Philippe Aghion | The Growth and Employment Effects of AI

Mary C. Daly:

Good morning and welcome. Whether you’re here in the room and I am delighted to see so many of you here or you’re online, we are very grateful to have you join our first discussion in the new EERN Network. Now, for those of you new to the San Francisco Fed, my name is Mary Daly and I’m president and CEO of the San Francisco Bank. I’m very excited to kick off our inaugural seminar of the new Emerging Tech Economic Research Network, more easily and lovingly called EERN. Now, we launched EERN in partnership with the Federal Reserve System Innovation Office, recognizing that emerging technology, including AI, are affecting the world around us.

At the San Francisco Fed, we think of ourselves as early students, studying issues like globalization, computerization, weather, and work from home long before they ever began to shape the aggregate economy. Our commitment to look ahead and see what’s next rests on the belief that understanding where the economy is headed is key to fulfilling the responsibilities that Congress gave us, including a safe and sound payment and financial system and a sustainable economy with full employment and price stability. In this context, EERN makes complete sense. It’s a natural next step in our commitment to learn before acting.

Now you might ask yourself, why is it called a network? Well, we did this for a reason. To truly understand the issues we face, we need to work together. We need to crowdsource. We need technologists, economists, business leaders, workers and policy makers to share information, debate ideas, and weigh the risks and rewards of these new innovations. These discussions that we will have are critical. Technological change is inevitable, how it affects the economy and society is not. That is a choice, that is our choice. So we built EERN to help us navigate these issues. But I know you came for our very special guests. So let me start. Let’s move on to the program. So as the very first speaker for our EERN virtual seminar, we are lucky to have the eminent voice in technology and the economy. Philippe Aghion is an economist and professor at the Collège de France. He will be discussing AI and the potential effects on growth and employment. Huiyu Li, who is with us here in San Francisco, is an economist and leader in EERN and a scholar of these issues in her own right. She will introduce Professor Aghion and lead our discussion. And with that, I’d like to welcome Huiyu Li, and I look forward to a great seminar. Thank you again for joining.

Huiyu Li:

All right. Hello everyone. Thank you for coming today. It’s my great pleasure to introduce our first speaker for the EERN Seminar series, Professor Philippe Aghion. I will call him Philippe going forward. I’ve been working with Philippe for almost 10 years, so I feel like I’ve earned that right. Yeah, so I’ll call him Philippe. So Philippe is a professor of Collège de France. He also holds appointments in two other institutions in SEAD and London School of Economics. Prior to all that, he was a professor in Harvard University in economics for over 10 years. He was also the president of the European Economic Association and also the president of the France Economic Association. So Philippe is a very prolific writer. He made seminal contributions in both innovation and contract theory. So today, he will have his innovation economic growth hat on. Philippe’s work has been cited by over 125,000 research papers. That is a lot in academia. So it’s not a surprise that Philippe is a recipient of many awards, including the John Von Neumann Award, as well as recently the BBBA Award for his research, Economic Growth Through the Lens of Destruction. So working with Philippe, I was fortunate enough to have a glimpse into what makes him so productive. He told me, cultivate your immaturity, which I can take means, that always be willing to explore new areas and also learn new tools. So I’m very excited that today, we get to explore together with Philippe, the potential impact of AI and Gen AI on the economy. The floor is yours, Philippe.

Philippe Aghion:

Thank you. In fact, the big secret is to choose your co-authors, and I’ve been extremely fortunate to have Huiyu Li as my co-author. She’s just amazing. And I think our joint work was enormously to Huiyu Li’s contribution. So it’s a great privilege to be your co-author, Huiyu Li. And thanks so much for inviting me to talk about economic impacts of AI. So the AI thing, I had been doing some work on AI and growth with Chad Jones, with a professor at the Stanford Business School, [inaudible 00:15:38] School of Business, Stanford GSB, and with Ben Jones, with the professor at Northwestern [inaudible 00:15:45]. But more recently, I was brought into AI because I had to produce this report for the French President Macron. And so we handed this report over to him four weeks ago almost, on what to do and how what could France do not to be left out of this important technological revolution.

So the AI revolution is really what we call a general purpose technology type of revolution because it affects all domains of activity, the economy, public services, the organization of work, the media, culture, everything. To a larger extent than previous revolutions, we had the steam engine revolution in the 19th century, then we had the electricity revolution in the 20th century. Then later on in the late 20th century, we had the IT revolution, and now, we have the AI revolution. The emergence of AI represents a non-precedented accelerator of generative AI. ChatGPT in particular represent a non-precedented acceleration in this overall AI revolution in particular because it simplifies the use of certain kind of tools to the extreme. And now, you can generate text, images, sounds at an extraordinarily high speed and with a stunning degree of realism.

So just to give you an idea of the acceleration, it took two and a half years for Netflix to reach 1 million users. It took two and a half months for Instagram to reach 1 million users. For ChatGPT, it only took five days to reach 1 million users. This AI revolution is unavoidable and it goes very fast and it creates both hopes and fear. The hope is that it would boost the growth process. We’ve been through a growth decline everywhere, and in particular, in the U.S. and with Huiyu Li, we try to understand the sources of this US growth decline since the early—

Philippe Aghion:

Sources of this U.S. growth decline since the early 2000s. And the hope is that the AI revolution will reverse trends and will boost growth again. And then we try to discuss that. The fear is that the AI revolution might create mass unemployment because it make a number of jobs redundant. We can now automate tasks that we would never have thought we would automate. And for example, generating images, synthesis of notes, even medical consultations, to such large extent, some of the tasks that are involved in those jobs might be now replaced by AI. And there are jobs like for example, translator or dubbing films, movie dubber that may become completely redundant once AI is introduced and generative AI is introduced. So there is this big fear of mass unemployment and there is what we call the existential risk. Is there truly an existential risk associated with AI that the risk that AI would completely substitute for human beings? So what I will argue here is that on the one hand, the belief that AI should boost productivity growth has some grounds to it, but there might be obstacles to productivity growth. So I will talk about both why we should believe that AI has the potential to boost productivity growth, but also we should be aware that there can be obstacles whereby AI may be an impediment to productivity growth. And I will talk about the two faces. It’s [inaudible 00:19:54] world you see there. And the second belief is that AI will create mass unemployment and destroy jobs, including skilled jobs. Here I will try to tamper that belief and say that things do not look as bad as one might think, but each time what will come out of my talk is that it all depends of the institutions and policies we put against the AI revolution.

I think all progress of humankind is always the joint result of technological progress and institutional change, and it all depends how institutions adapt or do not adapt to the technological revolution. And we will try to talk about which kind of policies would help maximize the potential positive effect that AI can have on growth and employment. So let me start with AI and growth. Why do we believe that AI should boost productivity growth? Well, because it automates not only the production of goods and services, but it also automates at least partly the production of ideas. And that’s very much a view that we pushed in this joint work with Chad Jones and Ben Jones. So in particular, why does AI automate partly and improve the production of ideas? Because it helps find solutions to complex problems. It facilitates imitation and learning and AI can become even self-improving and that’s something that you could not do before.

So I want to put many equations here. That’s a way and you just follow the idea there. Suppose you produced a flow of output Y with inputs X1, X2, Xn. You can call them tasks. And X will be labor if the task is not automated and X will be capital once the task is automated. You can use capital instead of labor. So at the end of the day, you can express capital Y as a function of labor input and capital input and it becomes like this. And the rate of growth of Y over L, which I call small Y, you can express it as the growth rate of A, which is productivity growth that would stem from innovation, divided by this one over alpha. But the alpha is the extent to which the overall economy is automated. Now, what will happen is that automation will boost growth for two reasons.

On the one hand, the AI will automate the production of goods and services, thereby it’ll increase a fraction of tasks that are using capital instead of labor, the alpha. That will boost the growth of per capita GB because you replace labor which is in limited supply by capital, which is on limited supply because we produce capital. But on top of it, you see AI improves the technology to produce ideas. And because it improves the technology to produce ideas, it increases the growth rate of A. AI increases the growth rate of per capita GB for two reasons. It increases the alpha, that’s the production of goods and services, and it increases the GA, which is the fact that it automates the production of ideas.

So now let me illustrate a little bit. There is this recent paper by Brynjolfsson et al. called Generative AI at Work with Danielle Li and Lindsey Raymond. And what they do there is that they look at a particular firm. So the field this firm is operating in is customer service. It’s a field with high AI adoption rates. And so what in fact is done is that you have a firm that belongs to the Fortune 500 company. It’s a Fortune 500 company. And this company advises small and medium-sized enterprises, U.S. SMEs on enterprise software. So the job mainly consists in answering the managers of these small and medium-sized enterprises, the question they have on how to install the software. So mostly there are Philippine employees of that firm. And when the SME managers call this firm, they are sent to these Philippine employees.

So now some of the Philippine employees do not have access to ChatGPT and then some of them have access to ChatGPT. And what this paper does is to look the difference it makes when these Philippine agents have access to ChatGPT, what difference does it make to how long it takes to respond to a question and also the quality of the response. So for example, you have a visitor who says, “My name is Alex. I’m super frustrated. I had customers calling me all day saying they can’t access their information on the website and I need this fixed right away.” And then the Philippine employee says, “I completely understand, Alex.” But then what’s interesting is that you have here a sample AI-generated response. Alex, I can definitely assist you with this. Can you please provide the email associated with your account? It’s nice to meet you Alex.

And what used to be done purely by manpower now has the help of AI of ChatGPT. So now what will happen is that the ChatGPT will suggest response to the Philippine agent. The agent may or may not endorse those response. And then we’ll get back to the small and medium-sized enterprise manager. So here shows the progressive diffusion of the ChatGPT among the employees of that firm. It’s progressive, so it’s good because you have each time a treatment group and a control group. You have the manager, you have the Philippine agent that just received access to ChatGPT compared to a similar agent who does not have access to ChatGPT to respond to the requests of the small and medium-sized managers. And what you see here is that you compare. So here you look, it’s a difference in difference.

I compare before and after my Philippine agent has access to ChatGPT compared to a similar agent that does not have access at all at that time to ChatGPT. And you see that ChatGPT induces a significant boost in productivity. The number of resolutions per hour jumps up by 14% after the first month and then to 25% after the second month. And so there is clearly an increase in productivity here in that particular case induced by the access to ChatGPT. So now that’s a very study and what you can do is to try to go from micro to macro. And that’s an area where research is very much left to be done. So the best we can do now is to extrapolate from previous general purpose technologies. We can say, well, you had the electricity revolution, you had the IT revolution. What happened with them? Usually after the first big discovery was made, nothing happens for 10 years say and then it starts boosting growth, you see?

And then you have the diffusion of the general purpose technology throughout the economy and that’s where really you can observe the boost in productivity growth at the economy wide. So if you just extrapolate from previous GBPs, you will say that AI and machine learning was invented about more than 10 years ago. It should be about the effect of AI should be seen in growth figures. And you see the dotted curve tells you how GDP should evolve without the AI revolution. And with the AI revolution is the green curve, and you see that the slope increases as of now for 10 years. And then after 2035, the slope of the GDP curve with AI becomes, again, the same as the slope without, except that you are at a higher level of GDP. So that’s the effect that we should… And what we predicted, if we take a conservative average of the electricity and IT revolutions, you would imagine a 1% increase in growth rate of per capita GDP over a ten-year period.

That represents a lot actually if that happens. But now remember that that doesn’t take into account the fact that unlike previous revolutions, or at least to a larger extent than previous revolutions, the AI revolution automates not only the production of goods and services, but it also automates, at least partly, the production of ideas. That mean that it makes the research technology permanently more efficient. The researcher can use ChatGPT or many other devices and can go much or can solve much more complicated problems and do the hard part of research delegated to AI and concentrate on the more creative part. And we believe that therefore the research technology should be durably boosted by the AI revolution. And that’s what we have here. That’s what we have, the green continuous curve. So automating production of goods and services at the red curve here on this graph, automating the production of ideas is the continuous green. And then when you add up the two effects, you get the blue curves. And so the blue curves, it tells you that not only you will have a big boost over the next 10 years, but you will still have a boost, you will have a steeper growth even after 10 years than you have without AI. So if you look this way at the AI revolution, you are extremely optimistic. You say, my God, we will have a big boost in growth, not only bigger and longer lasting than the boost induced by the previous technological revolutions. That’s the nice view. But on the other hand, and here you see what this shows. This shows what happens when you miss the AI revolution. You see, that’s what I tell my French colleagues. Here is the example of China. China was a leading power up to the 17th century, but then the Ming and Qing dynasties, and here I speak under the US control, decide to oppose to international trade for fear that the new technologies, the new imported technologies would pervert Chinese society.

So they went as far as forcing Chinese citizens living on the south coast of China to move 30 kilometers into mainland to avoid the bad influence of trade. And what you see is that whereas per capita GDP of China was only 10 or 20% lower than that of U.K. or France by the mid-17th century, China missed the effect of the industrial revolution that were initiated in Europe. And by 1900, the Chinese per capita GDP was 10 times lower than that of U.K. and France. We don’t want the same to happen with AI. That’s why [inaudible 00:32:20], but Emmanuel, you should not miss the AI revolution. You have to invest heavily in the AI revolution and also unleash data access and all those things that help new AI firms to come in and helps also the whole set of French firms to adopt AI because we can’t afford to miss the AI revolution. Now, it’s something I would like to say. I told you that AI has a big growth potential, but on the other hand, I would like to have a warning there.

And the warning comes from my joint work with Huiyu and Pete Klenow, Timo Boppart and Antonin Bergeaud. And what we showed in this, first we noticed that if we look at the average yearly TFP growth rate, it went up between ’95 and 2005. Bob Solow said we see computers everywhere, but not in the statistic. But that was right before it was here. It was right before the IT revolution showed up in the growth figures. You can see an upsurge of growth of TFP growth between ’95 and 2005 and then a big decline in growth. And you can see that this upsurge followed by decline in growth is true, is observed mostly in the IT producing sector, which is the black curve and the IT using sector, which is the gray curve. It has to do with IT a lot. And what we could see also is that the entry rate of all firms went down since the early 2000s.

And the story that we develop, and we argue with Huiyu in this paper called A Theory of Falling Growth and Rising Rents is that with the IT revolution, you had the emergence of superstar firms, the Amazon, Walmart, Microsoft, Google, et cetera. These firms pushed the growth process, they were better organized. And so as a result of the IT revolution, they could really expand, which first boosted the growth process in the U.S. But then the problem is that through merchant acquisition and uncontrolled merchant acquisition, they became tentacular, they became overwhelming hegemonic and they ended up discouraging entry of new innovators. And that’s why growth went down. So the problem is a competition problem, is that the competition policy in the U.S., and Richard Gilbert from Berkeley explains that very well in his latest book published at the MIT Press, the competition policy in the US did not adapt so as to prevent that these firms would become superstar firms and inhibit the entry of new incumbents.

And the problem with the AI revolution is that the AI revolution already from the start, the upper segments of the AI supply chain is dominated by few actors. In particular, the cloud is dominated by three superstar firms, Amazon, Google, Microsoft, and there is only one big actor on the market for graphic processes. So the big problem there is that right from the start, the upper segments of the AI value chain are dominated by the few number of large firms. And the fear is that the lack of competition would, from the start, be an impediment to growth, would suddenly limit the extent to which the AI revolution can boost productivity growth. So there is enormous promise from the AI revolution in terms of growth enhancement, but on the other hand, more than ever, competition policy has to be adapted. Has to, in particular, factor in much more than before entry and innovation in the way it is designed. Okay. So far, the way you design competition policy is very much focused on market definition and market share, not so much on the extent to which emerging acquisition would discourage future innovation and future entry. And more than ever, that change, that improvement in competition policy is needed if we want the AI revolution to fully deliver on its growth potential. Okay? So that’s really a strong conclusion that we have. Okay, so let me turn now to AI and Employment, that very much based on first paper with Céline Antonin, Simon Bunel, Xavier Jaravel on Automation and Employment using French firm-level data and ongoing work with Simon Bunel, Xavier Jaravel, Alexandra Roulet on AI and Employment, again using French firm-level data. Okay. So they’re very much… I know that some people in the U.S. have a very pessimistic view on automation and employment. I will not name them. They’re my colleagues, my co-authors, I like them very much, but I have to confess a profound disagreement with them on this. They think that basically automation is bad for employment. Okay. And what we did on our first paper on automation and employment in France, we have a comprehensive set of French firms and we look at the effect of adopting industrial equipment or adopting robots. And we see, we compare a firm that adopts robots or to other automation device, to similar firms that do not adopt it. So that [inaudible 00:37:45] even study. And you see that the firms that adopt automation, employment goes up. So here it’s kind of modern industrial equipment. Here, it’s robots. And why is that? And you see, it is because in fact, firms that automate, they become more productive and therefore because they become more productive, they become more competitive. The quality price ratio of those firms increases.

So their sales go up. What you can see here is that the sales go up on the left-hand side and what you can see on the right-hand side is their export sales go up. They get a world market that increases and therefore, there is more demand for their products and therefore they hire more. So this productivity effect more than counteracts the direct effect that you substitute machines for human labor. And that was true for automation. Of course, when you automate, the employment of competitors may go down. So that’s true. But on the other hand, it’s not a zero-sum game because your product becomes cheaper. You reach consumers that would never have gone neither to you nor to go to the competitors. So the total side of the market serves on those goods increases.

So it’s not that you just take, you steal business from other firms that do not automate. You also attract consumers that were not at all interested by this kind of goods. And that’s why the overall effect, even at industry level is positive as we find in that paper on Automation and Employment. So as I said, there are two effects. On the one hand you have the eviction effect, automation displaces task from labor to capital. But on the other hand you have this productivity effect. Automation increases productivity on existing tasks, which increases quality price ratio, thereby increasing sales, export sales, and thereby creating jobs. And that productivity for automation seems to dominate. Now, how about AI?

So what I’ve done here is two kinds of things, and that’s been done from the report. And we are now continuing the research. We have through the French Statistical Institute called INSEE, we have ICT French level annual survey and we have specific question on AI adoption in two surveys in 2019, 2021. And so the survey covers 9,000 representative firms with more than 50 employees. And we have even studies that compare between firms that adopt some AI between 2018, 2020 and similar firms that do not adopt AI by 2020. So we look only at firms that are both in the 2019 and 2021 surveys. So that boils down to 321 treatment group firms adopt AI before 2020 and almost 900 firms in control group that did not adopt AI in 2020. So we compare the employment of a firm that adopt to the employment of a similar firm that do not adopt.

Okay. So here it shows which AI technology firms that adopt use, mostly machine learning, analysis of written language, robotic process automation, those seem to be the main identifying objects of people from images, automatic speech recognition. You can see the main use of AI technologies. What’s also interesting is that through the survey, we can see that kind of why do you resort to AI, mainly IT security, marketing or sales, business administration process, production process. Those are the main ones, the main reasons why and then come business management, the reasons why you adopt AI. And now come the main result. You see, I compare a firm that adopts AI to a similar firm that does not adopt AI by 2020. And you can see that employment goes up, it’s exactly the same as with automation. And for the same reason you increase productivity, okay.

And it is true for males and females, but it’s not uniformly true over all jobs. For example, the effect on employment in administrative and commercial intermediate professions, for example, executive secretary, administrative service, legal service sales, those are jobs that can be put at risk. You see where. At least there is no obvious positive effect of AI. So you see overall, there is a positive effect of AI employment, but it’s not uniform across all jobs. There are jobs that do not benefit. And that’s very interesting because it echoes recent study by the International Labor organization, sorry, called Generative AI and jobs, the global analysis of potential effects on job quantity and quality. And this paper analyzes the exposure of various tasks and jobs to Generative AI, particularly ChatGPT.

The way it does is that a job is decomposed in a number of tasks and for each task you put a score less than 0.5 if there is a small risk of replacement by AI or a score between 0.5 and 0.75, if there is medium replacement risk of the task and a score above 0.75 is there is a high replacement risk on the task. And what’s very interesting here is this figure that they get. You can see on the top hand managerial jobs. So managerial jobs, you can see service managers, not extra classified. All the top graph is the managerial jobs. And you see that most tasks are in the dark blue. Those are tasks with very low risk of replacement, but they get some risk of replacement for some tasks which are the orange yellow. What you want is that you ideally for a manager or managerial task, you would like that some tasks be replaced by AI, which are the kind of boring tasks so that you can have more time to devote to more creative tasks.

And that’s it. The best tasks are the one at the upper hand where you have some yellow orange, but most of it is blue, dark blue. And those are the ideal tasks because those are the ideal of the jobs because those are the jobs where AI makes a difference, but for the better. Okay. Those who are on the bottom of this upper graph, AI makes no difference at all. Where AI really improves things is the jobs at the upper end where it replaces part of the task, but most of the tasks are not replaced. Those are for the managerial type. But then if I move down to the downward graph, to the lower graph for the clerks, there you see that most tasks are in the yellow orange. And so most types there are at risk and some of the jobs are typically at risk. And so that matches the study that we had for France where we saw that for these kind of jobs, the effect of AI was not positive.

Now, what we’ve done also in France is to classify, is to have the same methodology of the International Labor Office, but to look at the effect of AI on various types of job listed by the French Labor Ministry. We counted 222 types of job listed by the French Labor Ministry. On the Y-axis, you have the extent where a job is globally exposed to AI, is it concerned by AI or not? And the X-axis is the share of tasks that are hard to be replaced by AI on that job. And when you are high and left, those are the most vulnerable because the high end left here, those are jobs that are highly exposed and where the biggest share of task will be replaced by AI. For example, someone who does tele-sales, you see, an accountant or secretary, there, some of them will be at risk, you see, but the accountant and tele-salesman will be very much at risk.

When you have the high right, which is here, you have jobs that are exposed to AI, but in fact, very few of the tasks will be really replaced by AI. So those are jobs where AI will transform the job but will not put the job at existential risk, you see. And when you have the low and right, those are the most immune to AI. So that’s very interesting to see the various type of jobs, you see here you have accountant and tele-salesman, secretary. Here you have more like journalist, architect, a lawyer. Those are jobs that will be affected by AI but will not disappear with AI. And those are jobs that will not be affected at all by AI. So it’s very interesting to see this heterogeneity of jobs. AI has a globally positive effect, but it affects jobs differently, you see?

So now this graph is very interesting because it shows on the horizontal axis the risk of replacement of the job. And on the vertical axis it shows the cumulative share of total employment that this represent. So the way to read this graph is to start from the most at risk jobs. I told you accountant, tele-sales, those are high-risk jobs. They are here but they don’t represent enormous in term of share of employment. And then you move to secretary and you add the secretary, they are a bit less at risk, but the cumulative of accountant secretary gets you there and essentially you see, you get to write there and those are the one at risk. But the one at risk, they represent essentially 20% of the jobs and then the risk is much lower. You see what I mean? So the thing is that really the thing that very at risk represent here, if you read this graph from there to going down like 20% are really at risk.

So you have a global effect positive on AI, but you have 20% of the jobs that are at risk of replacement. Okay? So overall there is no existential risk of AI. I mean we don’t anticipate mass unemployment as a result of AI. The same way as the steam engine revolution did not cause mass unemployment. There was the Luddites movement, the Luddites in England were very much protesting steam engine revolution because their fear was, they were scared that to become redundant fully and nothing of the sort happened. Again when you had later on the electricity revolution, Keynes predicted mass unemployment, which never occurred. And the same with IT. And I think each time it’s because the productivity effect is there to prevent this mass unemployment from taking place. But still you need appropriate institution and policies for AI to boost employment. Okay. So I told you already that if you want to make sure that AI delivers on its growth promise, you need to adapt competition policy.

Well, in the same way, if you want AI to deliver on the employment front, you need adequate education and labor market policies. My model is Denmark. Denmark has a very good education system and on top of that, you have very good labor market policy. So let me talk about the labor market policy. In Denmark is what I call flexsecurity. So flexsecurity is that in Denmark, when you lose your job for two years, you receive very generous unemployment subsidies and the state helps you find a new job and they propose new jobs to you. If you refuse more than two jobs in your qualification, the unemployment subsidy go down. It’s a very good system because it accommodates creative destruction. Firms in Denmark are free to hire and lay off, but at the same time it protects employees, you see.

Here, it’s a work, it’s based on work by Alexandra Roulet with my colleague at INSEAD, and she looked at the effect of becoming unemployed on various health indicators in Denmark. She compared the health of a Danish worker in a firm that closes down so that the worker loses her job to the health of a similar worker in term of age, experience, education in a firm that does not close down. And it’s a difference in difference between the two kinds of workers before and after the first worker’s firm closes down.

And you see no effect of losing your job on purchase of antidepressant, antianxiety or sleeping pills. No effect of losing your job in Denmark on probability of having a circulatory problem. No effect on mortality. That’s very different from what happening in the U.S. where Anne Case, Angus Deaton spotted the Death of Despair phenomenon. They said, no, in the U.S. you have stress. When you are losing your job, you might lose access to health. It may cause disruption, family disruptions, loss of status. And that’s why people start taking opioids. They start taking sleeping pills. They start eating pizzas and become fat. And that’s the thing. And that in the U.S., you have the kind of sharp rise of the death rate, of the mortality rate of middle-aged, unskilled workers in the U.S. That’s very well described by Anne Case and Angus Deaton, no such thing occurred in Denmark.

So my view is that you see, okay, the AI revolution, it will globally create employment. You will need to retrain the 20% of the workers who might lose their job into a new job. But that you can do if you have a good education system because at school you learn to learn primarily and you have a good labor market policy like in Denmark where the state is there to help you to retrain and find a new job. And I think that is a very good kind of policies. So that’s what I want to say. The AI revolution is full of promise on growth and on unemployment.

But you need adequate institution and policy, particularly I would emphasize competition policy, education policy and labor market policies. So a lot of that I drew some kind of advertising from my book, Power of Creative Destruction, chapter six of the book, largely based on the work with Huiyu Li. And then there is another book where Huiyu Li has a beautiful chapter with Timo Boppart in this book, which is a [inaudible 00:52:49] book, the Economics of Creative Destruction, edited by Ufuk Akcigit and John Van Reenen with foreword by Emmanuel Macron. Again, thank you very much. Thank you.

Huiyu Li:

All right, thank you, Philippe for a very thought-provoking presentation. So I think it’s very interesting that you mentioned the role of firms in AI innovation. So a report by Stanford Human AI Institute finds that since 2014, industry has taken over academia in terms of producing machine learning models. And one reason they posited was that state-of-the-art AI systems just requires a humongous amount of resources in data, GP power. So now firm has that resource and they’re taking over. So given firms are now a main player on AI research, what role do you see for market or policy to make sure innovation continues?

Philippe Aghion:

My view there, you see, Huiyu Li, is very much, I had done work on that with Dewatripont, Mathias Dewatripont and Jeremy Stein, and then later on with Fiona Murray, Scott Stern and Julian Kolev looking at the importance of basic research as well. You see, there will be several waves in this AI revolution and a lot of … Many of them initiate in basic research. You see, for example, vaccines. Of course, we developed vaccine, Moderna, Johnson, Pfizer developed the new vaccines against COVID but RNA Messenger was invented in university. It was in a lab. It was basic research. I believe very much in the fact that you need freedom and openness at the basic stages of research. And then you get into increasingly more focused and firm managed research. But I don’t think it would be desirable that not to have a total role for basic research. In fact, when you look at the big actors in like Yann LeCun, for example. Yann LeCun, he was in my commission. He’s the head of research at Meta.

It’s very known, Meta, but Yann LeCun is a big figure there. He received the Turing prize. He started as a university person and he’s always very much linked to universities. He needs the interaction with university and they hired him from university. It’s no coincidence that Silicon Valley is close to Stanford and Berkeley, I should say. That’s how I said Stanford. What about Berkeley? Stanford and Berkeley. It’s no hazard. I mean, you need to have basic research nearby because you have this freedom and openness and then people can go into hybrid arrangements between university and labs and then go fully into private sector. But I don’t believe that the private sector alone can do everything. You need innovation chain and it starts with basic research and those … The early steps in the innovation chain are much better performed if they are under freedom and openness. That’s my answer to your question.

Huiyu Li:

Thank you, Philippe. The results you show for Denmark was very striking, and I guess it points to perhaps hope that we can reap the fruits of AI without causing a lot of pain. I want to compare a little bit with automation’s effect on manufacturing. I feel like for automation and manufacturing, they were replacing skills that are very specific and maybe those workers have very specific skills that were hard to retrain to another job. And also it was very geographically concentrated. When manufacturing plant goes down, the entire town gets affected. So right now, my prior is that AI, the types of skills it replaced will be less specific to a location, less specific to a particular industry. So maybe it’ll be easier to retrain these workers. What do you see on those fronts?

Philippe Aghion:

Yeah, I think that most jobs, I told you the jobs, I can see no more than 20% of jobs that will really fully be at risk of being redundant with AI. Most of the jobs will have part of the task replaced by AI but concentrated will have to be restructured and reorganized, but they don’t need to … They won’t disappear. But I think it’s important, of course … The schooling system is important. It is the combination of the schooling system and the labor market policy because at school you learn to learn. So you can decide to move from one sector to another. For example, having new sectors like helping elders, we have life expectancy that goes up. You need to help the elder population there. I think there is enormous scope for job creation. So I think that’s what will be. But it’s true that that’s where you need to have very active labor market policy to help those workers that are really losing their jobs to retrain and move to something else.

But I think that can be done. I think it can be done. If they start from a good level of education and a minimum education, they can retrain. In Germany, for example, many people have two careers. They go up to the age of 50 and then they become travel guides and they do things completely different. It’s amazing. They have two lives and they’re prepared for that.

Huiyu Li:

Yes, retraining is possible. Okay.

Philippe Aghion:

Yeah. But it requires a good education basis. You see what I mean? Because it’s at school that you learn to learn.

Huiyu Li:

Okay, so we have some questions from the audience. I want to make sure we get to those.

Philippe Aghion:

Yes.

Huiyu Li:

Actually, we had a lot of questions submitted to the audience when people registered, but also currently online.

Philippe Aghion:

I see.

Huiyu Li:

So let me first start on competition.

Philippe Aghion:

Why did you invite this guy?

Huiyu Li:

Yes. So the question is touches on competition. So some have argued that AI can be democratizing for business since everyone can use it, but then you also suggest that concentration of the core providers could move in the opposing way, could be barriers. Can you talk a little bit how these two forces may net up, how they may interact?

Philippe Aghion:

I think it’s very important exactly for the reasons that you … and together we spotted you. [inaudible 00:59:19]. It’s very important … I’m very much in favor of open-source AI. You see what I mean? Why is it important to have AI as open source as we can? Because first, it’s very important to evaluate AI. You see … And secure the models. You see what I mean? And so there are some AI … You know that there’s been some type of perverse use of AI and it’s been very important for AI to correct itself. So making it more open source helps AI to correct itself. It also helps personalizing models and adapt models to different contexts, in particular, linguistics. Also, open source is good for the environment because it avoids that each of us does retraining of its own model of foundational model. You see what I mean?

With open source, you can save on the energy use of AI, you see, which is a big issue, the AI and climate. And finally, last but not least, you reduce entry barriers. Anything which is towards data sharing facilitates the entry of new firms because sometime what you have with the competition is that you have firms entry that prevent entry of new firms. And that’s what you want to … So for example, you have … You see, you have some platforms, you see that the gatekeepers, they have many users, they abuse their position to prevent new potential entrants. And so what you want to do is to want to have known that they can’t prevent new entrants. So then we have to adapt data access. Also, if we see that in some merger and acquisitions lead to less entry and less innovation, we have to break up and to reverse the merger and acquisition, I think we should have a very more dynamic way of doing competition policy and be very much open source.

Huiyu Li:

So another big issue—

Philippe Aghion:

To prevent the phenomenon we saw in our own paper.

Huiyu Li:

Thank you. So another big issue in AI is that AI is not perfect. There are a lot of issues and uncertainties still with this technology. So when institutions, for example, government institutions that needs to be very careful and think about how to adopt these technologies, what suggestions do you have for thinking about how to approach these new technologies so that we reap the benefit—

Philippe Aghion:

No, it’s very important. Yeah, it’s very important, for example, to have good jurisdictions, good regulations. You should not have too many regulations, but you need to have good regulation, in particular protecting personal data. That’s very important. So there must be a policy to protect personal data, to avoid abuse. So there is a trade-off you because if you have no regulation, then very bad things can happen. If you have too many regulations, usually the big firms can survive the regulations and those create entry barriers to small entrants. So you need to find the right degree of regulation, you see, because if you have too many regulations, it goes against competition, but you have no regulation, of course, you have danger. I can tomorrow pretend I’m UE and do terrible things and I don’t want that to happen. So that you do through regulation also through me, my amount of openness that allows, as I said before, that makes correction easier. So that’s a very good question, which amount and which kind of regulations you need to make sure that you avoid abuse, but you don’t prevent entry.

Huiyu Li:

Okay. And you mentioned schooling being a very important aspect. So actually we had a lot of interest in this talk from the student folks and also from people about to enter the labor market. Do you have any advice for these folks about how they think about their training so that they’re prepared for AI as they try to enter the labor market?

Philippe Aghion:

Yeah, I think they should be acquainted with AI very soon. It’s very important when you enter a profession, whether it’s you want to be a lawyer or whether you you want to be a chemist or whatever, yet you learn about what AI does to this profession. And very interesting. AI does very different things to different professions, and I think it’s very important to get acquainted with AI, with the good and bad use of AI and with AI regulations. And I think that should be part of the curriculum. I think in France, in the report, we pushed for AI to be part of any university curriculum and adapted to the kind of things you learn. So the AI that you learn as part of a low degree is not exactly the same AI that you learn as part of STEM degree or whatever. And I think it’s very important to be acquainted with AI very soon and with the danger and how to use it or misuse it or not misuse it. And that should be part of basic … should be economic 101.

Huiyu Li:

So we also have questions from the audience about distributional consequences. So distributional consequences in terms of quality across countries, developing country versus developed countries within country across different income distributions and also within country across different areas. Is it going to deepen the rural-urban divide? Do you have anything you would like to share on these aspects?

Philippe Aghion:

Yes. Well, the Brynjolfsson study that I mentioned before has a very interesting conclusion. They looked not only at the extent to which AI increased the productivity of these Philippines employees that give advice to American small and medium services on software, but they looked differentially across productivity levels of the prior productivity and prior experience. And what was very interesting in this study that they showed that it was primarily the least productive employee that benefited the most and the least experienced employee that benefited the most from ChatGPT. You see.

So that’s very interesting because in that respect, AI reduced inequality. But on the other hand, it could mean also that you become one hiring policy, you become much more skill-oriented and it might increase inequality. So it’s very interesting topic, the effect of AI and inequality, but it’s not unambiguously bad. There are at least some respects, some dimensions in which AI may reduce inequality. There may be others like for example, hiring the skills. Do you hire more … Do you go more for STEM diplomas or less so. On the hiring policy, it may increase inequality, but once people are hired, it seems that it reduces inequality. So the subject of inequality and AI is a very interesting subject and by no means unilateral, by no mean unambiguous.

Huiyu Li:

Great. One more question about AI helping with the production of ideas. I think you show that AI, which could be different from previous GPTs, GPT being general purpose technology—

Philippe Aghion:

Exactly. Yeah.

Huiyu Li:

Not GPT, GPT that it could help with the production of ideas. So is there a concern that maybe there’s disinformation that could muddle this process of generative AI improving the production of ideas that there could be some vulnerabilities with AI that could actually make the production of ideas worse?

Philippe Aghion:

That you would generate bad ideas, you mean?

Huiyu Li:

Yes.

Philippe Aghion:

Is that what you mean?

Huiyu Li:

Yeah, bad ideas because you have disinformation, you have fake news, deep fakes. Could that actually make the whole—

Philippe Aghion:

Oh, yes, of course. You’re absolutely right, but they’re … absolutely. That’s why you need a regulation. That goes back to your previous question. You need a regulation, but you also need enough openness for AI to correct itself. You see what I mean? So that bad news, everybody would know and there will be social sanctions. So that’s why it’s important to have a mixture of regulation and openness. You need both to … because what’s interesting is that AI very quickly now spots the abuse and corrects the abuse. I told you that AI is self-improving, but it’s also self-improving in those dimensions. But it cannot be self-improving if you don’t have a minimum amount of data access. And if the system cannot … because the data … AI, it’s a computing power and its data. If you don’t have the data, you can’t do anything. So you need enough openness, enough open source, enough data access for AI to be able to improve itself and denounce the bad behavior. So it’s a mixture of regulation and openness that’s a response to it.

Huiyu Li:

Okay. Thank you. So one last question, Philippe. You are an economist and a professor. How do you think generative AI will affect your work?

Philippe Aghion:

Yeah, that’s a big deal. Probably you know, it took me a very long to learn a bit about [inaudible 01:08:34]. I’m very bad. I’m completely schizophrenic because I studied innovation and I’m very, very bad at adapting myself to new innovation. I rely on my co-authors a lot. That’s very bad. That will force me to myself adapt. And maybe now hopefully … That’s why I was very happy with the Brynjolfsson result that the least productive, the least experienced may benefit most from the generative AI. Maybe I may be one of those. That would be my hope because I start from very, very … My starting point is very low.

Huiyu Li:

Thank you very much, Philippe, for this.

Philippe Aghion:

[inaudible 01:09:10] in theory, but I’m very bad at it.

Huiyu Li:

Thank you. Thank you very much, Philippe, for this lovely presentation.

Philippe Aghion:

Thank you.

Huiyu Li:

And your answering of the questions. Okay.

Philippe Aghion:

Thank you so much.

Huiyu Li:

Thank you.

Philippe Aghion:

[Inaudible 01:09:24]. Great. [inaudible 01:09:35].

Huiyu Li:

All right. Thank you very much for coming to our first seminar. I hope you’ve had some interesting takeaways with you and maybe raising new questions.