Nicholas Bloom | The Impact of AI on Productivity

Thomas Mertens:

Good morning. It’s great to see all of you here and welcome to those who are joining us online. I’m Thomas Mertens, newly minted senior vice president and associate director of research here at the San Francisco Fed.

I’m delighted to kick off today’s EERN Event. As you know, EERN is the EmergingTech Economic Research Network that was started here in San Francisco. We try to get an understanding of the new technologies and how they’re being used in the workplace, how they affect the labor market, how they affect price pressures, and how they affect economic activity.

We then use those insights and pass them on to inform monetary policy decision making. Now, today’s program continues exactly in that tradition. We’re honored to have Nick Bloom with us today. Nick is the William D. Eberle Professor of Economics at Stanford University. He’s a world renowned expert on uncertainty, management practices, work from home, and the diffusion of technologies. And we’re glad he can join us today.

What’s really amazing about his work is that whatever he touches seems to go viral, not just in academic debates, but also in policy circles and in public discourse. Today, he will talk about the impact of AI on productivity. Now, some of us, including myself, might say, “Well, we know a thing or two, how AI impacts the workplace. We use it every day. We talk to our coworkers. We talk to our friends and family.” But Nick has data from more than 5,000 executives, in four different countries, across three continents. So that’s information on a much grander scale. And I’m really excited to hear about that.

Now, following his presentation, he will be joined by Huiyu Li for a Q&A discussion. Huiyu is my colleague. She’s a research advisor and the co-head of EERN. Now, as a reminder, this event is being livestreamed and it’s recorded and you can watch the recording following the event if you go on the website and you follow the link. And finally, please note, the views expressed here are those of the speakers and not necessarily those of the Federal Reserve Bank of San Francisco, or the Federal Reserve System. And with that, let’s begin. I’ll hand it over to Huiyu. Thank you.

Huiyu Li:

All right. Thank you very much for joining us for another EERN seminar series. It’s my pleasure to welcome Nick from Stanford University to our event. Nick was one of my advisors. I owe a great deal of debt to him for my economic training. So without further ado, let me welcome Nick to the stage.

Nicholas Bloom:

So first of all, I’m very happy to be here. Thank you, Thomas and Huiyu for the introduction. I feel like, I’m like, it’s here once a month, as folks know in the San Francisco Fed, Stanford, and San Francisco Fed. And in fact, Berkeley are very well integrated.

So, I’m going to talk about, as Thomas mentioned, the impact of AI and productivity. This is entirely data. There’s no theory at all in this. You’ll see it’s very data based and it’s primarily a survey. It’s with a lot of co-authors from the Atlanta Fed, from the Bank of England, from the Bundesbank, Kings, McCreary, and Nottingham.

Before I start, I’m actually going to try and embarrass Huiyu. It’s like a very British wedding thing. A British wedding, you’re supposed to try and embarrass the groom. So I’m not quite going to embarrass you, but it was sweet. It was fantastic to be invited out here. And Huiyu, I was on a committee.

This was a photo that actually I found over the weekend. I was looking for something in front of my kids and this thing came up on Google photos. I think the title is good times and it looks like we’re having fun, which is quite impressive given it’s a graduate PhD economics class. I don’t know if you can see, but in the back it says, “Learning about a new technology,” which is pineapples, which is a Conley/Udry paper. So it seemed kind of appropriate. We’re going to talk an even newer technology than pineapples, which is AI.

So the backdrop to this is, in some ways, a paper I have with… I know Chad, I saw Chad Jones spoke here very recently, with Chad Jones, John Van Reenen and Mike Webb that was just pointing out something that I guess folks have known, which is productivity growth in the US, and from work by Chad Syverson and Austan Goolsbee and others. In fact, across the OECD has been declining.

So this is on the right, you can see FRED, labor productivity growth has been gradually declining in the US since the ’50s. It was like 3 to 4% in the 50s. It’s kind of like bouncing around, but there’s a highly significant downward trend. And the question, I guess is, is this time different? As I’ll show you by the end, I’m going to give us a kind of fifty-fifty chance. It’s been declining for 75 years. I think AI is potentially genuinely different. That bar is very high. To point out, we’ve had computers, the internet, cell phones, a bunch of other, cloud, technologies that don’t seem to have reversed this trend.

So why might it be? Well, what are some of the predictions out there? So there’s an enormous range. On the left is Dario Amodei, who is the co-founder of Anthropic. And not surprisingly, he’s very positive in AI. He’s hoping to IPO his firm later this year, or may IPO later this year. He comes out with just amazing numbers. So it could increase productivity or GDP growth because it’s going to be almost all productivity growth to 10 to 20% a year. So this is like numbers we’ve never seen before.

And when you talk to tech bros in the valley, it’s the kind of numbers you hear. They’re like super enthusiastic and they talk about if your AI data’s more than a week old, it’s out of date, et cetera, et cetera.

On the other side is Daron Acemoglu, who’s my co-author and friend who has much more modest predictions. So he had a paper in 2024, which I should put in hindsight now is quite a while back. And he predicted 0.5% growth over 10 years. So that’s 0.05% a year, which is almost nothing.

The CBO is predicting the impact of AI and productivity is about 0.1%. So it’s kind of… I’ve never been in a situation, normally when you talk about impacts of productivity on productivity of new technologies, you’re maybe talking about 50% deviations in forecast, 100%. Here we have like 10,000% differences. So some people are predicting 10, others are one, others are 0.1.

So why did they get involved in this? I mean, working with a bunch of central banks like the Fed that clearly need to understand what’s going on, this is possibly the biggest driver of productivity growth and interest rates and monetary policy going forwards. And so, as I show you, we just did a very data-based, survey-based approach to try and figure it out.

What precisely did we do? We surveyed about, as Thomas mentioned, 6,000 CFOs, primarily, some CEOs, across companies, across four countries. Why did we survey these folks? Well, tech bros are going to tell you it’s like amazing. Everything’s incredible. The technology’s awesome, but the big issue as Chad and others have talked about is bottlenecks. So can this stuff be translated into the ground? And so we wanted to do a representative survey across retail, healthcare, manufacturing, education, et cetera. And my sense is the people that probably know this best are the execs in the companies that need to implement it.

So I’ll go through what the current data is. I’m not going to do an exhaustive survey. I’ll just give you a bit of a kind of background on what we currently know. And particularly, I’ll talk about the firm data and why we got involved in collecting data. I’m not sure this will ever come out as a proper full-on econ research paper. It’s really driven by policy. This is definitely policy first, research second. But it was because banks felt there was a bit of a gap in data. Then I’ll talk about our survey process and our results.

So in terms of current AI data, just want to be clear, there’s two different types. There’s on people, how much do individual people like Thomas, Huiyu, me use, versus firms? And they’re quite different. So I’ll go through them. We have much better people data and much worse firm data.

So what about people data? There are now a number of surveys that survey just individuals, for example, Americans, Americans age 25 to 60, for example, how much they use AI, how much they use it at home, how much they use it at work. And one stylized factor is it’s rising massively, not surprisingly. Second stylized factor is adoption in these surveys typically is like 30, 40%, at this time. So roughly half of the average American, a bit under half of the average American reports using AI.

Interestingly, I have a paper with Christos Genakos, that’s at Gallup, and Gallup has great data on politics. And I’m not going to talk about politics any more than this slide. Right now, AI and politics is kind of a light topic, but if you look at the data from Gallup, they know who’s using it and they know their political affiliation. And it turns out on the left, you can see Democrats are much more like to be using it, much more exposed to it. On the right are much more fearful in the impact of AI and job loss.

So I’m going to raise this. I think the whole AI topic is going to get a lot more complicated in a year or two when it starts colliding with politics. You already see a bit of this. Bernie Sanders was at Stanford two weeks ago, three weeks ago, talking about they should close down data centers because it’s pushing up electricity prices. So it’s already seeping in. It’s going to make it more complicated. I’m not going to talk any more about it, but just highlight that.

The other thing that’s kind of unusual on the individual use side, is this is a technology people started using more at home than at work. So I started work in 1994. When I started work, I was given a share of a computer and it had an internet connection, which is like a really big deal back then. I didn’t have a computer at home. I didn’t have the internet at home. So most technologies like this started at the workplace because they were just too expensive. Typical people didn’t have this stuff at home when it came out. AI is a bit unusual. If you look here, it says that more people are using AI outside their job than inside their job, which is another reason why adoption will be potentially quite fast, because in the workplace, people are very familiar with it.

So, I’m going to stop talking about people and talk about firms potentially because I think that’s where we get the bigger predictor on productivity. So what data’s out there? So firstly, the Census. I love the Census. I work with the Census. I’ve worked with the Census for like 20, 25 years with Lucia Foster. I’ve had many projects with the Census. They have amazing data. The Census to BLS, all of these statistical organizations are fantastic.

The Census has this survey called BTOS, Business Trends and Outlook Survey. Their data on AI, I think is a little odd. Not every single series they have, I’d say is ideal. If you look at it, this shows the AI adoption rate by firms is kind of low and it’s flat. So the latest data is 17%, and it hasn’t risen as much as you might think. Why that might be? One issue from BTOS, and I’ve done cognitive testing with the Census, I’ve been out in the field actually with them and figure… I spoke to people that fill in these forms, is the people filling in Census forms tend to be relatively junior. So they are not the CFO, the CEO, they’re kind of three levels down and they do statistical form filling in. And they’re great on filling in revenue because they can go to Excel spreadsheets and pull it. But on AI use, they maybe don’t know what’s going on and they don’t really have visibility of the business. So the numbers here look low. I don’t know why, but one sense is the folks filling it out aren’t that well-informed.

There’s another data set from a company called Ramp. So Ramp are a fintech. They’re doing very well. They have at least 50,000 customers in the US. They have many more abroad. What they do is they run your payments processing through them. So if you’re a customer of Ramp, you get Ramp credit cards and they… I know the CEO and some of the folks there and they’re growing like crazy. They decided to produce a data set of how many of their customers are using AI.

So how do they measure it? They look for recurring payments to say, Anthropic or OpenAI, in their customer’s transactions data. And if I see once a month you’re paying, your firm is paying $2,000 to OpenAI, I can figure out how you’re using it. So this is their data. They see much higher rates heading up towards about 50%. This is great in a sense, and actually we’re hoping to match this to the Census data and do research with the data, but there’s a couple of problems. On the one side, it’s a very narrow measure of AI. So it just looks at recurring payments for well-known AI providers, which is pretty narrow.

On the other hand, you can guess who their customers are. They’re like tech forward, very sophisticated, young firms. So one boss pushes it down, one boss pushes it up. It’s hard to know net-net what the impact it is. But overall, these are the kind of two most commonly mentioned data sets.

So why do we get into it? We wanted to look at adoption, in particular look at the impact on employment and productivity. So we have these existing panels. On the left, there’s something called the decision maker panel in the UK. It was set up in August 2016. So I was one of the co-founders who set it up after the Brexit vote. We were trying to figure out what the impact will be, and it’s been running ever since, and it’s heavily used by the MPC in the UK. On the right is the Atlanta Fed Survey of Business Uncertainty, which is a similar structure survey, it’s been going also for about 10 years.

How do these panels work? So if you want to recruit execs, you have to phone them up. It’s just like impossible. If you email them, I can see Dan here. Dan Wilson’s like a CFO of a company and I email him, he’s just going to ignore it. He’s not going to pay any attention to this. And so he’s not going to fill it out. I could of course pay him and offer him 50 or $100, which is traditionally how people would recruit execs. The problem is now that kind of paid exec online survey process is infested with bots. It’s very hard to know whether it’s him or if it’s someone masquerading as him or AI, et cetera.

So we instead had a two-step process. One is we get the registry of companies and we phone them up to make sure. So I phone up and I know the CFO is Dan Wilson. I get him on the phone. It takes three, four, five minutes, and he agrees to be in our monthly panel. And we ring up and in the UK, for example, it’s endorsed by the Bank of England. Mostly, we’re getting a kind of a 30% response rate. It’s kind of like if me as an academic was asked to fill in a survey by the American Economic Association, I don’t know whether I would or what, but I’d feel warmer towards them than anyone else. So we’re getting kind of a 30% response rate.

And then we push them into this frame and once a month he’ll get a short email, he’ll respond to the email with some survey questions and he may do it for three months, get bored, drop out, we phone him back again, say, “We’ve missed you. Where have you been?” This is kind of how this process works.

So we target CFOs. If we can’t get the CFO, we go for the CEO, and if you can’t get the CEO, you go to senior execs. To be clear, we target the population of all firms at 10 plus employees. So this is not like Walmart. They may be, I won’t mention who is who or not in our survey, they would of course be in it, but you should not think in your mind the typical firm in this is Walmart or General Electric. The typical firm has 250 employees. So probably the CFO is, I don’t know, at half a million dollars a year, they’re senior, but they’re not ridiculously senior. They’re never going to answer the phone and they’re never going to take part in the survey.

So before I show you the results, I’ll show you a couple of checks. So I always say it’s like food. If you’re going to eat food, you want to A, check the ingredients and B, check something about the place that’s serving it to you. They have some reputation. So I’ve showed you kind of the process a bit. What about the historic performance of these surveys? So here for the US and the UK, in the top row is the US, the bottom row is of the UK, just showing you for the last 10 years, our panels match of sales growth versus GDP on the left column and employment growth versus private sector, payroll from ADP in the top and a UK survey from the bottom.

So this isn’t a perfect match, but it’s pretty good. And why do I put this up? It’s because this survey’s representative. It’s like large and small, across regions, across industries. I mean, the sample’s a couple of thousand firms at this point is large enough to represent the economy. So overall, it looks reasonably good. The MPC certainly UK uses it a lot. The FOMC, I don’t know the exact details, but briefings are used based on this data.

The other perhaps tougher question is these respondees are used to making forecasts. Why is this? Well, it’s one of the upsides about CFOs, they’re numerous. And so, they can predict stuff they’re comfortable with to a reasonable degree of predicting. We can actually get prediction ranges. They can put probabilities on different percentages. Here’s two examples of predictions we get them to make every three months, which is sales growth, looking one year ahead for their company and employment growth looking one year ahead for their company.

And because they’re in a panel, because we can pull accounting data, we can wait a year and then see how they do. And this is prediction versus realization. It is pretty good. I wouldn’t say it’s perfect. Of course, the pandemic happens in the UK, there’s Brexit and there’s various shocks and things that throw them off. But overall, they have pretty good forecasting accuracy. It’s not bad. So this is a sample that represents broadly the economy and is reasonable at forecasting, at least historically.

So what are the results? So the first question is, in some senses, the easiest to answer is a question about yourself. On average, how frequently do you personally use artificial intelligence technologies in a typical working week? Before I show you the numbers, you should look at this and think for yourself. I know what I would sit on this, it’s a relatively simple question: Not at all; up to one hour; one to five; more than five.

So we survey typically CFOs, you should think of these people are probably in the early 50s, late 40s, they’re answering this question. What do we find? So most use it a few hours a week, but start on the left, perhaps surprisingly one quarter of CFOs in our survey. And I should be clear on the timing. So this survey ran November, December 2025, January 2026. It was evenly run over three months, so it’s pretty recent. It’s between kind of two to four months old. One-quarter just don’t use it at all. They just do not actively use it. Now of course, they passively use it. You use transcription and Zoom, for example, you’re passively using it, but this is kind of active, deliberate use of AI.

Of the other 75%, the most common is up to an hour. There’s then one to five hours, which I think would be where I am. And then there’s kind of five plus, which is to be quite honest where a lot of Stanford undergraduates are, unless they’re told they’re not allowed to use AI in class, in which case it’s zero. But I think most of the undergrads are using AI a lot.

Is this a high or low number? It kind of depends on your context. So when I talk to people in Silicon Valley, friends are in tech, friends are in VCs, they’re kind of stunned at how low it is. It’s like, “Who on earth doesn’t use AI? I mean, who is out there not using AI?” On the other hand, if you compare it to other technologies that really went mainstream three years ago, so ChatGPT was released in November 2022. This is an amazingly fast adoption rate. I’ll show you a bit of data on that in a minute. So it’s very quick, but it’s not instant. So not everyone’s using it, but 75% of people are using it. On the right, the average AI use of these folks is one and a half hours a week, which is material. Again, remember, these are kind of senior people running organizations, so perhaps it’s not that surprising they’re not using it more.

What about their firms? A separate question, how do their firms use AI? Again, similar numbers, about three-quarters of their companies explicitly, deliberately use AI. Is that high or low? Well, I think it’s very high compared to classic technologies where in a sense, I think perhaps this time is different. This is much faster adoption rates than things, for example, cloud. So sitting opposite someone at a Stripe event a couple of weeks ago and the guy opposite me just sold his company to Stripe and was saying cloud has about a 50% adoption rate of firms 10 years in. We’re seeing AI is 75% three years in.

The other thing that’s notable is the US is the highest. So the US is in light blue. The UK is kind of a bit lower. And then, you see Germany and Australia are further behind. I’m not that surprised. We’re sitting in the US, we’re in Silicon Valley out here, so the Fed is kind of… the San Francisco Fed is perfectly positioned to see this, be at the frontier of it. There is an adoption curve, but these gaps aren’t that big. So Australia’s lower, Germany’s bigger, but it’s not massively. It looks like the US, kind of a year behind. So this thing is spreading out pretty fast internationally.

So it’s hard to measure adoption rates on a like to like basis. This is data. Actually, I put it on LinkedIn yesterday because Akash Kalyani, my co-author, provided it, shared this for me, and it was like an amazing chart. So how do you measure the adoption rate of different techs? I wouldn’t say this is by any means a perfect measure. What is this? This is the share of job postings that mention this technology. You can think of good and bad reasons why that may not be ideal. I don’t know. If you lose your job to the technology, there’s not even a posting there. But for what it is, this is postings that mention the technology. And you can see here, AI has a much faster uptake than pretty much anything else. And again, this kind of maps with other data I’ve seen. The adoption rates have been very fast.

So what about impact? Finally, and perhaps the most important questions, what’s the effect on employment? What’s the effect on productivity? In order to get numbers from CFOs, even though they’re numerous, we can’t talk about total factor productivity. That’s not a concept they’re familiar with and who knows what we’ll get. So we asked about sales per employee. That’s kind of what… I used to work at McKinsey years ago and that’s kind of their business definition of productivity. So that’s on the left. And then, we asked about impacts and employment on the right.

So firstly, we asked about the last three years. So this is looking back to really the beginning of AI when it really came out, at least ChatGPT was launched. And you can see on the left, the basic answer is not much. So 90% of firms are just saying it has not had much of an impact on productivity. And you can see there’s about 7, 8% are saying a small effect, and then there’s a couple of percent that are saying large effects. On the right, the average effect on productivity is 0.3% over three years. That is 0.1% a year. That actually is almost exactly the CBO’s numbers, their own forecasts.

Should you be surprised? I don’t know. I don’t think so. So as you know, I’m at Stanford. We have a ton of AI. If you were to measure us, we use AI all the time. Has it affected revenue? Not really. I mean, tuition is unaffected by it, grants raised are unaffected. The sports revenue, everything is pretty much unaffected by it. Has it affected employment? Not really. So for us, we’re kind of pretty high adoption rate. I don’t see an enormous effect. There’s not many firms outside of the kind of techosphere that talk about it. I have friends who have software startups, SaaS startups, and they talk about impact, but they’re like 1% or 2% of the economy. They’re probably that top, far right thing that they’re kind of pretty much outliers. So over the last three years, we just have not seen on average a big effect on productivity. And if you look at the productivity numbers, they’re fine, but they’re not particularly amazing. So I think this matches aggregate data.

What about employment? Here is a very tight zero. So again, execs are just saying… looking back, to be very clear, this is looking back, not a prediction. I’ll show you predictions in the next slide. We just don’t see much of an effect on employment. You can see there’s a few percent that are saying small negatives, small positives, but if you look on the right, the net overall figure is 0% with a very tight standard error.

So over the last three years, if I look back, I just don’t see a big impact in our data of AI. Again, I’m aware that there are tech bros and SaaS companies, et cetera, that see enormous effects, but they are just not representative. The typical firm in here is some retail or a dentist or school or something that’s kind of in the broader economy and they’re just not reporting enormous effects.

When you look forward, things are going to look different. So here’s where the numbers get, in my mind, really big. So this is the forecasted impact on productivity in the next three years looking ahead. These numbers, to be clear, I think are absolutely enormous. So they’re nowhere like Dario Amodei numbers, but I’m not trying to sell a big AI startup. They are still very, very large. So if you look at the blue number, which is the US, the largest group of firms say up to 5%. On the right, if you take the average, you take the midpoints of these bins, American companies are reporting 2.25% increase in productivity over the next three years, which is three-quarters of percent a year. This is a really large number.

So again, productivity growth now is slowing to about one to one and a half percent a year. So adding three-quarters of percent will be roughly adding 50 to 100% onto baseline. This is an enormous number. And this is the sense in which I’m not convinced this time is different, but for the first time in my career, I’ve been an economist since the mid nineties, it feels like it could be. I know my idea’s getting harder to find as a historic paper. That era may be over. I give it kind of 50/50. So this is productivity. These are, again, really large numbers if you believe them. And to the extent to which we’re going to get any numbers, I think this is kind of our best bet just to ask execs and firms across the economy.

What about employment? Here, I actually think this is not much. I’ll just explain what they say and then explain why I don’t think this is a big deal. So firstly, what they do is tell us in raw numbers. If you look on the left, the distribution of responses, these firms are skewed… the most common response is nothing. The second most common response is small negative effect on employment in their company. Some are saying large negative. Some are saying small positive, or very few saying large positive. Over on the right, you can see for the US, they’re predicting in their own companies, 1.2% reduction employment over the next three years. The US labor market is about 180 million. This is roughly two million employees, so that’s a kind of ballpark figure.

So why do I think that’s not an enormous deal? There are two reasons. One is this is in existing companies. So just to be clear, there almost certainly will be new jobs created and new companies to offset this. Classically existing companies actually have slightly net job destruction, new companies net outward job creation. So A, this is only part of the picture. And I think if you look at the entire economy, it will almost certainly be less negative. B, this just is nothing in comparison to the scale of the US labor markets. If you look at JOLTS, there are five million jobs in the US that are destroyed every month. There are five million separations as far as quits a month. So over three years, you can imagine comparing five million a month for three years to two million over three, it’s just nothing. It’s like 1% of the churn. It’s just very, very small. So unemployment, at least in the next three years, I just do not see a big aggregate effect. Some additional churn, but nothing that’s particularly striking.

Finally, we also asked employees the same question. So do employees align with managers? We ask them the same question about AI use and impact on their firm, importantly, in their firms rather than in general. So what do employees say? So on the left, so this is 3,000 US employees. On the left, they use it about the same as managers and average 1.8 hours a week, so very similar. On the right, there’s a clear difference, first on employment. They’re net positive on employment rather negative. So the average American employee actually thinks it’s going to boost employment. And secondly, their productivity numbers are way smaller. So this is 0.9 versus 2.25.

So why the difference? One story is, look, employees know what’s going on. They’re doing the jobs, they’re on the ground, they’re kind of at the front edge, the C-suite is kind of being pushed by the board and they’re reading the media and they have no idea what’s actually happening on the ground. And so, employees are the people that are kind of realistic.

The other view, which I think I ascribed to you slightly more, is actually the C-suite potentially has a better view. A friend of mine was saying in his company, they’re adopting AI. When you talk to employees, they say, “This is going to be great. It’s going to make my marketing job a bit easier and do this and this and this.” And he said to the board, also more like the C-suite is thinking like, “Do we need a marketing function? Do we need this person? Do we need this role?” And so, in different views, the C-suite execs have the overall big picture and employees are kind of focused on smaller sub parts of this.

So to end, I think firms’ forecast, looking back over the last three years, we haven’t seen much impact. Looking forwards, the prediction for productivity, they’re nowhere like Amodei or Silicon Valley numbers, but they’re still massive. They’re really big numbers. Three-quarters of percent, if you believed it, is an enormous number. The impact on labor markets, again, in the next three years, I don’t see this large. I want to be really clear. This is one of these things where people overestimate the short run and underestimate the long run. I’m only really talking about the short run. Ten years out, you want to think about technology. I’m not talking about that at all. I’m just talking the next three year window.

Certainly for me, it’s updated my views. I’ve been, I’m not sure, a techno pessimist. My predictions are kind of aligned with data to now, but looking ahead, I’ve become more optimistic. And finally, we’re going to run these things every six months. As you know, in AI world, data depreciates at an incredible rate of speed. So within six months, this stuff is like historic. I was at an event three weeks ago and explaining this to some exec, in fact, the CEO of Ramp and he said, “How have you collected it? Is it more than five days old?” And I was like, “Yes, it’s more than five days old.” It’s like, “Oh, it’s out of date by now.” It’s like, “Oh my God, how do we do it?” We can’t keep up that much data. So we’ll rerun this again in June 2026. Okay, thank you.

Huiyu Li:

All right. Thank you, Nick. That was a fantastic presentation. I enjoyed it a lot. So you mentioned that your idea is getting harder to find the paper. Actually, I just Googled it. It had over 2,000 citations in Google Scholar. So it’s one of the most cited paper for understanding the productivity growth slowdown in the US. So I was wondering if you could give us some examples from that paper to help the audience understand why you had that view that productivity growth was not going to pick up.

Nicholas Bloom:

Sure. So that paper’s again, very empirical. I should say, in deference to Chad Jones, that was like Chad… I’m alphabetically first, but I think it was very unfair because Chad Jones was honestly the key driver of that idea. So this is very much his kind of research agenda, in particular. It’s kind of fascinating looking back at history, going back several hundred years looking at productivity growth because there are three phases. So if you go back far enough, you have to use British data because it’s hard to get American data. But if you look, productivity growth from Roman times to 1750 was incredibly low. So Madison has data, so it’s like 0.01. It’s just almost nothing.

And then what you see is the industrial evolution happens around 1750. And this is on British data, productivity growth is rising, rising, rising, rising for about 200 years. It’s accelerating. And then it hits 1950, that’s kind of the pinnacle. And then in either British or American data, what you see is it starts declining back down to now. So we’ve really only got two turning points in history, you got 1750 and 1950. That paper only looks at the 1950 bit onwards, but it says productivity growth is slowing down.

Why is that? The simplest example is Moore’s Law. So Moore’s Law, there’s different ways of doing it, but one is like the speed of chips is doubling roughly every two years. And if you put speed against time on a log scale, it’s this nice straight line. But the number of scientists involved in all the chip manufacturers is going up by about doubling every three years. And so the amount of effort to generate that is getting harder and harder and harder. And if you look at the whole US economy, you see productivity growth is slowing and R&D is going up. So that paper is saying it’s just getting harder.

And it’s classically a battle between the kind of apple tree effect. So like the low hanging fruit gets picked and it gets harder and harder to come up with ideas versus the standing on the shoulder of giants effect, which is other people make inventions and you build on it. And I think standing on the shoulder of giants was the dominant thing from 1750 to 1950, which is getting better. And then kind of the apple tree effect took over.

Erik Brynjilfsson’s my colleague at Stanford. He’s very tech optimistic. I have not been of that view until now. At this point, AI just seems so different. So I’m not going to say that paper’s wrong because it’s looking back and all the data’s correct. It’s less clear if it’s a perfect forecast going forward.

Huiyu Li:

Yeah. I mean, I think we talked about this a little bit offline that you really believed in the idea getting harder to find paper, but eventually your needle started to shift and that’s why you’re in the AI research field as well. Can you just take us through what are some of the things that shifted the needle for you?

Nicholas Bloom:

You know, it’s weird. All of my colleagues started going into doing stuff in AI and I kind of thought I won’t… As an academic, you know this, you don’t kind of want to be in the same area where everyone else is. And I thought, well, look, everyone’s rushing into AI. I don’t have any comparative advantages. I’m not going to do any work on it.

I ended up, this is really a data thing rather than a formal economic paper, is because of policy. So particularly the Bank of England, I mean, I know it’s a big issue in the Fed System with comments from Warsh and other views going on. It was like, what’s the effect of AI? It was literally an empirical thing. And it’s interesting because economists tend to forecast off history and models, which is great and it works very well. But when something’s radically different, it’s less clear how useful that is. And it feels like AI is potentially radically different. I mean, for everyone here, it’s definitely affected that class. In fact, so the class that you-

Huiyu Li:

The pineapple class.

Nicholas Bloom:

Yeah, that class. Just to give you an example, it’s really affected teaching for me for undergrads and grads. So two anecdotes. At Stanford for the undergrads, my friend, she’s in the maths department and has said the homework scores are going really up. They’ve gone up massively over the last three years, but exam scores have plummeted. Why? Because people are using AI to do their homework and then they don’t know it. And so they flunk out in the exam. And I had a different thing in that class that I teach, I lecture for half the class. It’s like a two-hour class. So I lecture for roughly an hour and then everyone’s read one paper and I randomize and someone will stand up and discuss it. And if Huiyu’s name comes up, she’d stand up and discuss it.

This year, everyone seemed to have the same clean looking discussions. And then I’d ask them questions on like, “On bullet point two, what do you mean?” And they’d be like, just reread it. And I’d say, “Well, what exactly?” And it was clear. Suddenly the light bulb came on to me that the students rather than reading the paper were feeding them into AI. And so it is just having an enormous effect.

I mean, Greg Thwaites, one of my… I have some other stuff we’re looking on the Phillips curve and we’re trying to do a GE (general equilibrium) model. And I know for people who’ve coded up Krusell-Smith, but it turns out he did an entire Krusell-Smith coding in like a week using Claude. So it’s clearly… I mean, everyone here must have the same experience. It’s really just affecting stuff on the ground. So as Thomas said, I just wanted to collect data. This isn’t a heavy duty academic paper, it’s just a data exercise.

Huiyu Li:

I mean, I think fact finding is very important at this stage because such a new technology that… Actually, one of the reasons our bank started the EmergingTech Economic Research Network is precisely that it’s a new technology. We don’t really know what’s happening and research tend to be a little bit lagging behind the real time data need for policymakers. So I think I really appreciate the data finding stage.

Nicholas Bloom:

You guys are in a great position here in terms of… I mean, it’s just so led by… It’s coming up Silicon Valley.

Huiyu Li:

Yes. Yes.

Nicholas Bloom:

And it’s a great idea to do this.

Huiyu Li:

Thank you. So there are many different kinds of survey data out there right now. As you mentioned, the Census results, the worker surveys, and just even measuring what you might think is simple like adoption rate, you have a big range of answers. I think you went over this a little bit in the talk already, but I was wondering if you could help us understand, for example, what’s the difference between Census adoption rate, which is lower than your numbers. I’m wondering whether it’s also related to the definition of adoption, like the way they ask the question versus you asking the question in your survey.

Nicholas Bloom:

Yeah, absolutely. So I presented this data to a bunch of tech folks. I was at an event. I’ve done two or three events with like, it’s non-academics, so the people like VCs and founders, et cetera. And they are often like everyone’s using AI all the time. If you ever use transcription or you go to an app, it’s powered by AI. So what do we even mean by… So you saw a question, I think it’s active deliberate use, but it’s a bit vague. So one thing, even just collecting data, it’s not that easy to be quite honest on what we mean by it. And then the other thing is who’s using it? Because if you look at any big firm. The thing about the Census is a bit weird is you have firms with 250 employees apparently have 18% AI use and in 250 employee firms, almost surely some people are using it somewhere. It’s just you’re not aware of it.

Huiyu Li:

Yes.

Nicholas Bloom:

So, in some odd ways, I prefer the CFO questions about how many hours a week are you using it? Because it’s like the one thing people can answer correctly is how much they’re using it themselves. It’s quite hard, I think, for people to get a sense of how much it’s used around the company. Our measure, I think, picks up basically CFOs that are deliberately paying for it. It’s in a sense similar to Ramp, but Ramp just has to look for recurring payments, which is slightly harder to pick up.

Huiyu Li:

I see.

Nicholas Bloom:

So, the adoption rates are clearly high. Why the Census numbers are low? I don’t know. I’ve worked with the Census many years and I did management surveys and they look great. We do cognitive testing. We’d go out and visit locations and test them out. I mean, I remember, this wasn’t the Census, actually the Atlanta Fed, but it was a great example of why it’s really hard. You have to test surveys. So we visited this location and they were like a hipster food company. They spent ages telling us how they were local food and it was all locally made, et cetera, and it was only one location.

And one of the questions were, “Are you a multinational?” And because we were testing them and they said, “Yes, we’re a multinational.” I remember I was testing it with this lady and I looked at her and was like, “Well, clearly they’re not a multinational.” I didn’t say anything because she was leading the testing and she said, “Oh, what do you mean you’re a multinational?” And they said, “Well, in our firm, we have employees that come from El Salvador, from Puerto Rico, from Mexico.” Okay. So at that point we changed the question to, “Do you have production facilities abroad?” You can’t predict quite how these things would go and how people misunderstand it. So that was another thing I’ve learned over the years is just pilot these things.

Huiyu Li:

Yes. I mean, from taking your class in Stanford, I understood that actually designing surveys is very difficult, making sure you have answers that are easy to interpret from an economics research perspective. So in terms of designing this survey, do you take advantage of all the experience of survey that you’ve done in the past for management, for work from home? Is it like similar sampling framework?

Nicholas Bloom:

So there’s two different populations, and things have changed a lot recently with AI actually on running surveys. So there’s execs and execs or doctors or populations that are kind of hard to reach, that paying these people has really fallen apart. So it used to be, I don’t know how many people have run surveys. Ten years ago, if you’re a pharmaceutical firm, you could pay doctors like $100 to take part in this survey and you’d screen them out by asking some medical questions. If you didn’t get it right, you just say like, “You’re not a doctor.” And so you could screen it out.

Now these things have totally fallen apart because they’re absolutely infested with AI bots and AI can answer it. And I was reading some stuff online, like people think, well, you’ll have a question that the font is in white and the screen is in white, but AI perfectly gets that. AI’s really good at getting all of these CAPTCHA trick things. It’s really hard to get rid of AI.

So for execs, I think is why institutions like the Fed are in a great position actually because it’s hard to survey these people otherwise. So you have a group of execs that would speak to you and they won’t, it’s very hard to get them otherwise actually. So this is why I’ve been working with the central banks, you have an enormous advantage. For individuals like the general public, you can still do online surveys where you’re paying people a dollar a go. Why is that not overtaken by bots? Because the companies figure out if you have multiple… You have to have different IP addresses and they have screeners and it’s just not worth it. So the money’s not enough. And often they’re paying in like air miles and in-game credits and stuff like that. So if anyone uses Cint or these surveys, they’re great. So if you want to survey random Americans age 20 to 65, you can still do it. If you want to survey execs, you guys in the central banks are probably the best way to do it. I don’t see an easy way to do it anymore actually.

Huiyu Li:

Yeah. Actually, I appreciate the perspective of executives. Our bank president, Mary Daly, gave a speech recently in February where she emphasized that her view is that, according to the speech, this is not my view, this is her view according to the speech, that for AI to be really transformative, they have to be organizational changes, not just employees experimenting with AI like marginal things, but you really need to transform the organization. So in your survey when you found these executives forecasting gains in productivity in the next three years, do you have some idea of like, in what way are they planning to achieve those gains?

Nicholas Bloom:

Yeah, the organizational change is a critical thing. So just to be really clear, the tech behind AI now is just amazing. I mean, I was in an event three weeks ago with the CFO of Anthropic. There were a bunch of AI companies they were talking about. It’s incredible what’s happening. So I’ll give you kind of two anecdotes on org change. So the classic anecdote economists love is the introduction of electric motors. So if you go into New York, I always wondered why loft style apartments, like why would anyone build a factory with like a really tall factory? It’s really expensive to build like a six-story factory, which are now loft-style apartments. Turns out, I didn’t realize with water and steam power, it was all mechanical. And you’d have one rotating pole with belts off it. So if you go to a factory that’s built powered by a steam engine, you have one steam engine with a big pole that rotated around and loads of belts.

So if you look at factories like 150 years ago, they’re tall and thin. So you have lots of belts coming off it. When the electric motor comes in, it was completely different. It was lots of little motors spread out. So that’s why modern factories are one story kind of slab type buildings, which are totally ugly. And it’s what now modern factories look like.

Now, I raised this because there’s this J-curve effect that when the electric motor came in, it took 20, 30 years to have an effect. So you had to reshape the whole buildings. And so initially they just replaced the steam engine with an electric motor, still with a central rotation, that didn’t do much. So that’s the classic sense in which the real obstacles are adoption.

I can give you another anecdote just for us at Stanford, which is I’m involved in admissions, PhD admissions. In fact, I’m chairing admissions this year and it’s an insane process with like 700 applications, thousands of letters, transcripts in many languages, and it would be really natural to use AI. In fact, AI would be perfectly made for this because you have people… Thomas may write repeated letters and you could compare them over the years or Dan or anyone else. But we’re not allowed to because Stanford doesn’t want to use AI because it doesn’t want us in the New York Times with some things saying Stanford University is using AI for admissions. It’s just like, we’re just not going to go there. So as a result, this is more of the obstacle.

So my sense is the next three years, exactly as you say, is driven by obstacles. Like the Fed System. I think because of security, you can’t… Can you use Claude on your computers here? I don’t think, can you? Or maybe you can. I know some organizations have issues with bringing data in and out. There’s all these obstacles that 10 years will be worked out, but this I think is what’s going to drive the next two or three years.

Huiyu Li:

So in your previous research about IT, I think you have a paper about America, US doing IT better.

Nicholas Bloom:

Yes.

Huiyu Li:

So since your survey has this cross country dimension, do you see any differences between how a different country might be able to achieve the productivity gains and what are some of the important factors for making a country successful?

Nicholas Bloom:

Yeah. So we see Germany. I mean, don’t know if any Germans here, but Europe has many more regulations on this, so it’s going to be slower.

Huiyu Li:

Thomas is German.

Nicholas Bloom:

Here we go. So our friends in the Bundesbank are like ugh. But because AI is like everything on warp speed, rather than being five years behind, they’re one year behind because everything’s happening so fast. So in the June version, we’re trying to… Well, hopefully we’re going to get the Bank of Mexico, we’re talking to the Bank of Korea, et cetera, to try and spread it out. But the US is clearly frontier and you do see the UK’s a bit behind and you see some adoption. We’ve only got four countries, so all of our four countries are pretty wealthy and developed. But my guess is if you look at the bank… One of my co-authors is Jose Barrera. Did you overlap with Jose?

Huiyu Li:

Ah, Yes. Yes.

Nicholas Bloom:

So Jose is at ITAM in Mexico City. He works a lot with the Bank of Mexico, so hoping to get them involved.

Huiyu Li:

I see. On the employment results too, it seems like US for the next three years, there was a forecast of slightly higher employment decline for the US than a couple of other countries. Is there a reason why US might be more pessimistic?

Nicholas Bloom:

I think it’s just a bigger impact. My personal read of the data is it has not affected employment much. I know there’s a literature claiming that it’s reduced hiring of new graduates. Kind of two thoughts on that. One is that’s actually empirically really hard to tell what’s going on. So the classic way this literature operates is it looks at the hiring of new graduates pre and post November 2022. And the notice is that hiring of new graduates, particularly in IT-exposed sectors, has dropped a lot post November 2022 and attribute that to ChatGPT or AI. But the problem is something else happened, as we all know, the interest rate tightening cycle in tech, over-hired and went into a hiring slowdown and so it’s very hard to tell what’s going on. At least our survey suggests it wasn’t AI so I think AI hasn’t had an impact much today. But I think looking forwards, it’s going to start to have a much more impact.

Huiyu Li:

Yes, that would be consistent. We had the chief economist of Google here for an event.

Nicholas Bloom:

Oh, yeah, yeah.

Huiyu Li:

That was his view as well that it’s probably not AI right now.

Nicholas Bloom:

It’s great. He came to talk in class actually. Fabien, yeah.

Huiyu Li:

Yes, Fabien, yes.

Nicholas Bloom:

I had a Zoom call with him and before he told me he was working on it, I was like, “This is the result, I’m thinking it’s interest rate changes.” And he’s like, “Aha, I have a paper on it.”

Huiyu Li:

It’s very much aha.

Nicholas Bloom:

The other thing that’s really interesting talking to tech firms is there’s two views on the impact of AI and hiring of young folks. I talk to a lot of undergrads and not surprising they’re really nervous about the effect. One view is the classic view that, oh, it’s really bad for undergrads because like my friend, she’s a law partner. And she said, “In our law firm, we’re just hiring less new joiners because I’d have them do basic case work, prepare slides, help on briefs. And I can do that in AI.” She’s like explicitly, “AI is just doing the stuff that these new associates would do.” So that sounds like bad news, and that’s the main view.

But there is a second view, which is young people are AI native, and they’re much faster at getting up to speed in it. For example, IBM, Cloudflare, have both explicitly said they’re going to ramp up actually hiring of new entries and interns because they’re much more familiar with it. And it’s not clear to me net-net which would dominate.

Huiyu Li:

I see. This is actually a good segue to take some questions from the audience. We have both pre-submitted questions and questions from the audience. First one, did you get any insights on how leaders were estimating the productivity gains of AI? This is kind of related to something I’m thinking about too. When we talk to business leaders, sometimes I think there’s a different idea of what is productivity gain. They think more in profitability.

Nicholas Bloom:

Yes.

Huiyu Li:

Rather than productivity. So how do you think about this when you conduct these surveys?

Nicholas Bloom:

So again, if I come back to my example with Dan, he’s the CFO, he’s probably on a million dollar… I don’t know why I keep picking on you, Dan. I’m sorry. Dan and I know each other going back 20 years. We have a lot of overlapping work. So, he’s probably paid a million dollars a year, he’s pretty busy. We phone him up, we’re very happy he’s in the survey. We can’t ask more than five minutes, to be honest, these monthly meetings. We don’t ask any how or why, we just ask for numbers. I think things like the Beige Book are actually great, where you can sit down and talk to someone for 20, 30 minutes. I think the Beige Book’s really useful for that.

So it is also true when I talk to execs, having presented this, they often talk about increase in revenues and reduction of cost and they all get frustrated that everyone’s focused on reduction of cost. And they often say, “Actually, what it’s doing is making my product better and faster.”

There were two calls yesterday with two big tech firms, both of which you’d know, I won’t mention them because it’s a bit sensitive, but one of them is ambivalent about AI. It’s massively affected their business actually and they’re completely re-swiveling around.

And the other seems incredibly positive because the head of sales, she was saying, “I have this 2,000, 3,000 people in my sales group and they’re focused on small and medium businesses. And the businesses traditionally were just too small to bother to put a person on. But now at the tail end, we’re starting to use AI and also AI plus human.” And so she can see her funnel getting much bigger and she says sales is going up. So I think there’s the costs go down and revenue go up, but everyone’s focused on the cost go down. But from talking to business, they’re much more typically focused on the revenue going up.

Huiyu Li:

I see, okay. This is a question that came through the pre-submitted questions. We actually get a lot of pre-submitted questions before the event, so this is about the optimism about AI from the employee side that you found in your survey. I think there are also other surveys like Gallup that found that many respondents seem to be quite pessimistic or have a sense of uncertainty or fear around AI. Why do you think your results might be different from some of these other surveys?

Nicholas Bloom:

It depends critically how you frame the question. And a totally different thing, I remember a friend of mine who was quite senior in the Conservative administration when the independence vote was coming. I don’t know if people remember, but there was a vote whether Scotland would get independence from the rest of the United Kingdom. I think it was 2012, 2013. And my wife’s a Scot, so it’s very relevant to us as to whether we’d be carrying two passports rather than one as a family unit. But there was a huge debate. And in the end, the Treasury said about whether 16 and 17 year olds can vote. So classically, only 18 plus can vote. And the independence group wanted 16 and 17 year olds to vote and they negotiated, yes, 16, 17 years can vote as long as we get to design the question. And the Treasury’s view is designing the question matters much more than who votes. And in the end they won in the sense separation didn’t happen. And so it’s a lot like designing the question is pretty critical.

I’ll tell you what we do, and then what we do is we basically say, we try and design something that’s neutral, which is, for example, what do you think the impact of AI will be on employment? And actually, you can’t give negative numbers. You just have to use a two point question. So you say, will it be positive, no impact, or negative? And if they say positive or negative, so they say positive, we say, how positive? Will it increase it by one to five percent, six to 10, 11 to 20, et cetera. So you can classically get 15 bins or something across it. That is neutral, and so those are where our numbers come from.

You can set a question up another way, which is more like the Gallup one. If you looked at the Gallup data, which was, “Are you worried that AI may replace your job?” And that is a pretty leading question. And so there you get loads of responses of, yes, I’m terrified. So they both can be true. It can both be true that people on net think it’s going to increase employment and they’re worried it’ll replace their job. But one of these things from running surveys, the framing matters enormously. So we try and frame it as neutrally as possible. So if you look at the questions, it’s like positive, zero, negative, and then how much.

Huiyu Li:

I guess there’s also selection to who answers the survey. I could see people who might stop and talk to Gallup about their views are the ones that are particularly worried about AI.

Nicholas Bloom:

You’re totally right. So another big… Stefanie Stantcheva has a bunch of great stuff on this. You don’t want to tell people upfront what the survey is about. Again, if you say it’s a survey on Brexit, you can imagine who’s going to take part it in, the people with extreme views… You get the Yahoo problem. If you look at restaurant ratings in Yahoo, they’re either fives or ones because only people that hate or love it actually bother to go and there’s almost no three stars. With surveys, ideally you have a very vague introduction. So for our panel, we just call it the decision maker panel or the Survey of Business Uncertainty (SBU) and it just rotates through.

Huiyu Li:

Okay. Yes. And then this survey, it sounds like you will be continuing doing this survey. What are the plans like in terms of frequency as well as like what kind of questions will you be following up?

Nicholas Bloom:

So we plan to do every six months because otherwise people get a bit bored of asking AI questions. In the end, actually in the US, so in the SBU, we put in a question on expenditure. So it’s something else that I was amazed about. I sat next to at a dinner the other day, an Economist journalist, Callum Williams. I don’t know if people know him. He lives up in the city and his wife’s a VC. And he was saying there’s no numbers on AI. Anthropic may list later in the year, but no one knows Anthropic’s revenue because no one knows how much companies are spending on it. And Anthropic isn’t releasing data and it’s moving so fast and he says it’s like a data vacuum. So it’s really hard to actually get just even those basic numbers. So we just asked how much is your firm spending, which is a really low level question, but it’s very hard to get at.

But yeah, roughly every six months, because otherwise you just get AI question fatigue, so that’s the plan. And we’re going to, I think the idea is to ideally keep the questions the same because you have comparability. It’s another one of these issues that ideally you want to change to keep up with things as they’re happening, but if you do that, you can’t compare. So at least in June, we’re just going to use the same question.

Huiyu Li:

So in your survey Mamba check you did was that the forecast of the executives actually matched actual sales growth very well. So that’s very reassuring. Do you think that with AI being very new, some of this relationship could break down? I’m sorry, that was a leading question. I should have asked that without framing it that way.

Nicholas Bloom:

No, you’re exactly right. On the one hand, I think for things like sales forecast, you can say in a sense, you’ve got the wisdom of crowd. So if you survey… We had 6,000 people on these, so between 5,000 to 6,000, depending on the question. So the idiosyncratic errors will wash out because we’re equally weighted, but there could be a common mistake. So it could be every exec’s over optimistic or under optimistic, or you could be hit by a macro shock.

So we did this because I actually struggled to think of a better way to get data. This is like a glass is half full. I don’t think our data’s hugely reliable, but it’s hard to think of a more reliable way of getting hold of it. Because in the next three years, I think the big thing is not the tech, it’s the obstacles on the ground, like security and a person doesn’t like it or there’s some issue with PR, et cetera.

So, we just try to get thousands of execs in actual companies that are having to implement it. If you look 10 years out, I’d probably rather talk to technologists because they have a sense of the longer run horizon. But yeah, it may be that they mass get it wrong. For example, there’s a massive recession. It becomes easier to implement AI because it’s easier to fire people, for example, that would be like a common error.

Huiyu Li:

I see. Would you be able to include younger firms or startups? Because now I think AI is making it easier for even a single person to start a company. And I think you also mentioned on the job creation destruction side, the startup might pick up some of the job losses.

Nicholas Bloom:

Yeah. So we try and refresh, but we’re probably slightly underweight, because we look at 10 plus employees. It turns out, I mean, as you know this isn’t fair, but output and employment is dominated by 10 plus employee firms have something like probably 95% of employment and probably 97%, 98% of output. We focus on that, they’re a bit easier, but we’re slightly underway on startups. So yes, we would maybe get a more positive take, but it probably wouldn’t have a big effect on aggregate forecast because those are young firms. I mean, again, this is why I think I’m comfortable in a three-year window. If you looked at a 10-year window, you’d miss, I mean, Anthropic. In fact, the amazing thing, I didn’t know this event, Anthropic’s first dollar of revenue was in March 2023. It’s like, this is unbelievable. This is a company worth almost a billion… Wait, trillion dollars. Yeah, that’s an odd billion. I’m getting my numbers wrong.

Huiyu Li:

Yeah, these days I saw trillions now.

Nicholas Bloom:

Yeah, and they had no revenue three years ago. So yes, we’d miss stuff like that.

Huiyu Li:

Okay. So we have a little bit more time left. So just to switch gears, one question is, how do you see AI will affect the way you do research?

Nicholas Bloom:

Academics spend an enormous amount of time talking about it. I mean, I’m at the stage now, I’m doing some coding, not as much coding as I used to do. So I use AI a bunch. Oddly, I use it to do tax advice, health stuff, get advice. On the weekend, our fridge stopped working. And I was like, “Why is the fridge not working?” And I punched into GPT and it had eight things or something. And I was like, “I’ve done that.” And then one of them I hadn’t thought of and it worked. I was like, “Thank God, our fridge is back work, we haven’t lost all our food in the freezer.”

But in terms of research, it is clear that younger academics are using an enormous amount of Claude code. So like the example with Greg on Krusell-Smith, so anyone that’s coded up these micro to macro models, this Krusell-Smith is just a nightmare. I did it in my job market paper and one after that, because you have to have, you solve these multiple iterations. It takes a year to two to solve it, and they’re now doing it in a few weeks. My sense is it’s not going to… I don’t know if you’ve seen on social media, all these graphs of NBER working papers that have to change.

Huiyu Li:

Yes, yes.

Nicholas Bloom:

I don’t think personally it will make a huge impact on output because the biggest constraint is thought time, at least for me. You can’t write a hundred papers because you’ve got to think about them and make sure they make sense and present. And Tim Bresnahan always used to say writing papers is a kind of group activity because you present at seminars and people give you feedback, and why this is great because you get feedback on your questions. You think, oh. And so you can’t accelerate that bit, I don’t think.

Huiyu Li:

Yeah. There’s still the time constraint.

Nicholas Bloom:

Yeah, so I think it’s the most impactful technology I’ve seen, but I don’t think people are going to write entire papers off AI. I hope not. I’ve certainly refereed papers that have made up references. I mean, everyone must have this experience of, I refereed a paper that cited two or three of my papers. I’m like, “I don’t think I wrote those papers. I’m sure I didn’t.” And I went back, okay, I think I know what’s going on here.

Huiyu Li:

Okay. Unfortunately, that’s all the time we have for today, so thank you all for attending both online and here in the room. We have more EERN events scheduled this year, so please stay tuned. Our next event will be April 15th with Ben Jones from Northwestern University. He will present about AI and research and development. Thank you very much.

Nicholas Bloom:

Thank you.