I built an artificially intelligent tutor! 🤖🎓 How’d I do?

My edu-geekery dial went to 11 last week when OpenAI announced custom GPTs. And here I’ve been talking about how you can use ChatGPT for learning! 🤓

I’ve been building an AI tutor ever since. Turns out it isn’t easy.

Here are a sampling of the things I wanted it to do:
🤔 Start by evaluating the learner’s knowledge through adaptive questioning (a necessity for effective tutoring)
🙋‍♀️ Use a Socratic Method-style approach to elicit knowledge while learning new information (to ensure an active learning experience)
🙊 Deliver new information in bite-size chunks rather than the long-winded answers we’re used to with ChatGPT (to avoid cognitive overload)
🧐Use checks for understanding to reevaluate the learner’s knowledge (this is essential for scaffolding new information)
🧠 Use spaced repetition to aid knowledge retention (to keep the information in long-term memory)
👩‍🏫 Base its process in educational research (because that’s just good teaching)

I started by describing this approach via the GPT builder’s chat interface. I uploaded research and explained how to incorporate it. (It was like that scene in The Matrix where Keanu goes “I know Kung Fu.” I fed in concepts, and the GPT instantly synthesized its new learning and described how it would incorporate it.)

But when I tested the result, the GPT often didn’t follow its instructions!
I’d tell it never to lecture, but rather to ask a series of adaptive questions to gauge understanding and then deliver information in digestible chunks. It would respond in its own words (typically a good way to evaluate a learner’s understanding!) that it would ask questions to evaluate, then alternate between delivery of bite-size information and checks for understanding. Great answer!

But then it would just spit out a typical ChatGPT-esque long-form answer about the topic rather than evaluate my understanding. 😤

I tried a combination of approaches to improve it: from engaging in questioning about how it could improve its answers (channeling the Socratic Method and Stacey Moore‘s Leadership Prompting approach) to telling it specifically how to behave. It would improve… and then it would slip back to its old methods. I found myself wondering how much priority the instructions of the custom GPT get vs. the underlying ChatGPT algorithm.

But you can teach an old AI new tricks! My custom GPT works as intended… sometimes! 😅

So: hello world! Meet MyAI Learning Guide. If you have a ChatGPT Plus account you can try it right now and if not you can watch my video demo!

Feedback, please! What is it doing well? Are there specific improvements or research to incorporate? And what are your thoughts about AI-guided learning?