogenous across sectors. Workers have lost out in cases where a GPT has been a perfect substitute for their work, such as the agricultural workers during the Industrial Revolution who were replaced with automatic threshing machines. On the other hand, workers have generally benefited when a GPT has helped to augment the type of work they do.
Similarly, GPTs have led to the loss of some industries, but also the creation of entirely new ones. The invention of railroads eventually killed demand for horses, while electricity did the same for whale oil. But with the advent of railroads also came the need for workers to drive the trains, man the carriages and work at railway stations.
BOOSTING PRODUCTIVITY: THE EFFICIENCY GAINS
So, if past GPTs are anything to go by, the AI revolution will deliver huge changes across economies, including a big boost to productivity. In this section, we think through the channels through which this might occur and what we might expect in terms of the size and speed of productivity gains.
The most straightforward way in which AI will boost productivity is via one-off efficiency savings (by which we mean doing more with the existing resources or doing the same with fewer resources).
In some instances, these will be achieved by AI directly replacing humans because of AI doing more efficiently what humans currently do. In other instances, the savings will be achieved by AI helping humans to become more productive in their existing jobs, freeing up time to spend on other, potentially more productive, tasks. Moreover, AI could free up resources other than human capital; for example, the use of shared driverless vehicles would allow land used for parking to be utilised more productively.
Chart 3 helps to illustrate the channels through which these productivity gains might be achieved. In general, output per hour worked can be increased in three ways: i) raising the amount of capital per worker ii) raising the quality of the workforce and iii) raising multi-factor productivity i.e. the efficiency with which labour and capital are combined.
Deploying AI, which is basically a new form of capital, has some impact on productivity through raising the amount of capital per worker. This is most likely to be through the addition of new AI software, although AI might also slow the depreciation of non AI-capital by enabling more preventative maintenance.
But the bigger impact of AI is set to be the boost it gives to multi-factor productivity by, for example, facilitating better working practices. There are many examples of this with past technologies. When electricity replaced steam in factories, workers no longer had to be arranged around the centralised power source, enabling more effective production lines. The invention of the steam engine and the need to manage railroad schedules led to the adoption of standard time, helping people to use their time more effectively. And the development of email and other digital communications facilitated outsourcing and offshoring. An example of how AI might improve working practices is by using predicted demand to improve how firms determine and vary their staffing needs.
Chart 3: How Labour Productivity Growth Can Be Achieved
Source: Capital Economics
The potential efficiency gains resulting from all this are huge. Because AI is still at a relatively early stage, studies of its effects are scarce and, because the technology is developing so quickly, soon go out of date. Nonetheless, those that are available point to a significant boost. For example, an April 2023 study of call centres, showed that access to AI assistance increased the productivity of agents, as measured by the number of customer issues they resolve per hour, by 14%.
Admittedly, in some cases, the application of AI is a zero-sum game and, while boosting productivity for some firms at an individual level, will not do anything to help at the aggregate level. This includes the use of AI in targeted online advertisements and in the automated trading of financial instruments.
Equally, though, it is easy to imagine the deployment of AI in some sectors having a far bigger positive impact. For example, in the time previously taken by a lawyer to draw up and check one contract, they might be able to manage ten contracts which had been provisionally drawn up by AI. There have been countless studies that crunch laboriously through each type of job, classifying what share of each could potentially be automated by AI and with what effect. But there is so much uncertainty that precise estimates are meaningless. A simple back-of-the-envelope calculation based on some plausible assumptions makes the point that this could have major implications for productivity across a wide range of job type.
BOOSTING PRODUCTIVITY: AIDING INNOVATION
A more uncertain, yet potentially even more important, way in which AI might boost productivity is if it helps to further technological progress. This could occur if it serves as a new general purpose “method of invention”. For example, the greatest impact of optical lenses was to aid R&D via microscopes and telescopes, rather than helping people see better.
AI could sift through the vast amount of research already published to search out further research avenues. There are literally millions of scientific papers published each year and there is no way that human researchers can digest everything in their field. But AI can, and it can look for potential relationships and suggest the most viable research streams. Indeed, one of AI’s advantages is thought to be its ability to follow unexpected threads and “think outside the box”, drawing researchers’ attention to patterns and connections which they might otherwise overlook. Meanwhile, AI can also help by using its knowledge to predict the results of real-world experiments.
AI has the potential to boost innovation across a very wide range of fields, from drug discovery to education to transport. Imagine all the potential applications of using AI in the field of material science, where it could choose the best new chemical compounds to solve real world problems. It could help to develop new pesticides for the agriculture sector; new corrosion-resistant metals for the transport sector; or compounds that absorb pollution to aid the green transition.
Using AI to boost innovation would not just give a one-off boost to the level of productivity; it could boost productivity growth permanently. We could even end up with a virtuous cycle whereby AI helped itself to improve at innovation.
Admittedly, none of this applies if the pessimists are right and there are simply no new good ideas left to be had. Economist Robert Gordon argues that there were only a few truly fundamental innovations (including instant communications and using power on a mass scale) and that these have now been made. If true, AI or no AI, the scope for further technological progress is limited.
More likely is that new ideas are still possible, but as science becomes increasingly complex, it requires more effort to find those ideas. The world can still make technological progress, but to do so it needs to increase research effort to offset the fall in research productivity. According to one NBER study, the US must double the amount of research effort every 13 years to offset the increased difficulty of finding new ideas. If this is true, then AI is still a powerful tool to aid innovation. But at least part of its power will lie merely in preventing productivity growth from slowing further as research productivity continues to wane.
But what if there were still some things that AI could not do? Explosive growth would still be possible if spending shifted towards the AI-intensive sectors in response to their fall in relative price (so-called “Baumol’s cost euphoria”). For this to occur, real spending would need to rise more rapidly than relative prices were falling, to push up the share of nominal spending in these areas.
More likely, though, is that the areas that AI was able to automate fully would not be perfect substitutes for those which it cannot. (Indeed, as technology has made manufactured goods cheaper, consumers have spent a greater share of their nominal income on low-productivity services.) So, as part of the economy became increasingly automated, the unautomated, less productive part of the economy would grow as a share of GDP, limiting overall growth. Growth would still get an initial boost from automation; but over time, it would slow as it became increasingly constrained by those parts of the economy that were essential yet hard to improve.
THE BARRIERS TO ADOPTION
In theory, then, AI has the potential to be a real game-changer for productivity growth. But potential gains are one thing, realised gains are quite another.
The productivity boost from past transformative technologies has generally been more drawn-out and less dramatic than might have been expected considering the importance of the inventions. And economists have been puzzling for years over why the digitalisation of the economy over the past couple of decades (including the advent of mobile internet, cloud computing and the Internet of Things) has been accompanied by such weak productivity growth. Nor is there any evidence yet of any boost to productivity growth from early-stage AI. Despite the recent fuss about ChatGPT, AI has been around for a while; it is almost a decade since reports began to predict an AI-related surge in productivity growth. Yet productivity growth, particularly in advanced economies, has been sluggish. Indeed, since pandemic-related distortions have washed out, productivity growth in G7 countries in the past couple of years has been below its average since 2005. (See Chart 4.)
Chart 4: G7 Output Per Worker (% y/y)
Source: Refinitiv, Capital Economics
One potential explanation is mismeasurement. There is certainly something in this. But research generally concludes that this accounts for only a small part of the productivity paradox.
An alternative explanation is that the full effects of GPTs, including AI, cannot be realised until waves of complementary innovations are developed. To make the most use of the Internet, for example, we also needed the development of cloud computing and big databases. To make the most use of AI, we will need other complementary innovations. One example already developed is computer vision, whereby computers analyse images and video, which can be used in applications from autonomous vehicles to medical diagnostics. There will be other complementary innovations which we cannot yet imagine.
The other key reason why technology does not translate into immediate productivity gains is that, even if the technology is available, many firms do not deploy it immediately or effectively. However, deploying AI is not just about plugging in some software and letting it do its thing. Companies need the accompanying capabilities, such as databases, data management systems and IT specialists.
Moreover, making the most of a new technology – AI included – requires significant organisational and process changes to make full use of the accompanying capital investments. For example, adapting to online selling was not just about setting up a website; retailers had to change all their logistics to include new warehouses, delivery channels etc.
Indeed, most of the factors that have slowed the exploitation of productivity gains from previous technological steps forward are just as applicable now as in the past. These barriers can be broken down into those that are internal (and within a firm’s control) and those that are external (which a firm cannot directly influence).
Of the internal barriers, the most important for many firms will simply be the costs required, especially in the near term when they will be highest. Indeed, a key uncertainty regarding the speed of adoption of AI is how quickly costs fall.
As for external barriers, regulation/regulatory uncertainty is likely to be key. Governments may place limits on how AI is deployed in practice. And firms may face, or fear facing in the future, costs associated with data security and privacy regulation.
The upshot is that the pace of AI deployment is not simply dependent on technical feasibility; instead, it will depend on various factors including costs, regulatory barriers and the skills of the workforce. Policy will also play an important role in all of this. Indeed, AI is most likely to deliver a big boost to productivity if several factors come together, including a rise in investment, reskilling of the workforce, and a balanced regulatory regime.
QUANTIFYING THE BOOST
So where does this leave us? While AI’s potential benefits are large, a “big bang” of productivity gains seems unlikely; it would be more realistic to expect any boost to come through gradually, as in the past.
We think that the impact of past technological advances is the best guide to the likely impact of AI in the coming years and, given the nature of AI, the ICT revolution is probably the best one to look at. Data show productivity gains during the ICT revolution reached 1.5% per annum for the US, so that seems like a reasonable guide to what is achievable for a country that is at the forefront of developing and deploying the technology. It is even possible that the gains from AI are bigger than this; whereas many applications of the digital revolution were about improving the consumer experience, we think that AI has many more business use cases that will directly boost productivity.
One thing to note about the ICT revolution, however, is that the US was by far the biggest beneficiary in terms of productivity gains. The largest euro-zone economies, for example, saw little boost, while some other countries (including the UK, Canada and Australia) were in-between – experiencing some productivity uplift, but not on the scale of the US. This reflects differences in the policy environment as well as economic structures that affected the adoption of new technologies.
The lesson is that productivity gains from AI are by no means guaranteed and will depend on whether countries have the factors in place to help them utilise AI effectively. Below, we explore in more detail the extent to which countries away from the technological frontier will benefit from AI.
AI is far enough advanced, with sufficient business cases, that we can be confident that there will be a significant productivity boost at some point. However, while AI’s adoption lags may be shorter than for previous technologies that required significant replacement of physical capital, there are still various reasons why its adoption is likely to be slow and fitful. As with past technologies, the productivity boost is therefore likely to be more of a slow-burn than a miraculous surge. Any productivity boost is more likely to be a late 2020s and 2030s story than something to be expected over the next few years.
To the extent that the past GPTs are a good guide to what to expect, we might expect a boost to annual productivity growth in countries which successfully and fully deploy AI of 1.5% for a period of a decade or two. However, not all of this will necessarily be picked up in the official GDP figures. And the extent to which countries successfully utilise AI, and therefore benefit from productivity gains, will differ significantly.
Jump to the table of contents
Jump to the section "Blowing bubbles? AI's financial market implications"