Are We Investing Too Much Money on AI? Oxford Economist Weighs In
Carl Benedikt Frey, the Oxford economist who co-authored the Future of Employment study and wrote "The Technology Trap," asks whether the trillions being invested in AI will pay off. He compares AI with the First and Second Industrial Revolutions and the computer revolution, and discusses stagnant growth, AI's unpopularity and what governments in Europe and the United States should do. Stephen Horn and Laila Rizvi conduct the interview.
Transcript
Trillions of dollars are being poured into the AI promise of a global economic transformation. We're being sold a productivity revolution, faster output, greater growth, and unprecedented prosperity. Okay, but when? Real, tangible gains are barely showing up in the data. So is the return on investment actually going to be smaller and slower than promised? And could these billion-dollar bets bankrupt the companies making them and take the global economy down with it?
Globally renowned economist Carl Benedikt Frey co-authored the famous Future of Employment study, which garnered a thousand headlines about robots taking over our jobs. But his book, "The Technology Trap," showed that revolutionary technologies can take generations to pay off. Look back, during the Industrial Revolution, machines transformed everything while working people's wages stood stagnant for 50 years. And the railways, they changed the world, yet ruined the people who funded them.
So today, Stephen and Laila are asking Frey, are we being promised or expecting a little too much from AI? All right, let's get to it. Looking forward to your conversation today. Do you think that we're expecting or being overly promised on all the things that AI can deliver?
I think the idea that it's a magic bullet, and we'll save everything, is not going to be the case, but that's never been the case with technology. And I think you have to acknowledge that AI is one of the technologies, and maybe one of the most important so far, but there are going to be subsequent things that we can't even imagine that AI leads to.
I mean, I think it'd be interesting. And, you know, we talked about it, in effect, being a journey, you know, taking time for the benefits to come through, and that's what we've seen. And I think that's what we are seeing now. We are seeing benefits coming through, we are seeing productivity increases, and we are seeing return of investment. But it's such a short timescale. You know, we're talking about two or three years, where we're talking Industrial Revolution, you know, was 300 years ago, right?
Yeah.
So I think it's a fair point, and we'll have to see how this plays out. And I think, coming back to something we said before, nobody knows the answer.
Right. Well, we hope that our guest today does. So let's get to your fascinating interview with Carl Benedikt Frey.
Do you think the trillions of dollars being invested in AI will pay off in the next five years?
I don't think it will pay off for the firms necessarily that is developing the technology. I think a more interesting question, though, is, is it going to pay off for society and for the economy? My hunch would be probably not, because I think that even though AI has a lot of promise, we also need these institutions to adjust to this new technology. And if we look historically, it has always taken many years and even decades for new transformative technologies to really reshape society and the economy.
And you've said that previously, that the productivity growth from AI won't actually match that of the computer revolution, but in this day and age, it's the companies that benefited the most from the computer revolution that are taking this bet on AI, like Amazon, Facebook, Google, Oracle. Do you think that they are going to be wrong in the size of the bet that they're taking?
The enormous bets that are being taken are essentially being taken on a promise of general artificial intelligence, or superintelligence even. They're being taken on a promise that the leader in the field is essentially going to capture that market. I struggled to see a situation where all the firms investing enormous resources into this race is going to come out profitable on the other side of this.
Can I take you back a little to a question from your book, "The Technology Trap," where you were talking about the First Industrial Revolution, and the time that it took for society to benefit from that revolution? I grew up in Manchester. And, you know, when I was growing up there, most towns still had, they were derelict by now, but hundreds of mills that were a sort of testimony, if you like, to the reshaping of the industrial process that came out of the Industrial Revolution.
So you say it took a long time for society to benefit. I put two points to it. One would be, potentially, society as a whole didn't benefit. You know, you still could go a hundred years on from the Industrial Revolution to see abject poverty in those, you know, cities around the North of England. But also, whilst society as a whole took time to benefit, there was income inequality straight away. So there was a lot of money created very, very quickly. So are we likely to see a same thing with AI, where it takes a long time for society as a whole to benefit, but there's income inequality on the way?
We've been through very different episodes of technological change in the past as well. And so the First Industrial Revolution that you mentioned in its early days essentially centered on the mechanization of textile production. And artisans lost their jobs as the factory system proliferated. There were many people that lost out. And what you see on the macro scale is that the labor share of income declines during this period, and wages at the lower end of the income distribution are probably declining as well.
The Second Industrial Revolution, though, was very different. And so what you see is that technology, electricity, and the internal combustion engine is almost lifting all the boats. You see very broad-based improvements in standard of living, and the income of the statistics also actually understates the transformation that takes place.
Well, I think a key difference is that what we saw during the early stages of the Industrial Revolution was essentially technology taking the form of automation. What we saw with the Second Industrial Revolution were new industries, new products, and new activities being introduced. And so think about it this way. So if all we had done since 1800 was automation, we would have productive agriculture, we would have cheap textiles, but that would be about it, right? We wouldn't have airplanes, vaccines, antibiotics, computers, et cetera. So most growth and material prosperity comes from doing new and previously inconceivable things.
And I think a key question going forward is whether AI will mainly take the form of automation, and more closely mirror the First Industrial Revolution, or whether we will see it develop new products, new industries that create new jobs and activities, that create more broader shared abundance like the Second Industrial Revolution. That is still an open question, and it depends on our choices that we make.
One of the things I'm really interested in, though, is where that innovation is going to come from. You know, because a lot of this conversation comes around the hyperscalers, about the people creating the frontier models, but traditionally, that's not where the innovation that's going to see that societal transformation, that you are going to talk about, comes from. It comes from people taking that technology, and then reinventing their businesses, and creating new businesses. So where does that innovation come from?
Often comes from the people using the technology, right? And so when you get the powerful productivity tool, whether it's a computer, whether it's the internet, whether it's AI, you can do one or two things. You can use it to do what you're already doing a little bit faster, and you can take on more projects as a result of that, so you can drill more holes, or you can use that tool to dig deeper and develop new types of products and ideas. What you choose to use that tool for will probably depend on the incentives you face.
And so I'm an academic. In academia, the incentive is publish or perish. What do we see since the computer revolution? Well, we do see that, actually, breakthrough innovation is in decline. We see a decline in breakthrough patents as well, because we see that at any given point in time since the computer revolution, scientists and inventors have more projects ongoing at any given point in time. And the more things you have ongoing, the more thinly your attention is spread across those projects, and the less likely you are to push the boundaries.
Do you think, because intelligence is so cheap and available to almost every consumer, that we are getting more kind of AI slop, whether that's, you know, a slop when it comes to academic papers or work, or essays? If we had more scarcity, do you think people would be more careful with how they use intelligence, and use it more innovatively? Do you think that's part of the problem?
No, I think that's certainly true, and I think it's an extension of the computer revolution that way. And we're already actually seeing that in academia. We are seeing more projects, but projects of lower quality being produced.
In my view, I'm a business person, and in my view, a lot of innovation comes from process, about how you innovate process, and how you put AI at the heart of process. And I don't hear that conversation anywhere. I hear the conversation about, "Are we going to Mars? Are we going to have trillions of dollars here or there? And who's going to win the race?" And I don't hear the stuff about how are companies actually going to take this stuff, use this stuff, and create new things from it. And I feel that unless we have that debate, you know, we're just going to be looking for headlines all the time.
Because it's not very sexy, right?
Well, I do think that that is a concern. Although there are many things happening in the world that does not necessarily make good stories. And so even though those things are happening, people are talking less about them. And one thing that we do see in the data is that AI-native firms, for example, they are much leaner than...
Right.
...their non-AI counterparts, which does suggest that there are firms that are setting up, that are really building their companies around this new technology.
All right, thank you very much. After the break, we ask Carl, what happens if AI doesn't deliver productivity growth soon? But first, we want to hear from you. Is the AI industry just full of over-promises? Will we be underwhelmed? Let us know in the comments below as we take a quick break.
All right, Professor Frey, diving straight into it, what happens to society if AI doesn't deliver productivity growth soon?
Right now, a key concern is stagnant growth. I think one underrated reason is that an economy that isn't growing eventually becomes a zero-sum game. The only way I can improve my standard of living in an economy that is not growing is at expense of somebody else's. If people see success in such an economy, they feel that it's come at their expense. And that's a very unhealthy situation to have as a society. It creates very nasty politics. And so in my view, that is the greatest concern, because, frankly, AI is the best hope we have for a revival in productivity.
But I feel as if that's become more of a norm as AI has developed and improved, and these companies have become bigger, and more investment has gone into them, people are much unhappier. They do think that their lack of success is due to other people's success, AI taking their jobs, AI companies trying to replace them. So has AI actually done more harm than good in that context?
Well, AI is clearly a tremendously unpopular technology. Democrats and Republicans cannot agree on many things, but they do agree on that AI is really bad. It's hard to think of any other technology where the developers of that technology are saying, "Well, best case, this will take your job, but worst case, it's going to kill you." In part, because the developers of this technology are actually concerned.
I think there is also some hype around it, maybe because some of these companies do fear that their position in the market may not be as strong as they would like. They've poured a lot of resources and investment in the technology. The best bet they have is maybe that the government takes a stake in the company. Well, if you have the government on your side, maybe that's the best moat you can get going forward.
Though I would suggest that both governments in the US and the UK are so unpopular that having the government on your side may not be-
Yeah, it will just make you even more unpopular.
What you actually want. Yeah.
Can I just pick up on the point that you raised about government in your latest book, you know, "How Progress Ends?" And, you know, you talk in the book about the sort of dichotomy, if you like, between innovation, innovative companies, and bureaucracy. And one of the things that I would kind of say is that governments have no idea what to do with AI at the moment. People talk a lot about AI, but there's very little coherent policy sort of coming out. So what do you think the way forward for government is at the moment to try both to foster innovation, but at the same time, you know, they've got one eye on inequality? What should government do?
I think in terms of government policy, it depends on really where you are. So if you are at the frontier, that your strategy is not necessarily going to be the same one as if you are catching up, right? And so I think for Europe, the best bet right now is on the application layer and doubling down on open-weight models, the way that China has done very successfully.
If you're at the frontier, and you're the United States, I think your biggest risk is that even though you have a lot of companies in the race, they're very highly invested in large language models in particular. And right now, it's not clear that that is going to be the future of AI. It might be small language models, it might be world models, it might be something else. And so what you want to do in such a situation is hedging your bets, and having more firms trying different things and investing in different approaches to AI.
Professor Frey, what do you think we need to do for AI to actually deliver productivity growth? Because we're not seeing it really in the statistics. There's not any robust data, but then the argument on the other side is that it's just too short of a period of time, where like the frontier models that we have now, which are now capable of doing actually decent-quality knowledge work, have really only been about for the past six months. Otherwise, the hallucination rates and the quality of the output was just too dire for them to actually reliably do quite high-stakes knowledge work. So how long do we think we'll have to wait? Or are we not going to see that until something changes in what we're doing?
So I think there is one line of hope perhaps, that if we just wait, you know, the productivity gains will come sooner or later, because with the electrification, it took four decades, with computers, it was a little bit shorter. So if we just hang in there, the productivity gains will come. And, you know, I think there is some truth to that. I'd be very surprised if we don't see AI in the productivity statistics at all a decade or so from now.
But I think then there is the second question, which is, how sustained will that productivity growth actually be, and how significant will it be? Because it matters whether AI will unleash something like the Second Industrial Revolution, where the United States, for example, saw sustained productivity growth for, you know, half a century in 1920...
Yeah.
...to 1970, interrupted, I should say, by the war. Or whether we will see something like the computer revolution in 1990s, which petered out after a decade.
And I think the answer to that really depends to what degree we'll see new firms coming in, building new products and new industries around AI. Right now, it's mainly being used as productivity tool by individuals. Firms are not yet restructuring the way they operate around it. Although we're seeing that incumbents will face more competition now from AI-native firms who've actually built their businesses around AI from the onset. And I think that might be one indicator that we actually might see the productivity gains a little bit sooner this time.
So let me pin you down on this a little bit, because I've been listening very closely, and I think you said that AI is our best hope. And yet, at the same time, you know, I'm not feeling how hopeful you kind of like actually are, you know? So it's a very good argument. You know, you talk about the applications of the technology, you talk about how the combustion engine, you know, 50 years of growth. So how hopeful are you that we will see the benefits of AI, you know, in our societies and in our economies sustainably?
I'm an academic, and I'm Swedish, and so people usually don't bring me on for my, you know, hope and optimism. But I do think that hopefully, we will have learned some lessons from the computer age. One of those lessons is that institutional change is going to be needed for these productivity gains to be realized. Decentralized organizations, for example, did much better during this period in terms of realizing the gains from the computer and the internet. And what we're seeing is, right now, is something similar with AI, which is that firms that are adopting these technologies are becoming leaner.
And hopefully, we'll learn some lessons around some of the negative effects as well, including the rise of social media and the risks. I think, you know, I hope that we don't take those lessons too far, though, and that we end up regulating AI the way we regulate pharmaceuticals, for example, which would create a similar market structure where you have only a few firms that are able to bear the costs of product development, and you have startups being dependent on partnering with those firms in order to bring their products to market.
Humans have, in the end, shown remarkable resilience and remarkable flexibility and ability to learn from history. And so that's the best hope that I have, that we will actually harness the advantages of these technologies. But I certainly would not take it for granted. If I took it for granted, my latest book would not be called "How Progress Ends."
That's pretty hopeful, I think. Professor Frey, we're going to end on our quick fire questions. So mine is, what is something that is accepted as universally true in your field that you actually disagree with?
I think it's seen as almost universally true that the best predictor for the future of work is what machine capabilities are. What's missing in there is that we need to consider more what humans actually want. Do we want the self-service checkout, or do we want the cashier? In that setting, we might be happy with a cashier. But in other settings, you know, it's hard to imagine going to giant, you know, a restaurant that's a giant vending machine, where we do want the waiter involved. And so, you know, understanding what consumers want.
So the question I always ask is going to be a difficult one, I know, because you've already said this. Looking at AI, what makes you the most optimistic?
I'm really quite optimistic about the potential for AI in medical discovery. I think we do have an extraordinary opportunity for curing many diseases, including, hopefully, cancer.
So what did you think of that? A lot to digest there, huh?
Wasn't it a lot to digest? You know, there was just so much in there, because we're going back hundreds of years, aren't we?
Yeah.
We're learning the lessons all the way across. But I find it really, really interesting. But the thing that's going to kind of stick with me is the question that I asked him, really, which is, we've got all these problems to do with AI, and then it's getting bigger...
Yeah.
...and bigger and bigger and bigger. On the other hand, it's our only hope.
Yeah, I mean, you can't take a step back, as in I think we are too far in, although it's only been three years since ChatGPT came out. But yeah, we're too far in, as in you got to think of solutions. And unfortunately, I think AI is a key part of those solutions. Like you can't go back to the Stone Age to try and fix the problems of today. You have to...
No, that's right.
...have modern solutions for modern problems, right?
Right, and the example he gives of, you know, America in the 20th century with the internal combustion engine is a really good one.
Yeah.
And it did drive, you know, economic prosperity kind of upwards. But I would argue, without the Model T Ford, that wouldn't necessarily have happened. And he talks about sort of AI companies first.
Yeah.
You know, AI-first company, should I say. But where are those companies? Where are those companies going to come from?
Yeah. Well, I did want to ask him...
Are they not employing?
What is an AI-first company?
Right.
As in, does it exist? Do we have an example...
Right.
...of one...
Right.
...which is in its infancy?
Yeah.
And I mean, I don't know what that looks like, and it's certainly not even the people that are investing all this money into building the AI.
So where's that going to come from?
We don't know what they look like, but one thing you do know is they won't employ many people.
Yeah, because they're going to be AI-first.
Because they're going to be AI-first. So that's not going to...
Yeah.
...make them very popular, is it?
Yeah. No.
So you've got all these kind of, you know, and the other point that he made about the role of government, you know, again, it didn't seem like a very clear roadmap to me. And I don't really see any governments around the world actually sort of grasping the nettle on that one.
Yeah. Well, a lot to think about.
Intriguing interview. Thank you both so much for that wide-ranging conversation. So with these final thoughts, we would love to hear yours. Is AI a little over-hyped? Are we expecting too much too soon? Weigh in below, and leave your comments. Thanks for being with us today. We're back next week with an all-new episode of "Agents of Tech."