Your data shows higher perceived automation risk among younger workers and people without higher education. Walk us through who that person actually is. What do they do? What is their background? Where do they live?

It’s quite general; age is the strongest predictor. And then you see it in occupations where pre-ChatGPT people were already more thinking about it. Where tasks are kind of, you can imagine the robot arm coming in and doing it. In those sort of occupations, the fear doesn’t speak so directly to what would happen. I’d love to redo this survey now and see how the distributions have shifted. Now I expect that white collar would be much more in fear as well. Big shift, but again, also towards younger.

I think it’s largely because they’re the ones who are most exposed to technologies and most familiar with them. They’re more and more likely to be using technologies. And they have the least job-specific skills that they’ve already acquired. I think a lot of people who have been sitting in a job for a long time might think, oh yes, this job has things that can be automated, but I have all this institutional knowledge, firm-specific capital, client connections, and whether true or not, for people who’ve been around the block a bit, they think these components are much more important, which can not be taken over by the AI.

You’ve found that automation moves workers wanting to unionize and demanding redistribution from their government, but it doesn’t move them to retrain. And this is exactly what our policymakers are telling us. What do you think is the reason for that gap?

I think retraining is something that, number one, takes time, effort, typically costs money, and creates a lot of uncertainty. If I say today I’m going to retrain to become something else, I not only have to learn that, I have to pay for a course, but I also don’t know what job I’m going to find afterwards. Am I going to find a job? What’s that job going to look like? That’s a jump into the void that people don’t take so easily.

And there’s also doubts about how easy it actually is. People, as they get older, anyone who’s tried to learn a foreign language, knows it gets a lot harder as you get older. That’s why I think we don’t see this immediate response of people saying, oh yeah, I’m going to re-skill, change horizontally, something similar but different, or vertically, that I really go and upskill. And because, hey, maybe that’s also not useful with AI.

I wasn’t too surprised that people were not planning to respond on that margin so much, just because of the cost and the lack of certainty involved with it. Ex-ante, I didn’t expect the response to unionization, especially because unions in the US are not the most popular thing and most prevalent. If we had done the survey in Europe, I might have expected that more, but it is an indication that people are attached to the status quo. We know from many studies that status quo protection is strong. People don’t like something that has been taken away from them. So one way you can protect your job is basically to join a union.

I’m not surprised about the request for redistribution. That was ex-ante probably our biggest hypothesis. Because that has always been a bit the case when things fail, people return to the state. But especially in recent times with COVID, it has become more common that the state is seen as someone who has to provide a solution. We see that now with every energy crisis. Prices are going up and immediately there’s requests for handouts to finance this, caps on energy bills. State-provided solutions. And that’s why that, for me, was very expected given current trend.

You made me think about another one of your papers. I don’t remember the name of it, but it was about workers who switch occupation earn less, right? It got me thinking that perhaps the reason for them not retraining or learning new skills is because it’s hard to make moves when you’re afraid. Do you think fear is the biggest reason, or are there other things involved?

I think it’s both the uncertainty. If someone wrote a contract with you and said, “if you learn this, then you get that,” then people would much more willingly do it, maybe. But since it’s extremely uncertain, I might do this retraining and be in a worse situation than I was. I have a gap in my CV, I paid a cost, and I don’t find anything. And maybe my prior experience does not count for this job. I switch occupations. I do not have that firm-specific capital. I do not have that occupation-specific capital. I start from zero, where I’m competing with people who are younger than me and have newer skills.

That’s why someone might not want to hire a new applicant. I often hear that story, for instance, for women re-entering the labor force after having children. If they stay out of the labor force for a while, it’s often problematic that the employer looks more to hire someone coming straight out of education rather than someone who has been detached from the labor market and education for some time. Since there’s neither the guarantee nor the statistical increase in your mean, both the mean and the variance are speaking against switching occupations with uncertainty. That’s why people are not just like, “oh, I’ll go and train to learn something else.”

I’m going to need you to grab a crystal ball, if you have one. If a 20-year-old came to you and asked, given everything your data shows, what should they be studying right now, or what would be useful in the labor market in let’s say 2035? What would you tell them?

We have to hypothesize about this. But anyone right now who presents their projections as the truth, either they’re going to be extremely lucky or they overestimate their predictive capacity.

If we look back, every technological revolution is different, but there’s patterns we can learn from history. When there’s been big technological innovations, it’s not always panned out exactly as one might have predicted. If you look at when computers took off and were introduced, it’s not like they’ve replaced everything that a computer could do. It depends a bit on whether there’s high demand for that job, whether it’s constrained in its supply. So it’s just a cost, but it might anyway increase demand.

Let me give you the example of, say, a lawyer. That’s a typical one that one thinks they’re in trouble with, with ChatGPT, or with whatever LLM. But in the end, it’s not 100% clear, because most places do not want to, often literally cannot, rely on LLMs. So maybe lawyers will just become even more productive. What can happen here? Maybe one lawyer can now do the work of 10 lawyers. That would mean we’re going to need fewer lawyers. But at the same time, maybe many more people are going to want access to lawyers because now they can use them. I typically didn’t use a lawyer for something. Now I’m like, hey, wait a minute, now I can. Lawyers are becoming much more productive, maybe I can hire a lawyer as well. So you have those things pushing in two directions. On the one hand lawyers are going to be substituted by LLMs; on the other hand, it might increase demand for lawyers.

That goes for a lot of different professions.

One thing that is obvious is, well, manual labor, in many cases, is harder to replace. The typical example of the plumber and the carpenter. Those skills are just harder to replace. But at the same time, a lot of knowledge work is still going to be required. People will have to verify codes written by LLMs. You still need someone to orchestrate all the agents. So I don’t think those skills become useless, say, coding. To me it hasn’t changed so much.

I would say don’t bet on single white-collar skills. That, I think, is risky. Betting your career now only on translation, say, is risky because these things are very good at translation. But if you’re very good at translation, maybe there’s still a job there for you to verify what LLMs do or to reinforce the LLMs. There’s still going to be demand for some jobs.

In some sense what has changed is, yes, you should know how to use AI skills to enhance whatever you’re doing. But it’s unclear to me what are the exact skills that are going to be winners or losers. Whatever you learned, if it’s transferable, I think you can still find a lot of places where you can use it. It’s a lot about transferable skills and resilience. Transferable skills and resilience that we’re going to need.

All of the universities I’ve spoken to, I have some friends who are in uni right now, professor friends, the attitude towards AI is really mixed, but usually it’s not positive, right? There are a lot of integrity issues, a lot of skill development concerns that are extremely valid. How are you, or Cambridge, or whatever other university you’re aware of, preparing students for a vastly different market, where a lot of jobs are non-existent right now?

I think it’s very heterogeneous responses. There’s like an ignorance response of, “okay, we have to block this all out, students shouldn’t be allowed to use it, make sure everything is done without it,” and I can see how that can help in some cases in skill acquisition and learning outcomes, because if we all have super easy access to LLMs answering things, we don’t go through the hard process of thinking. It’s a bit like, if I take the car everywhere, my legs are getting weak. If I take the elevator up the stairs always, I don’t build muscle. And that’s how our brain is as well, if we always have the access to the easy answer, we don’t go through the hard, painful thought process. But that hard, painful thought process is what makes our neurons connect and remember things. So I can see how sometimes people, in that, “oh, we have to get it out of everything,” where they’re coming from. And I definitely see that value. We have to make sure students learn how to work in a group, how to get to the hard solution through thinking hard. It’s just so easy, you click it in, you get the response, but you don’t end up internalizing the knowledge.

At the same time, these tools are there. We should learn how to use them. People are later going to be in the job market, where you might be better off using them, so you should know how to use them. So those are two really hard things to bring together. On the one hand we still have to make life difficult and make us think, just to not become vegetables while a machine does everything. But at the same time, use these for knowledge enhancing. Everyone will say things like, “oh, it should be like a bicycle, you know, that you still ride,” but that’s a really hard equilibrium to find. And I need to work with it.

One way I sometimes try to do it is take it head on. I give them a task where they’re supposed to use it, but that makes them contrast the response of different algorithms and maybe figure out, why did one answer this, why did one answer that, why is the one code quicker than the other or something like that. But it’s hard.

But I think both camps are right. The ones who say we should take it and use it for everything, and those who say we should ban it for everything, I think both have a point. But in the end what we need is both.

If you could look 10 years out, what does the picture look like for universities? For entry-level job-market positions? What data is most worrying right now about workers? If you could wrap that together.

I think the good news for university is that students still like to go there for social reasons and credentials. LLMs might be good at teaching at some point, but students are still somehow better at coming to class. Now you can take MIT courses on YouTube, but not everyone has figured that stuff out, and people still build networks with each other at university. The consumption value of university, I think, is still there, and that’s what they have to survive on a bit, and the credentials as well. But not just in-class learning, audience interactions, peer networks, and so on.

The worrying thing about jobs, I think the data is not clear. Entry-level coders have gone down, but we can’t really separate it from the business cycle and so on. Some jobs will disappear, but it might also increase the entry to entrepreneurialism. I like to think of positive things sometimes; entrepreneurial people should, because the old businesses, they don’t have AI integrated in their workflows, and they’re not going to be able to do that. You can’t really take an old business and so much as shove AI in and that’s going to work. There’s a lot of room for entrepreneurs to build something from scratch, having these workflows in it. But not everyone wants to take that risk, and a lot of people want the security of a given job. But a lot of those are still there, teaching kids in school, that’s not going to go away.

Of course there are going to be some losers, who, I don’t think they are human losers, but they lose out because they’re learning something that right now is going to be less in demand in the labor market. That’s why we have to think more about transferable skills and resilience and adaptability. Also, what economists call non-cognitive skills, interacting with humans a lot. Maybe there’s going to be so much more value in trust, which in business is really important. People want to trust who they do a deal with, and that comes from communication and things like that. So one of those things cannot be replaced and will not be replaced. Knowledge generation, yes, you’re at risk and so on, but I think there will be demand for other things.

I don’t think the universities now have to go and try a

nd predict what will be exactly the skills that are at demand. It’s about thinking of education as a package, and I don’t necessarily mean values or morals, but I mean more like a package of skills that are transferable and resilient.