AI and not letting real life get in the way of (learning) theory: Part 1

It's often said that technology helps in one of two ways: Either it makes it quicker to do things we were already doing, freeing up time and headspace, or it allows us to do things we simply couldn’t do before. However, I've found that lots of uses seem to sit somewhere in the middle of those two. Things we could, in theory, always have done, but in practice, it just took so long that it often wasn’t worth it. One such example is high quality comparisons.

We use comparisons in teaching all the time. They’re one of the easiest ways of making abstract ideas feel more concrete, especially when students can see clearly what’s changed and what hasn’t. When I was a maths teacher, this was relatively easy. You can put two equations side by side that are identical except for one term, and students can immediately see what that one change does. In science, the logic is similar: change one variable, hold everything else constant, and observe the effect. Economics, however, has a small complication:

Reality.

Economic phenomena live in messy, noisy contexts. When we try to use real‑world examples, we often end up changing lots of things at once, whether we intend to or not. That makes it harder for students to see the specific mechanism we’re trying to teach.

This came up very clearly for me this week while teaching development economics, specifically the Harrod-Domar model. I was teaching a mixed Year 12/Year 13 class and I wanted to make a fairly small point within a bigger lesson.

When we talk about high capital-output ratios in developing countries, one explanation is that capital can simply be harder to use productively. Geography and landscape can reduce the choice of suitable capital, increase the cost of installing it, and raise running costs as well. Some machines just won’t work well in certain environments. Others can work, but only if you first build a lot of infrastructure around them.

I wanted to show this visually. In my head, I had two images. One of a developing country landscape where it would be genuinely difficult to whack a factory down. And one of a developed‑country industrial site with a nice patch of earth, decent roads nearby, and reliable electricity.

I could, in principle, have searched for those images. But finding a pair where the differences lined up neatly with what I wanted students to notice would have taken a long time. And even then, there’s a good chance the images would have contained lots of distracting detail - things students fixate on that aren’t actually the point I wanted to make.

As a quick example, take a look at the two images below. Why is more ice cream being sold in the first picture than the second? 

(Scroll down!)

I was intending to show that ice cream shops are busier in summer, but you could have instead thought about vans vs parlours, seaside vs park, well maintained vs a bit shabby etc.*

Back to my capital:output ratios. I was able to avoid all of this misdirection by asking AI to create two contrasting images with clear specs so that students almost couldn’t help but focus on the relevant differences.

The ease of building in one place versus the difficulty in the other was immediately obvious. For what was, in the grand scheme of the lesson, a fairly minor point, this suddenly felt worth doing.

That said, I do have some hesitations, because the place I’d generated doesn’t actually exist.

I’m generally very relaxed about this sort of thing elsewhere in Economics. When I teach price elasticity of demand, I pick nice numbers. When I teach the theory of the firm, I often reverse‑engineer figures so the cost and revenue curves behave themselves and come out in pleasing, intelligible shapes. Real firms’ cost curves don’t look exactly like that, and I’m comfortable with that compromise. I’m teaching a model of reality, not reality itself, and if the model helps students understand how firms behave in general, it’s doing its job.

Here though, it feels different.

Rather than converging on something like an average of real‑world situations, here I’m deliberately exaggerating the contrast.

That creates two potential issues.

The first is about nuance. Development economics is an area area where flattening complexity can be particularly problematic. Countries are internally diverse, conditions vary hugely within regions, and labels like “developed” and “developing” already do a lot of heavy lifting. When we simplify too aggressively, there’s a risk that students walk away with a picture of places as uniformly constrained or uniformly advantaged, rather than shaped by a messy mix of geography, institutions, history and policy. That matters, not just for technical accuracy, but because development is so often taught and discussed through moral and political lenses as well.

The second issue is that I’m helping students understand a world that doesn’t actually exist. That gives me pause, but I think I’m mostly comfortable with it? The message I’m trying to teach does exist. I'm using a world that doesn't exist to help students to understand the world that actually does exist. I’m not really claiming that this is what a developing country looks like, I’m showing why you can’t just whack a factory anywhere and expect it to be equally productive as another place. That mechanism is real, even if the image is a constructed exaggeration designed to make it visible, and if it was a cartoon rather than a photo, maybe I wouldn't have even considered all of this.

In that sense, this isn’t so different from other parts of Economics. We simplify demand curves, tidy up cost structures, and choose numbers that behave nicely, all in service of helping students understand a general principle. The difference here is that we are moving that one variable as far from reality as possible. So, while it's useful, it's still worth handling with care.

This is probably where there’s something I need to learn from our colleagues in Geography. They spend a lot of time showing how places differ, while also being explicit that no place ever perfectly fits a checklist of characteristics. Perhaps the answer isn’t to avoid these kinds of illustrative examples, but to frame them carefully, and to return to nuance once the core idea has done its work.

I don’t think this means we shouldn’t use these tools, but we need to be clear about why we’re using them, what we want students to attend to, and what we’re deliberately leaving out - at least for now.

In Part 2, I want to look at the other side of this. Not the ways AI can strip reality away, but the ways it can actually help us make comparisons that are more grounded in real life, not less.

(The irony here is that I actually did use AI to create these images so I didn't have to faff around finding royalty free ones, but you get the gist: This could have been what I would have found doing a google search for 'ice cream seller on hot(/cold) day.')