Generative AI in Education

A few years ago, when GPT-4 was still the frontier of generative AI, I got an early look at what this technology could mean for education.

During my senior year at Indiana University, Phil Jordan and I presented at IU's Statewide IT Conference about the opportunities—and problems—we were already seeing with generative AI in education.

Phil approached it from the instructor's perspective. I approached it from the student's.

At the time, much of the conversation around ChatGPT on campus was understandably about cheating. Students could give a model their homework and ask it for the answer.

I had started using it almost the opposite way.

In one of my Python courses, I gave ChatGPT the table of contents from my textbook and had it generate practice problems around the concepts we were learning.

I'd solve them myself, give my answers back to the model, get feedback, and keep practicing.

I remember doing this for about an hour before class one day.

Then I went to class and, for the first time in a while, I actually understood what was going on.

I was following the instructor. I was raising my hand. Concepts I'd been struggling with finally clicked.

It showed me something that has stuck with me ever since:

AI didn't have to replace the work. It could give me more opportunities to do the work.

But there was another part of the presentation I'd almost forgotten about: **the equity gap.**

The free models available at the time were noticeably less capable for the kind of personalized practice I was doing. GPT-4 produced better exercises, understood my responses better, and gave me more useful feedback—but accessing it required a paid ChatGPT Plus subscription.

That's what the slide in this photo was about.

If AI could meaningfully improve the way a student learned, but the best versions were only available to students who could afford them, then we weren't necessarily leveling the playing field. We could create another advantage for students with more resources.

Looking back now, even GPT-4 and the models we were comparing it against belong to an earlier generation of this technology. The capabilities have moved enormously since then.

But I don't think the questions we were asking have become antiquated with the models.

How do we teach students to use AI to become better at something rather than use it to avoid learning it?

And if these tools genuinely improve someone's ability to learn, how do we make sure access to the best learning tools isn't determined by who can afford them?

The models changed.

Those questions didn't.