Merryn Talks MoneyWhy AI-Driven Productivity Is a Decade Away in the UK
Economist Diane Coyle, currently the Bennett professor of public policy at the University of Cambridge, analyzes artificial intelligence’s potential economic impact—from people’s jobs to their productivity—on this week’s episode of Merryn Talks Money.
Merryn is getting her own newsletter.
Sign up here: https://www.bloomberg.com/account/newsletters/merryn-talks-money
Bloomberg Audio Studios, Podcasts, Radio News.
Welcome to Maren Talk's Money, the podcast in which people who know the markets explain the markets.
I'm Maren sum Zetweb.
This week, we're focusing on artificial intelligence and the role we really hope it's going to play in boosting productivity. For a while, it seemed that the consensus was that AI was going to revolutionize everything. Last year, Goldman Economists estimated that AI would increase annual global GDP by seven percent over ten years. The IMF predicted that AI has the potential to reshape the global economy. Look at pretty much any report from any boog organization and you will see something very similar, the idea that this will change everything. Here in the the labor government clearly believe it too. They spent a good part of the recent Investment Summit positioning themselves as a champion for AI, called it an opportunity, saying that the country needs to run towards it. Not everyone is so convinced that AI is a silver bullet for all our productivity. Rose and I'm afraid that we are cynics on this podcast. We rather believe that too, and so does this week's guest Diane Coyle Professor. Dame Diane Coyle is the Bennett Professor of Public Policy at the University of Cambridge.
The first book of hers I read.
Was GDP, A brief but affectionate History, and I often refer to that even now. Her latest Worse Cogs and Monsters What Economics is and what it should be? That explored the challenges for economics in the context of the digital transformation.
I particularly liked the title.
She's got a new book coming out in April called The Measure of Progress, and maybe we'll get her back on to talk about that then. For now, her current research focus is on productivity and on economic measurements.
Earlier this year she.
Authored a very good article that caught our eye, Don't bank on AI being a quick fix.
For elusive growth? Diane, welcome to Merin Talks Money.
It's a real pleasure.
And now one of the reasons that we asked you to come on, apart from all these wonderul books and writings that we've talked about, was an article that caught our eye earlier this year called Don't bank on AI being a quick fix for elusive growth? And rather like a lot of other people on this podcast, we've talked a lot about how everything's going to be absolutely fine as soon as productivity pops up, and some somehow, one way or another, technology and AI in particular, will do that for us. So we will go from productivity growth being as a miserable one or one and a half percent across the Western economy is up to two and a half maybe more, and all the problems will disappears going to happen?
Is it? Well? Technologies do drive productivity growth over the longer term, but they tend to be much slower than people often imagine when there's a lot of new and exciting news about what's coming along. And the classic example from economic history is electricity, which is a nineteenth century technology fundamentally, but it didn't affect productivity growth until at least the nineteen thirties. And AI might be a little bit quicker, but I think they're still looking at a decade or so before it starts to have a measurable impact on productivity, and there are lots of reasons why that uptake is slow. So I get quite concerned about the silver bullet claims that you referred to, because it raises expectations and they get disappointed, and then you get a sort of backlash against technology. AI is an amazing technology. It will be able to do some extraordinary things for us, but it's not going to fix productivity. Is ye.
Should we go back briefly to electricity, because that's very interesting and is that? Was that an infrastructure problem? Because obviously connecting everyone electricity is a huge infrastructure problem. Got to be expensive, it's difficult, and you can't immediately see the impact of it. So you get that presumably Jacob effect right where you've got to spend all the money up front and the productivity doesn't appear until everyone's connected, and then it comes. So that's why it takes many decades, did take many decades.
There's several reasons for the delay. Infrastructure is certainly one of them. So with electricity, it was building out the networks. With AI, it's building the data centers and the computational power, but also actually getting the data in order because AI eats data, that's its fuel, and a lot of the data that we have isn't available, isn't interoperable, isn't in a good shape, isn't high quality, And so there's a big challenge in that sort of intangible infrastructure that we're going to need for AI. But there are other investments that are needed as well. In the electricity example, it was building new kinds of factories because you can have a dynamo on each machine, and the assembly line was part of what made the electricity revolution create productivity gains. But then also once you had to build new factories, you needed to build new transportation networks to get workers to them. And even then there's a reorganization of work that's needed, so the different kinds of jobs in factories and so on. So it's infrastructure directly involved, it's other kinds of infrastructure and reshaping the economy. And then it's what actually did people's jobs involved from what are the skills that needed for that? So all of those things need to come together, and that's why you get delayed.
Let's go to the optimistic end first, So go back to that bit and talk about the main things that people expect AI to do for us. So if you divided into generative and predictive, it's really generative AI that is the thing that we really expected productivity to gain gains to come from.
Should they come.
Yes, and the models are still improving dramatically over time according to the benchmarks that the industry itself uses to tess that. So I've just come back from a week in Silicon Valley, and every time I've been there in the last couple of years, people have been saying, oh, well, the thing that the next generation of models will do for us is absolutely amazing, And then that has been true. So I think technical advances are still extraordinary. But then to think about the economic impact, you need to think about what will that actually do in specific activities in the economy. So we've got the language models, and they are clearly able to do many kinds of administative task, so they could create productivity that way. There are visualization models, so let's think about what tasks involve visual inspection, or where visualization would help improve productivity. There might be things about three D real time visualization that are going to become much more possible. So mapping from what the technology can do in general to what are the actual things in a workplace it might be able to do for people is part of the journey.
I suppose what I'm really trying to ask is, as an ordinary worker, an ordinary person, what is it that we will see change?
What will happen around us? I mean, we know.
What's happened with digitalization and the rise of the smartphone, all those kinds of things. We can see the difference around us. We understand how that productivity changed, right. We understand how the smartphone changed how we work. We understand how our computers have changed things. We understand how cloud computing has changed the way we work and operate, etc. But as generative AI moves forward, what is it that we will we will see and we will feel what will be different for us?
Let's think about it in two ways. What products are we going to experience as consumers that will make things much better? And for there, I'd be looking for time saving things, things that I don't like doing that AI and robots might be able to do for me. Hans Rostling a long time ago did a wonderful ted talk about the impact of the washing machine on people's lives and how extraordinary it's been. At the moment, AI can't do that because it's not embedded in robotics. It's abstract, it's not computers. Work is starting on that kind of embodied AI and down the line, if that can do my laundry for me with a super washing machine that will also fold and iron clothes. That's going to be something quite amazing in terms of daily life applications in medicine, so a lot of opportunities for medical scanning, monitoring, tailored personalized medications. Huge potential there to improve the multi of people's lives. So that's one way to think about it, and then the other way to think about it is all the process innovations it can help us to introduce at work that will save time do boring work much more cheaply, and then they're free up the time for workers to do things that are going to add more value for their customers. So if you think about the medical examples, well, it saved doctors time writing letters so that they can spend more time with their patients. If you work in manufacturing, is it going to speed up those processes or help you track waste or energy use in better ways than are currently available. And that's the kind of process innovation that allowed the just in time revolution and manufacturing productivity in the nineteen seventies and early eighties. So thinking about it in those two ways will give us some clues about where we can hope for ultimately productivity gains.
My AI reverse the trying trend from a consumer's point of view of having pretty much all the work that used to be done by companies when you buy something, or book something, or rain something being outsourced to you, and the maddening difficulty that that caused with me, might say is one of the big downsides of the product of the technology revolutions.
So far, it might. It depends on the incentives of the companies to do so, of course, but we've all paid that time tax of getting into voicemail hell and not being able to get talk to a human being.
Or chatbot and talking to a chat but I'm begging it to let you talk to a person so you can get to the end of the process.
But not happening.
So one model might be what's happened in retailing, or is happening in retailing where you used to go to the shop and somebody would scan your groceries and possibly even put them in a bag for you back in the old days, and increasingly that labor has been outsourced to us, and so now we've got the automatic checkouts where you do your own scanning and packing and you've got to persuade the machine that you haven't put the item in the bag the wrong way. But now there are stores where you just take the stuff off the shelf and go away with it, and so that then would be labor saving. How we measured these in productivity obviously has a different effect because the automatic checkouts and the shops are going to increase the measured productivity of the retailers because it's not measuring your unpaid labor in doing that scanning and packing yourself. So we don't know yet whether that AI enabled or you walk into the store and take it out is going to last. Maybe not, but that would be the kind of journey we might hope for. Elsewhere, there are some call center experiments with AI where it actually seems to be improving the experience that consumers have because the AI can train call center workers and gives them better scripts so they become better at doing what they're doing. But I think you know that's what we need to see, that's what's going to help us.
Interesting, let's go back to what you just said about the measurement of productivity, because it's a very interesting point that these making us do the work improve the productivity of the firm, and if you then take that work back in house using AI, it may make no difference at all to what looks like productivity. So the change has happened, but it's not caught in the numbers. And this is one of the things that you specialize in writing about how the numbers do not count to the change as digitalization and technology moves.
Toward, the numbers really entirely miss the digital revolution in our experience as consumers and workers. So there are all kinds of phenomenon that have been enabled by the digital from global production networks that rely on communications and logistics that have been enabled by digital to the way that many manufacturing companies wrap services around their products it's called servitization in the literature, to all the platform models and these free goods, and how do we take account of the zero price that we pay for the free goods? But what are the other implications of all of that just not captured in the economic statistics. It's always been the case that it's really hard to measure the major impacts of big innovations in statistics. The statistics are good at small changes, but not big changes, And so economists would also look at things like improved life expectancy or the quality of life as well to supplement GDP figures. But at the moment, this extraordinary change in how we lead our life since the smartphone arrived in two thousand and seven is pretty much invisible. And that's a lot of what my next book is going to be about.
Okay, And what are the solution as to that? How do we pick it up?
How do we change the way we measure I mean, in a very minor way. We get very irritated on this podcast a lot by GDP measurements because we're maddened by people constantly talking about GDP instead of GDPP ahead. But that's a very minor problem in the context of what you're talking about. So how do you reframe the measurement of economic progress?
Well, a lot of it is about collecting new data and new statistics, and we've got to use the technology itself to help us to do that. It is slowly starting. So on things like global production chains, the impact of a tariff on imports is now going to be different than it would have been in the nineteen sixties because to manufacture anything, you've got to import a lot of the components, but the data are starting to capture what's called value added in trade, how much do you need to import to manufacture particular kind of export. So that's slowly starting. But there's all kinds of new days to collecting credit card data, using administrative data like taxes, using mobile phone data, webscraping, and the research into how to develop statistics using these new methods is kind of in its early days. So I think we're probably in for a I don't know, twenty year gap before we've got a settled set of statistics on the digital economy.
Okay, So if we were in optimistic mode, might we say that, in fact, living standards are improving much faster than our current statistics tell us they are.
I think they're improving in some ways and for some people, faster than statistics tell us, But in other ways they are less good at telling us what's going on. The obvious example people always give is that statistics don't tell us anything about the environmental costs of recent economic growth, which are going to hit productivity because places will get flooded much more than they used to be. Agricultural productivity will decl line because soil quality and biodiversity is declining, so they will have real economic consequences, none of which are measured. So there are pluses and minuses, and we need really a broader framework to think about all of these things that are outside the traditional standardized manufacturing economy for which these statistics were devised in the nineteen forties.
Okay, that wasn't the answer I was looking for.
I was looking for you to tell me that the statistics were hiding progress, as in effect, you're telling me they're hiding a type of decline.
They're hiding both the hiding decline and progress. And if you think even about things like medical innovations, something that's now pretty basic, like the ability to do a really straightforward cataract operational hip replacement operation, which has become a much more standardized and straightforward process because of advances and technology. They're not big, shiny advances like generative AI models, but they're really important improvements in the quality of people's lives.
Well, let's move from that, just briefly, because no one's well on this for too long, to the NHS and the possibility the possibility that there are ways that AI might be able to help us with the NHS in a way that constant large lugs of cash do not.
Yes, So it's obviously one of the hopes for AI that it will enable big improvements in public service productivity and particularly in the NHS, And we already talked a little bit about some of the potential there. There are already quite well established gains to be made from having AI screen test results and provide that information to doctors or making decisions, a lot of potential for administrative simplicity. But to me, it's a question of organizational than it is of technology. It boggles my mind that it isn't already the case that the NHS has not mandated all of the hospital trust and GP surgeries to use interoper technology and standardized forms of data. So even getting to the basics of how easy is it going to be to adopt AI, we're not there yet. And it is clear that there's been underin investment in very basic technology. My husband was in hospital last year and we noticed that the nurses had to spend all of their time with their backs to patients, typing things into computers with these really user unfriendly kinds of interfaces that were very out of date. So is there potential for AI to fix that, sure, but that's going to need some investment of money, but also time to restructure those information flows. Within the NHS. There's an issue about data and data sharing, and there is real conservatism and risk aversion across the NHS and linking up people's data. Recent announcements about the NHS the NHS app bringing that together for individual patients. That's really welcome, but I think that fundamental lack of trust in the possibility of data sharing. How secure will it be? Who's going to profit from that? Will it be private sector companies or will you? Will patients see some benefit? That's a really big hurdle to overcome yea.
And not helped by the way that public trust in the NHS has fallen very dramatically over four or five years. Yeah, last four or five years, So how would one then trust them further?
Absolutely? And then there are questions about organization and financing. So there's lots of innovation in health tech. People are coming up with some amazing new products, but they've got to be paid for. So what's the business model that will enable the NHS to invest in those kinds of new innovations that might save them a ton of money, but their budgeting structures don't easily allow that, and then who's got the authority to make decisions? So AI is an information technology. It uses data and it creates useful information which will allow people to make decisions that can improve productivity. But that's going to be decentralized across the public services and including in health. They're quite hierarchical for reasons of accountability and also training. So how can the skills needed and the authority needed to take decisions be delegated. So there's no point giving nurses amazing information through AI products if they can't then do something with it. If they still have to call up doctor to get them to take the decision, they're still going to be bottlenecked there. So there's a whole range of barriers alongside this huge potential, and obviously the state of the NHS and the public confidence and it means this is one that we've got to crack as sin as possible.
I imagine briefly, when I was introducing the current government talking a lot about wanting the UK to be an AI leader and how important that is and how they can help the state can help drive that. What do you think the state's role is here in driving investment into AI or improving the sector in the UK.
So multiple roles. Actually, we do have a really good AI sector at the research frontier and in parts of AI, so the government has funded a lot of that research and should carry on doing so. Except it's got broad benefits and want to enable more startups. There's an issue about how do those companies grow. Many startups in tech still hope that they would be brought out by a big American company, and we need to be able to grow our own and understand which niches we have real advantage in and can export in. So that's again something that government needs to help with because public markets are making it very difficult for companies to grow across the board. But then there's a whole area of things about coordinating activities de risking investments. Is, for example, using public procurement in the NHS to encourage innovation and say there will be a market for new kinds of AI products. That it doesn't have to specify which ones are going to be brought or exactly which technological approaches needed, but the fact of it's called an advanced market commitment. In my world, and the fact that there will be a market through public procurement can also really help encourage us better technology and you know, public services can drive the take up of technology, get things to scale, make it cheaper for other users, demonstrate effectiveness. So this is an area where I think government action can make a big difference as long as it's realistic. So I think if the government to say we have this silver bullet and it's going to change things overnight, that that's very dangerous and they'll be backlash. But if it's got a realistic approach in you know, strategic industrial policy kind of framework, that could be really powerful.
Okay, interesting, and what about the concerns around competition and the idea that what with the you know, the amount of investment, the amount of data.
Except with that AI.
Progress required that you have a situation where a big tech this turns into big AI and we have the same problems that we've anti competitiveness problems that we've had with big tech and the aim as well, where's the government's were all there to preempt that.
So it's a big issue. We've now got the Digital Markets legislation that allows the CMA to take action around certain kinds of behavior by big tech companies that have got a dominant position. If you're looking at generative AI and the large models, that the scale needed at the moment is ginormous, so that problem of dominant domination of markets becomes even worse. So more of the same in terms of monitoring what's happening and looking out for things like algorithmic collusion between the different models remains important. There's a lot of scope for competition at the level of applications that use large models, and that's a nascent market with generitive AI, so there's a lot more scope to make sure that is competitive.
The other thing that I think is quite important to talk about is the energy hunger of AI. And you mentioned earlier that day is the fuel of AI, but of course that the fuel of AI is fuel. And you also we talked earlier about the externalities and bringing those into the equation. So if you start to look at the productivity gains that might come from AI, you kind of have to bring all that into the equation.
Absolutely, And among the most extraordinary stories recently have been Microsoft reopening the three Mile Island Nuclear plant and Google commissioning a whole fleet of small modular nuclear reactors. So I've seen charts where the prediction is that the use of AI as it grows will end up using all of the electricity generated in the United States, which clearly is not going to happen. So the energy supply might increase. There'll be a lot of incentive for people to come up with much more energy efficient types of models and AI processes, and that is also starting to happen. But otherwise it's a real constraint on the use of AI because this might not be the most efficient way we want to use our energy. There are lots of other uses. And when you read stories about development along the End four corridor not being possible because there's no scope for new grid connections due to all the data centers, then you start thinking, well, perhaps there'll be political reprioritization here. AI is not going to sweep all before it, so we certainly start need to start thinking about the energy efficiency. I don't like the way that Google, when you do a search now gives you an AI summary at the top, because that means you're getting the same information as before, but with turn times the energy.
Use you know, I hate that too. I've only noticed it relatively recently.
It comes up and you look at it and you know that that bit at the top that you didn't ask for is using. I'm not entirely sure what the figures are, but people keep telling me that a generative search takes a thousand times the energy of an ordinary search. As you look at that and you know you can feel the energy uses, that doesn't feel good.
It's not good. And so one, you know, really key imperative is keeping an eye on that energy use. And I think governments will actually face choices about about their energy grids, which are you know, public, key public policy decisions.
You've seen that in some of the EU comments already on AI have been looking at the energy factor. But I suppose the selver aligning, if you're that way inclined to think it's the self aligning. Is this resurgence of nuclear energy and the drive forward with SMRs et cetera. That seems like a good thing.
Seems like a good thing. And obviously nuclear is controversial with some people, but in a fully renewable world will need some baseline generation which nuclear can provide a zero emissions.
And if the very rich, very cash rich tech companies are prepared to be the ones who go first and finance some of the research into this, and that that seems like a finally a positive of the vast amount of cash that they've earned over the last decade.
It does.
Yes, Yeah, we've talked.
A lot about things that AI might be able to do that it could do. What we haven't talked about is things that people believe AI can do, which it can't. What is it that we might be expecting that is very unlike.
That's a good question. What do you have in mind?
I do know it if you're particular in mind.
I just every time someone talks about AI and they talk about, well, it can do this, where it can do that. It can change the NHS, it can redrive productivity, it can change the way we educate people. It can change the way we manage the time. It will change It'll completely change the way we run our agricultural economy, for example. You know, you hear nothing but transform, transformation or optimism. And because I'm not an expert in this, when I hear people telling me it's going to change the way we farm, it's going to change the way we go to space, it's going to change the way we mind mind for metals on the moon, etc. I'm perfectly happy to believe all that because this is not my area of expertise, and it seems entirely possible that if you can manage data this effectively, and if you can have models that can that can genuinely learn to do things that we haven't yet learned to do, then all these things could be possible. But I also know, because you fall victim to optimism all the time, that inside this belief that it can do anything, there's bound to be things that definitely can't do well.
It might do all of these things if you again look back to those historical examples. The railways enabled urbanization because they meant that cows didn't need to live in the city centers to provide fresh milk. They could live out in the countryside, and the railways the milk train would bring in the fresh milk. Who would have predicted that when the first railways were built. So, yes, the tech might do all kinds of amazing things. The question I always ask when I hear the hype is how is it going to do that? Given that this is an information technology, So what you need to talk about is how does having a different, better, faster flow of information, change what we do to produce the amazing outcomes that you're talking about, and a lot of the hype merchants actually can't answer the how question.
Well, it's one of those things we will gradually find out, won't we.
Yeah, And I mean as always that one of the main barriers is people don't like changing what they do. So anything that requires fundamental change and now you do your job with leadir life can can be very slow. It's partly fear because the kind of upfront cost of using AI is so high. Still, there are a lot of things for which it's still much cheaper to use the traditional forms of production and traditional human labor. And so even if there are amazing changes down the road, as we were saying earlier, it be pretty slow.
Do you worry about the job market? Do you worry that?
I mean, obviously, as you say, it will be slow. But in ten years will we have an unemployment problem or will the economy do it? It always does simply create you in different types of jobs, always has done so.
In klind towards optimism in terms of number of jobs. I don't think this will be a sudden change in the way that the wave of de industrialization in the late seventies and early eighties was, which obviously policy could not cope with. That are just too many people became unemployed too quickly, So I think this will be more gradual and ours adjustment in the labor market. I guess my bigger concern is about the quality of work and what people get paid. Because the benchmarks for AI are set in terms of how well does it match what humans do. That creates a kind of bias towards substituting for humans doing what they do at lower cost, and instead it would be better to think about benchmarks that improve outcomes that both humans and AI can deliver together. But if we go down the substituting for humans route, then the impact on wages and job quality would be something to worry about.
What's your greatest hope for AI when you look at ten years, What's the thing that will have been most transformative that will change Not things we can't say, the things think we can see, things we can feel.
I would hope that AI will have saved us a lot of time doing things that we don't like doing, to create time for us to do things that we really enjoy. So that this might be we've got more leisure time because technologies have, over history increase the amount of leisure and reduce the amount of work that we have to do. Or it might be that our work becomes more enjoyable we can focus on the more satisfying parts of that. So if I were a medical person, I would hope that AI had taken over all of my ADMIN and left me with shorter shifts but more time to spend with patients and give them the quality of care that I would want to give them.
Goodness, I tell you, well, we'd all like that.
Diane, thank you so much for joining ask It.
I'm fascinating, Thank you, Thank you lovely to talk to you.
Thanks for listening to this week's Mirin Talks Money.
If you like our show, rate to review.
And subscribe wherever you listen to the podcasts and keep sending questions or comments to.
Merror Money at Bloomberg dot net. You can also follow me and John or Twitter or X.
I'm at Marinus w and John is John Underscore X. This episode was hosted by me marenzum zet Web. It was produced by Someasadi production support and sound designed by Moses and m and special thanks of course to Diane Coyle
