Teaching Statistics in the Age of ChatGPT

One of my favorite websites to use while teaching statistics is Seeing Theory, by Brown University. It is a wonderful website, because it allows people to “see” statistics.

Visualize concepts in statistics, to use the technically correct term, but you see what I mean.

One of the many reasons I like this website is because it presents a fun, interactive way to “get” what statistics is all about. It is one thing to talk about flipping a coin, it is quite another to actually flip a coin 1000 times. Or roll a dice, or understand what a probability distribution is, or to (finally!) “get” what the Central Limit Theorem is trying to get at (beware, though – every time you think you’ve “got” the CLT it has a way of revealing an additional layer of intrigue).

This past summer, as I’ve mentioned before, I was teaching school-going students courses in economics, statistics and public policy. I have made use of Seeing Theory in the past, but with the advent of ChatGPT (and especially ChatGPT4), I figured it might be a good time to not just show “cool” visualizations, but also actually try and build them.

And so we did! As we covered a topic, I would ask my students to “build” a working demo using ChatGPT (or Bard). I would nudge and prompt the students to well, write better prompts, and if necessary, step in and write the prompts myself on occasion. But for the most part, the work was done by the students, and we were able to get simple working demos of some stats concepts out of the door.

The “whoa, this is so cool!” moments were worth it in and of themselves, but it is my ardent hope that the students understood the concepts a little bit better for having seen the visualizations.

A great example is the Monty Hall problem. Run a simple Google search for it, if you haven’t come across it before. In my experience, some students tend to not “get” the explanation the first time around. Until this summer, I would get around this problem by asking them “what if it was a million doors instead?”, or if all else failed, by actually “playing” the game using three cards from a deck of cards.

But this time, we built a demo of the problem! So also for Chebyshev’s inequality, the expected value upon rolling a pair of dice and a simple way to visualize what regression does. The demos won’t satisfy professors of statistics or professional coders, for you could add so much more – but for young students who were trying to internalize the key concepts in statistics, it was pure magic.

And the meta lesson, of course, was that they should try and do this for everything! Why stop at stats? Build working demos for concepts in math, in physics, in geography. And if you know even a little bit of coding, try and build even better demos – both I and my students were relatively unfamiliar with coding in general, so we stuck with simple HTML.

But with AI’s new coding capabilities, it is clear that teaching (and learning) can become much better than was the case thus far. If you wish to disagree with me about the word “better”, I look forward to the argument, and you may well end up having more than a couple of points. But the classes were certainly more interactive – and at least along that one dimension, they were certainly better.

I hope to do much more of this in the months and years to come, but for the moment, do try out some of these demos, and let me know how they could be made better.

Thank you!