Peter Nunez has just finished his undergraduate degree in economics at UCL and is spending the summer building an AI tutor for students of CORE’s The Economy 2.0.
We caught up with Peter to find out more about the project, and how the tutor differs from standard LLMs.
Hi Peter, thanks for joining us. Could you tell us a bit about your academic background?
I’ve just finished my undergraduate degree in economics at UCL, and I’m heading to LSE next year for an MSc in Econometrics and Mathematical Economics. I’m currently working as a research assistant at UCL. I’m interested in applied microeconomics, particularly labour economics. My undergraduate dissertation was on education and migration.
There’s no big story behind why I got into this. It’s more that this is the world I see every day, and learning economics helps me make sense of it. By second year I knew I wanted to pursue research.
What was your history with CORE as a student?
I was taught with The Economy 1.0 for my introductory economics course, and that was my first time doing economics. I hadn’t taken it at A-level. My professors often said that not having A-level economics as an alternative benchmark was a good thing, and I agree: I really enjoyed first year. The Economy 1.0 was introductory economics to me, without any other frame of reference.
Where did the idea for the AI tutor come from?
CORE director Wendy Carlin had the idea. The starting point was that AI is now essentially omnipresent in education. Everyone is using it. You can even see ChatGPT usage drop during the summer holidays. But current AI use in introductory economics isn’t optimal.
If you talk to an LLM about intro econ, it defaults to one of the standard introductory textbooks. Wendy’s example is that it always defaults to market theory. That’s not what CORE teaches, or what we want to teach. The goal was to build an AI tutor that’s grounded in CORE’s content, and that behaves in a way that genuinely helps students learn.
What does helping students learn mean in practice?
Used incorrectly, AI in education can give you an answer without actually teaching anything, and can make you feel like you’re learning when you’re not, because it’s doing the thinking for you. We wanted to build something that behaves like a tutor, not just an agent that hands over answers.
We modelled the pedagogical behaviour on existing study-and-learn agents, which are already well documented. The most immediate thing you notice is that our tutor gives short, focused answers. It is not going to spit out a two-page response when you ask a question. When you ask a standard LLM something, you get a wall of text, and it is easy to skim and think you have understood it without processing much. A real tutor wouldn’t do that. They would try to understand your confusion, push you in the right direction, and guide you to the understanding you need. This is what we want the AI tutor to do.
How have you built your tutor?
We built the tutor on the built-in agent builder in Copilot, with equivalent versions planned in ChatGPT and Gemini. The idea is simple. You give the agent a system prompt and sources, and it behaves accordingly. At its core, this is really a prompt engineering task. I was a bit worried about that at first, but it has been surprisingly effective, partly because study-and-learn agents are already a well- documented model to build on.
What are the biggest challenges?
The main challenge is optimisation. There is a token limit on the system prompt, around 8,000 characters, and fitting all the behaviour you want into that limit is the real work. My first draft, written by hand, got the tutor to exhibit one of the behaviours I wanted, and that alone used up the entire system prompt. There were many more behaviours I still needed to add. Fitting all of them in has been the real struggle.
Who is the tutor for, and how do you make sure they can access it?
The tutor is for undergraduates studying The Economy 2.0. The equity angle was one of the goals from the start. Copilot is free for anyone with an educational Microsoft account, which most students at most universities have. That is a lot cheaper than the £20 a month a student would pay for ChatGPT Plus. Not everyone can, or wants to, pay that.
Deployment is a little tricky. Right now I can only distribute the agent directly to people at UCL. But if we send other universities the system prompt, the source links, and everything else needed to set up the agent, someone at each university can build it themselves and share it with their students. On the CORE website, we are imagining a drop-down menu where you pick your university and are then sent to the right version of the agent.
Where is the project at now, and what’s next?
The first complete version of the tutor is done. Next week we start testing, and we’re figuring out how to integrate it into UCL’s introductory economics module, ECON0002. In testing we’ll get qualitative feedback on how students feel about the tutor, and we’ll see transcripts of their conversations with it: how they prompt it, what they use it for. That’ll tell us what to change before the next academic year starts, and how to advise students on the best way to use it.
We are working out the specific tasks and activities that will show students the benefits of the tutor. The target is to have everything ready for students in October, when the new academic year begins.
Why would a student use this rather than just ask ChatGPT?
That was actually the first question I asked Wendy in our initial meeting. My concern was that frontier LLMs like ChatGPT and Claude are going to be technically smarter than what we can build. But those models default to the standard introductory economics content, and if they are doing that, they are not giving the answer CORE wants.
The big sell for our tutor is improved learning gain, and there are two dimensions to that. The first is CORE grounding: our tutor is specific to CORE content, so it doesn’t default to market theory or other standard introductory content, and it’s going to be right more often for a student studying The Economy 2.0. That especially matters for the macro material, which diverges from the standard introductory curriculum more than the micro material does. Micro diverges in meaningful ways too, but the gap is largest in macro. Concepts like wage-setting and price-setting are central to CORE but aren’t universal in introductory economics.
The second dimension is the tutoring behaviour: our tutor asks the student questions rather than handing over the answer straight away. For example, when I asked the tutor why the optimum in a constrained choice problem is where MRS equals MRT, it gave a small piece of information and immediately started asking questions back. My concern is that students might find that kind of exchange frustrating rather than helpful, and the open question that testing is designed to answer is whether the combined value of grounding and tutoring behaviour is enough to overcome that.
How do you plan to convince students it is worth using?
We have a few ideas. The obvious one is demonstration and comparison: show students an answer from a frontier LLM like ChatGPT or Claude next to a conversation with our agent, and let them see which gave the better response. We’re also considering an in-class task where students do one exercise with ChatGPT or Claude and one with our tutor, so they can decide for themselves whether it works for them.
There is also just a familiarity hurdle. Copilot is less familiar than ChatGPT and other tools for most students. I never had it open on my laptop until I started this project. That will settle over time. It all comes back to the same question, though: is the increased quality of the answer worth jumping over a few barriers? That is what we need testing to tell us.
Thanks for your time, Peter. Good luck with testing.
