This series is on AI’s disruption of entry-level work and what universities should be doing about it, or should think about it. Give us a quick picture of what you have been doing at CORE right now, and how that ties to that question.

I think there are a number of ways into this, and it’s probably good if we step through them.

What we did at CORE over the first decade and a bit was to create a new way of teaching economics at the starting level for students. The reason we did that is because the way it was being taught didn’t meet the needs of either students or employers. I don’t know if you did an introductory course, but the traditional way of teaching was to just have some models and teach models and then often provide trivial applications of those models.

What students coming into the classroom really want is a set of tools that help them understand how the world works. So, I look out the window, and I want to be able to have a sense of how I can explain what’s going on.

That was our whole motivation, and the learning model that we used was actually the way that one of our colleagues told us that people learned to play the tabla, Indian drums, where people start from the complex coda at the end of a piece. Beginning students start by learning the complicated but really rewarding final passage. And they practice and practice and practice that. And then they move a few bars closer to the beginning and they learn those passages, and then they can go back towards the end, and they get a great sense of satisfaction.

We wanted to bring that into the classroom and say, let’s start with a complicated question that you actually care about. It will be too complicated to begin with, but that’s where we’re going to start. We’re going to work back from that and think about what kind of data is a good way of conceptualizing in quantitative terms what we are interested in. And then we say, what kind of models from economics shed light on that question? And then each time we go back to that complicated question with your new powers, your new superpowers from economics, you get that same sense of satisfaction.

That’s how we approached bringing together new content with new pedagogy in the CORE project. And we brought a whole community of researchers on board because researchers are actually working on the problems that we face. It’s just that it often hadn’t crossed their minds that their research could be brought into the introductory classroom.

At that time we were also responding to dissatisfaction of employers who said, okay, we hire these economics majors because they have numeracy skills, they’re good quantitatively, but there are a lot of things about what they’re like when we get them that are not terribly impressive. So, for example, they can’t apply their models to a new context. They’re most comfortable with something that is very abstract. They don’t work very well in teams, they don’t show much empathy. They can’t reason out of the box.

Employers were already saying that more than a decade ago, but nevertheless, economics majors got the premium because of their quantitative skills. When we started thinking about the challenge from AI, it was very reminiscent of the complaints employers had all along, which we were responding to in CORE, but this time there’s a competitor, a direct competitor for the graduate, who can do the analytical, quantitative, algorithmic problem-solving better than the graduate economist.

The question then was, how good is the AI at the other things? Another complaint of employers was that many economics graduates were pretty bad at writing. The 2026 AI is definitely better at writing than the typical economics graduate. And we don’t really know about whether it’s any good at teamwork. We presume it doesn’t have much true empathy. It is certainly patient.

This really focused our attention in the project. And we said, OK, we’ve got the goods in terms of a strong pedagogy and strong content, which is going to become super important for a reason I hope will become clear. We’ve really invested over the last decade and more in creating curated, high quality, pedagogical material from the best research there is, and making that available to undergraduates. And it’s this point about curated, selective, research-informed material that’s pedagogically sound, in other words, it can be taught and understood by undergraduates, that stands in contrast to just getting an answer from an LLM.

This crystallizes thinking about what is the role of the university in the era of AI. Part of the role of the university is to take the best of research from the world and curate it and make it available to students. And that is where we can out-compete the LLMs, because the LLMs will give you, as we all know, tremendously plausible answers to your questions, but they will often not give you the right answers or the correct reasoning.

In economics, that’s particularly interesting because of the training material. The training material in economics is multiple decades dominated by the main textbooks, popular books, economics journalism and the blogosphere. The research literature is a small piece. There’s a whole load of economics out there that will very rapidly default to a simple model of market clearing (how the world works) and that markets work well (so intervention is bad), rather than the more sophisticated view that comes from contemporary economics and that CORE reflects, which is that understanding how markets work is essential, but that many of the most important markets cannot be characterized by the supply and demand model with market clearing.

If a student just goes to the LLM to get their answer, they may be misled by the very simplistic economics which dominates the training corpus. It’s what I call “econ slop,” and there’s a lot of it out there.

I was teaching from January a class of 800 first-year undergraduates. And I also had 20 teaching assistants working on this class. This was a really interesting window on how rapidly things are changing, and pinpointed some of the key problems that students are manifesting.

One of the most important problems is that many of them do not know how to read independently. They rely on getting answers, getting summaries of material, and they have lost confidence and practice in their ability to read critically and independently.

If we think about the importance of reading in the evolution of human knowledge, reading was a really major step. But that ability to read, to reflect, to exert the effort to understand what you’re reading, that’s really hard. It’s often frustrating in economics. It’s not always that easy to understand how a model works and what role is being played by the assumptions. But if you’ve forgotten, or you never really got the idea that you have to exert some effort to learn and that you will feel confused, and that feeling confused and uncertain is absolutely at the heart of the true learning process.

What was happening, I think, is that students have got into a pattern where they weren’t reading, they often weren’t reading for pleasure, and they weren’t reading sustained pieces of text as part of their university study.

So one of our big projects is to create a tutor that will help students regain their skills at reading. It might sound very basic, but I think it’s absolutely fundamental. A student came to me, for example, saying that he was worried he’d forgotten how to write independently and asked, what should I do? And I said, well, maybe you should read a bit more. Read a novel. And he said, oh, well, I haven’t read a novel since Harry Potter.

I think it’s very important to encourage students to be reflective, and to worry about their own capabilities. And these are very, very good students. These are outstanding students. So there’s no question that they have the capabilities. They have the capabilities, and they have sufficient introspection to be worried.

In the “Has AI Eaten the Economics Major?” episode at the Centre for Economic Policy Research, you described AI as a cognitive sparring partner for students. You’ve mentioned the tutor that helps them read, you’ve mentioned issues coming from LLMs. Biologically thinking about it, it makes sense, we’re made to conserve energy. If you can just ask ChatGPT to do it for you, it’s a no-brainer. But what does that actually look like, the cognitive sparring partner? What does that look like in a CORE classroom?

One of the projects is the one I mentioned, which is creating the CORE reading tutor, which will literally serve students chunks of text and then in an encouraging way, allow them to ask questions. And if they ask a question that’s out of scope, they’re told, no, you have to stick to the text. So if they ask who James Bond is or something, then they’re not going to get an answer. And to move on to the next chunk of text, they have to demonstrate that they have actually internalized the material.

That’s one thing. The other thing that we’re doing that I hope touches on what you’ve asked is that, as in many classes that have large numbers of students, we use multiple choice questions. CORE has a really big test bank. One thing we did was to put all of our, maybe 3,000, individual questions, into a whole range of LLMs and see how good they were at getting the correct answers. There was a wide range of results.

Now, that’s beginning to give you a way of helping students learn and to understand the way LLMs can both help and hinder their learning. We use the jagged frontier model of Ethan Mollick. LLMs are good at some things and not good at other things. So what we can do is map this region where LLMs are good at certain tasks and answering certain types of questions.

Using our test bank, the calculation questions, all of the LLMs would get the correct answer. So if students are having difficulty answering those kinds of questions, then using an LLM as a Socratic tutor can help them learn those techniques. The kinds of finance type of problems, where you just have to solve net present value problems, for example, are algorithmic. So LLMs can come on board as a helpful tutor.

And that’s really great because, if I’ve got 800 students, I don’t want to waste time in the classroom on those things, because many of the students can just do them, and they’ll be bored. And some of the students will take ages to get the hang of it, and they won’t get the hang of it in a lecture and will be frustrated. That’s where the material is inside the frontier and LLMs can boost student learning.

Then there are some questions that some of the LLMs get wrong. These are ones that require rather subtle reasoning going beyond simple market clearing models, for example. And what we’re doing is to help the students by showing them the answers, the wrong answers from the LLMs, and the wrong reasoning from the LLMs, to help them understand this jagged frontier, and the fact that to understand some material, they cannot rely on the LLM.

But how do they figure out what LLMs are good at and what they’re not? This is what we are working on at CORE. With questions that are outside the boundary, we will divert students back to the CORE textbook (free online at www.core-econ.org), and, for example, to the reading tutor, which will help them go through that specific material.

These are the sorts of tools that we’re trying to build. I hope students will learn that answers are cheap. Good questions are scarce.

Students will have to learn to figure out whether their questions are the kind where they can get straightforward help from an LLM (practice at solving routine optimization problems, for example), or whether they’re the questions which, to answer, you are going to need to exert your own effort to get your brain around it. They’re the frustrating questions, but the ones where getting to the answer builds the scaffolding for the next step of deeper learning.

One example, which I used first with my teaching assistants in the weekly training session, and then with the students, was like this. I gave them a simple question: in this context, does a curve shift up or down, and why? By the way, I brought paper and pens into the classroom. And they were kind of curious, what’s this about?

The first question was quite straightforward. They had to say just this thing, the curve shifts down, and they had to give a brief explanation. There was a lot of reluctance to write anything. The TAs, all very smart PhD students, initially wrote in very small writing with reluctance. We worked through that one, did another one, and they became more forthcoming. The third question was in the same format, but next to the answer of whether the curve shifted up or down, it said “incorrect,” and then it had six lines of text giving an explanation for this, and it said next to the explanation “incorrect” and that it came from an LLM. The job of the TA in the training session (and the students later on) was, what’s wrong with this explanation?

It was very striking that of the group of 20, some of them turned over the paper and covered up the answer, because they’d already figured out that if they were going to work out what was wrong with the answer, they’d better figure it out for themselves and not be lulled into going along with the likely smooth, beautiful answer from the LLM, which they knew was incorrect. So they covered that up, concentrated on the question, and quickly found out the mistake in the answer. The ones that looked at the answer first found it much harder to figure out what was wrong with it.

And that’s a lesson to the students, that just getting an answer cheaply is going to sometimes make it harder for you to learn. Because it’s so plausible.

This is, I think, a way of illustrating how you can use AI very constructively. You need to redesign the way you teach, the way you interact with students, to create this cognitive sparring partner, which can give students training in the exercises like the algorithmic ones, but where we need to help them is to learn to push back when the LLM gives beautiful fluent answers to questions that are testing higher levels of understanding. Another CORE project is to create a learning environment in which only CORE’s carefully curated research-informed material is used to generate questions and guide students to develop and practice interpretation and reasoning skills.

CORE has been in 500-plus classrooms around the world for over a decade. What part of it has aged the best in this AI moment? And what part are you rewriting now?

Well, we never stop. We published a new flagship text in two volumes, The Economy 2.0, just last year, reflecting criticism from students and from our community of teachers and researchers as well as new models we have developed to get closer to both the world and the research frontier. So we’re always improving things and figuring out how to do things better. I think CORE’s model of knowledge creation and community building has aged well.

We have two other projects, two other big parts of CORE, which you may not have looked at. One is called “Doing Economics,” and this is pretty AI-proof. It’s a way of getting students handling real data and figuring out how to do data visualization and data interpretation, starting from a position where they don’t have any confidence, for example, with a spreadsheet or with using R or Python to code.

They can choose which of those tools to use. But the AI doesn’t really help you, because we provide the code if you’re using R or Python, and we provide walkthroughs to use with Excel. You have to go through the steps yourself to progress through the project.

We run that exercise with our students at UCL, and it takes them a lot of time, but then they feel a sense of satisfaction that they’ve really learned some of these basic skills, which almost every graduate economist has to be good at, data visualization and data interpretation. They have to write up a short piece, a bit like a short leader piece in the FT, using what they’ve learned from their own data work in the Doing Economics project. That’s golden. They have to sweat to do it. Or at least it would be just as much work to get the AI to do it. So it’s not going to be worth it.

Another part of what the CORE project does, which again I think stands the test of time really well, is what we call “Experiencing Economics,” another ebook. This is a set of classroom games and experiments. We want students to feel what it’s like to be put into a whole series of economic situations where they’re the actor and they have to make decisions.

For example, we play a public goods game where there’s a question of making a contribution to the public good, and standard old Econ 101 would tell you that no one would ever contribute because in their own self-interest they would always choose to free-ride. And we find, consistent with the research, that in every classroom where you try this, you get the result that people contribute to the public good. And then the question is, why? And the students have got to think about that.

When they have to decide how much to contribute in the game, there’s no way that AI is helpful. They have to figure it out, just decide what they’re going to put in. And then you play several rounds and the contributions to the public good go down. So that’s another great question to them: why is that? Why, when we played this for the ninth time, did you actually contribute less? And then you introduce punishment for free-riding into the game and the contributions go up.

Just like Doing Economics, CORE’s Experiencing Economics is a great way of helping students understand some really central concepts in economics. AI doesn’t help them with that. Although it may help us to create versions of this game that can be played asynchronously.

So these are pedagogical choices that we’ve made that really survive the presence of AI. I’m pretty confident that we’ve taken good decisions. And that we’re also very much sensitized to how we need to meet the challenges, especially the ones I’ve talked about: how to read, how to write, how to think.

Thinking takes effort. And as you said, we’re kind of programmed to take the path of least resistance. Most of them go to the gym and they realize, okay, I’m doing this and it’s an effort, but I get better at the challenges I face there, and may even start enjoying them.

The great thing about the CORE project is we have teachers in university classrooms all around the world. The person who’s designing the reading tutor, Tamoghna Halder, is based in Bangalore at the Azim Premji University.

That’s a very good university, and it puts a huge amount of effort into recruiting very, very talented students with huge potential who have big gaps in their school education. So the AI reading tutor is also a method of opening up opportunities to very smart people. They’re going to be drawing some people into leadership positions who would never have had a chance without this technology.

On Marketplace in February, you asked why economics majors got a labor market premium, and how that premium is being undermined. Your Hong Kong colleagues said consultancy and finance there are now hiring philosophy and history majors over econ. So what did the major actually certify, and what would you teach undergraduates now to defend it?

I think we’ve covered this. Essentially the major teaches quantitative analysis and algorithmic problem-solving skills. Building spreadsheet models and coding up data analysis are examples of skills where econ graduates would out-compete other social science graduates and get a premium. That’s now done, we know, better by the AI. On the other hand, the reading, comprehension, and critical thinking capabilities are often stronger in students trained in good philosophy and history programs. So that’s where their comparative advantage is coming to the fore.

And it’s all handled in Bloom’s taxonomy, if you look at Bloom’s taxonomy of the different levels in the hierarchy of knowledge. So that’s the touchstone for thinking about how these different capabilities are categorized.

I feel like it’s such a transformational time right now for universities. I studied history, and we were given large amounts of books to read in a very short amount of time. I would be like, I can’t read all of these. It’s a week and it’s five books. And they were like, well, your job is to find important information in them and draw conclusions. I could do that in 20 seconds now. In Political Quarterly last month, you said, answers are really cheap, but what’s scarce are good questions. In that world, what does an economics degree actually certify in 2035? Or does the credential not survive in its current shape?

So 2035 is kind of a challenge. And I think it’s a very good question. I’m very confident, actually, that the CORE material, because we work on it all the time, along with the pedagogy and the organizing principle of teaching economics as an integrated way of understanding the world, will continue to provide value for students.

Economics’ secret sauce, its USP, is being able to provide causal interpretations of what is happening. And to take into account what we call the general equilibrium effects of policies and of shocks. How are people going to respond to a well-meaning policy change? A politician might think, okay, I can reduce inequality by increasing taxes on the rich. Well, that might be true. But what if the rich just hire more tax lawyers and evade the taxes?

Questions of understanding human behavior and strategic interactions is at the heart of CORE’s actor-centered way of understanding the economy. That will survive as a valuable and rewarding education when combined with the ability to interpret data and to tease out causal effects, to better understand which kinds of policies work.

Teamwork and empathy are qualities that are going to become increasingly important for graduates. And traditional university teaching has been quite bad at that. Business schools do better, I think, in getting students working together.

As an example of what can be done, in week one, when students arrive at UCL in their economics induction session, they’re each given a GPS location, and at the end of the session, they have to make their way around London to find their group at the GPS location. Having found each other, the students have to work as a group within a week to create a video linking that location to the first part of the CORE textbook (hopefully encouraging them to read as well as to work together).

We’ll be doing more activities like this by 2035. But the other final aspect of that is that current narrow disciplinary boundaries, I think, will have to become more porous. Students will have to know more history, they’ll have to know more science, they’ll have to know more psychology. And economists really rely on those things. History, psychology, political science, sociology, we use them. But we don’t really acknowledge it. And we don’t help our students learn to be part of multi-disciplinary teams, which will include AI as a key resource.

I think that is going to be absolutely central, and that universities will have to become places where we take these really, really smart young people, put them together, probably for doing exercises, at least one of which they’ll have no access to computers and they’ll have some paper to solve their problem. Building teams and building trust and helping students to learn to communicate across cultural barriers and so on. These capabilities are going to become more and more important as we move towards 2035.

*This interview has been lightly edited for clarity and readability. The interviewee reviewed and approved the transcript before publication.