Will AI transform economic growth?
What would need to be true for AI to significantly raise the growth rate?
What impact will AI have on the economy? Will it shake advanced economies out of their productivity slowdown? Will it steal our jobs, or will it leave us busier than ever? Will it leave us all living a life of luxury, or will it enrich only a small elite?
I don’t know the answer to any these questions, because I don’t know what AI will and won’t be able to do in future. I don’t even know if *anyone* knows the answers to these questions. The people who build AI models ought to know best, but share prices give them a strong financial incentive to tend towards optimism.
But although we don’t know how AI will develop, we can say a lot about what economic impact it might have. There have been transformative technologies — and technologies that have fizzled out — in the past, and they can give us clues as to whether and how AI might change the economy.
What I want to do here is not predict the future, but set out some tests or markers for how AI might change the economy. What are the routes by which AI might increase our prosperity or cause economic pain? What would need to be true, in terms of AI’s capability and use cases, for it to transform the economy?
Before I dive in, here are a few ground rules I’ll try to follow.
First, this isn’t meant to be a pro- or anti-AI piece. So much of the debate about AI is focused on whether it’s the greatest ever technological breakthrough or totally useless, whether it will kill or save us all. Those things might be true, but I think there’s also a huge space in between triumph and disaster which is worth exploring.
Second, try to focus on economic impacts. It’s possible that AI might wipe out humanity or inflict other enormous damage. That debate is worth having, but economics is a poor vehicle for discussing existential risk to humanity, so it won’t be part of the debate here.
Third, try not to extrapolate too much from the present. It would be a mistake to judge AI only on what it can do now — remember those predictions that the internet or mobile phones would never take off? But would also be a mistake to assume that it is guaranteed to follow a path of exponential improvement, or that it will be as useful as it seems.
Because this is a big topic, I’m going to break this down into a small series featuring three separate posts:
1. Will AI transform economic growth?
2. Will AI steal all our jobs?
3. Will AI make the few or the many rich?
You’ll only get AI’s impact on economic growth here, but hopefully I’ll find time to write the other two pieces soon.
1. Will AI transform economic growth?
In most advanced economies, the pace of economic growth has slowed down over the last decade and a half after decades of steady expansion. There are many possible explanations for this slowdown: ageing populations, restrictive fiscal policies, misdirected investment, energy scarcity and so on. But the underlying reason may be that we’ve stopped inventing and applying new, useful technologies as quickly as before. In this context, AI is an exciting proposition — a new technology that might break rich countries out of their economic slump.
The economist Thomas Philippon recently proposed a new way of rationalising our recent economic slowdown. Looking at very long term economic data, he concluded that economic growth is normally linear, not exponential, over time — that is, the rate of growth gradually slows as it proceeds. Except, according to Philippon, for three major transformations, when new technologies transformed the rate of growth in the most advanced economy and set economies onto a new, higher growth trajectory. The first two of Phillippon’s transformations were in Britain, in the late 1600s (when the initial conditions for the industrial revolution were set) and then in the early 1800s (when steam really took off). The third came in the USA around 1930* and is linked to electrification and widespread use of oil. Could AI be part of a fourth transformation, which permanently shifts the growth rate upwards?
Even if you don’t accept Philippon’s analysis — and I’m very much on the fence** — it is still common to view long run economic growth through the lens of general purpose technologies (GPTs). A general purpose technology is one that is so widely used that it touches on many different parts of the economy, not just the sector they originate in, and has a big enough impact to raise the rate of economic growth. If it is to live up to its billing, AI will at the least need to become a GPT (and no, having “GPT” in your name doesn’t count).
What does a general purpose technology look like? Take electricity, one of the most obvious and important GPTs. The electricity supply industry contributes less than 1% of the total UK economy (outside of energy crises), but electricity’s influence on the economy is many times greater. The breakthroughs enabled by electricity — cheap light, the internet, much of modern healthcare — are far too numerous to list here. It would be easier to list things we can do without electricity than things we use it for. Well over a century after Edison’s breakthroughs, electricity is still driving innovation in the economy; much of today’s shift to net zero, for example is built on electric cars and heating.
One of the key qualities of general purpose technologies is that they enable other breakthroughs and uses to be built on them. We didn’t just switch from gas lamps to light bulbs and then stop — we have continued to develop new uses for, and indeed new general purpose technologies built upon electricity ever since. This is a key economic test for AI: will it lead to breakthrough upon breakthrough, use case upon use case, in many parts of the economy? If it is to have a very big effect on economic growth, it will have to. You cannot judge the economic impact of AI solely by the turnover or share price of the AI industry itself.
In this light, it is worth asking how more recent candidates for general purpose technologies have fared. Computers and the internet are obviously general purpose technologies. They are used in most aspects of our lives — shopping, gaming, banking and so on — and they continue to have breakthroughs — including generative AI — built upon them. The internet has been with us for over 30 years, but it’s possible to argue we’re still at a fairly early stage of its exploitation.
The strongest recent candidate for a general purpose technology is probably the smartphone. Smartphones are widely used, and they’re a lucrative industry in their own right, but they also have wider applications. The location services built into smart phones are potentially transformative, and their portability and flexibility offers the prospect of a wide range of uses. However, the case against smartphones as a GPT is: are they really that transformative? Mapping software aside, are they really just a more portable computer? I’ll leave that one to the reader to decide.
What the internet and smartphones have in common is that they have coincided with periods of relative economic disappointment. They have been the champion technologies during an era of slower productivity growth, certainly relative to the post-war years of rapid oil- and electricity-fuelled growth. There are various reasons this might be, one of which I’ll return to momentarily, but it helps make a useful point: a technological transformation does not always result in an economic transformation.
***
Another way to think about AI’s impact on economic growth is through a supply and demand lens: does it let us produce more stuff per hour, and does it let us make stuff that’s useful?
The productivity side of AI feels pretty intuitive and immediate. AI can do things in seconds that can take humans hours. If the only measure of productivity was how many words you could write per hour, AI would be utterly transformative. Its ability to write code, to generate pictures and videos, to create songs is clearly very striking. But the economy also features more difficult tasks than writing.
The bigger tests are things like: can AI help us diagnose illnesses more rapidly? Can it do legal or accountancy work in place of humans, or at least save lawyers and accountants time? Can it increase the rate of production at manufacturing plans? And do the time savings — which are obvious — outweigh the costs, such as re-designing processes or checking for mistakes? I’ll make no attempt to answer these types of questions — I don’t think anyone can really know until an AI has actually done these things — but they seem to me both like things AI plausibly could do, and also very difficult for AI to do. Economies are characterised by greater complexity as they get richer, and this could make it harder for AI to complete processes from end to end. Or it could make AI even more valuable.
But as well as productivity, we also need to look at the other side of this: will the things AI makes actually be useful? Innovations only grow the economy sustainably if people actually value the things they produce. Some things AI might be able to do — diagnosing illnesses, automating administrative work — are obviously useful. But a lot of the immediate outputs are less obviously useful. Being able to produce thousands of pages of writing might be useful for navigating England’s planning system, but might not be the most direct route to raising productivity here. Being able to produce huge numbers of songs may not be all that useful if they’re all mediocre — we tend to listen to a few songs we like many times rather than lots of different songs once.
So there’s a second test for AI here: it not only needs to increase productivity, but it also needs to be able to produce — or help produce — useful things. The more useful things it can do with as little human input as possible, the greater its economic impact is likely to be. It would be good practice for proponents of AI to be more specific about the exact use cases they envisage, and how they see them working. As Soumaya Keynes puts it, the “scale, scope and speed” of these use cases will determine how an economic impact AI will have.
But even if AI produces a whole range of new, useful stuff, there’s a further challenge: can you get people to pay for it? This is one of the issues that the internet, smartphones and other recent technologies may have suffered from in economic terms: we might value them, but do we pay enough for them?
The giants of the internet era have found ways to make money from their products in a variety of ways. Google and Meta have focused on advertising spending by businesses, and along with Amazon have become the key intermediaries of the internet. Apple and Microsoft have taken the more traditional route of selling goods and services to consumers and businesses. But for those one step down the internet food chain — the news websites, the musicians, the app creators — getting paid for digital goods and services has been harder. There are plenty of people who make a living on the internet, but these markets seem less valuable and more precarious than one would hope. There is a vast amount of consumer surplus associated with the internet, which is great but doesn’t always help to pay the bills.
This is a challenge AI will have to tackle. Will people pay for AI models themselves? My guess: consumers broadly won’t, businesses will and that will be enough for a handful of AI companies to make it big. Much more importantly — and this is my third test — will it be easy for people to get paid for the things they make with AI? I could argue this point either way. There are plenty of goods and services AI could help produce — in healthcare, professional services, manufacturing — that have easily captured markets. But if a lot of the output of AI is intangible — or if the barriers to using AI are fairly low, allowing market prices to collapse — it is possible it may suffer from the same payment problems as the internet economy.
Finally, related to the question of usefulness is another nagging question: will AI be able to do useful things safely, in a way that’s acceptable to the public? The promise of self-driving cars looms large in my mind here. If we’d given the self-driving cars we have today to the Victorians, they would surely have used them. But we have set a much higher bar for their use, and developed complex, human-operated road systems that they need to integrate in to. This emphasis on safety may be a good thing — it seems a perfectly reasonable thing to prioritise as you get richer — but it probably does hold back raw economic growth.
Given AI also comes with scope for significant public concern over safety (as well as job losses and a concentration of wealth), it could also be vulnerable here. I suspect AI might be slightly more insulated from this problem than self-driving cars — after all, much of the time you might not know you’re using AI. Nonetheless, the economic impact of AI may depend in part on how much we really want the growth, and what we’re prepared to stake to get it.
I don’t, of course, have a definitive answer on what the economic impact of AI will be. But I think it is pretty clear what AI will have to do to significantly raise the rate of economic growth. First, have an impact across the whole economy, not just within the tech sector. Second, be useful as well as productive; the more useful things we can do with AI, the bigger the economic impact. Third, find a way to make people pay for the fruits of AI. And maybe try to get the general public on board, to avoid getting banned or sparking a backlash. If it can do all that, AI might be the best thing since electricity, or steam, or whatever your favourite GPT is.
Notes
* 1930 is a surprising year to associate with an increase in the growth rate, but long term growth rates move in mysterious ways. It’s also worth noting how Phillippon’s timeline doesn’t fit at all with the classical view of the industrial revolution (that it started around 1760).
** My own view is that Philippon’s finding is interesting and plausible, but it relies on fitting a pattern around a large amount of data, and is hard to back up theoretically
*** Remember Phillippon’s model has TFP growth as linear, so the rate of TFP growth has been slowing ever since the 1930s in his work