Part 1: How a ChatGPT Experiment in eLearning Came to Be

This is the first in a three-part series documenting a project I worked on with Jennifer Chien, integrating generative AI into an eLearning module to deliver personalized feedback. The full case study was later published in Training Industry.

A fascinating research case study began over a comforting hot pot meal with a friend.

Writing is not just a tool for me. It has been the engine behind almost every project I've taken on in recent years, so it felt natural to document this one as I went. By the end of this series, you'll see the whole arc: how the project started, how we built and tested it, and what 17 experts taught us about putting AI in front of learners on a sensitive topic.

But first, how it began.

A goal, and a hot pot dinner

A great portion of my childhood was spent on my bed, turning page after page, absorbing every word as fast as I could. Readers often become writers, and I did: silly stories as a kid, melancholic poetry as a teenager, research papers through college and grad school, and eventually published travel pieces in motorcycle magazines, including the print issues of Rider, which still feels special to hold.

My interest in motorcycles cooled. My interest in being published did not. So when I set my annual goals at work, I added one: submit an article for publication by the end of June, this time in my actual field, instructional design, instead of motorcycles.

I didn't make a plan. I just knew it was on the list.

Then I was having hot pot with my friend from grad school, Jennifer Chien, when she said she'd love to publish something together. She already had a topic in mind. She sent me a webinar by Garima Gupta on integrating ChatGPT into Articulate Storyline 360 to provide feedback on text a learner types into the module. When I opened the link, I laughed: I had registered for that exact webinar and missed it because of a conflict. The serendipity was hard to ignore.

I said yes.

The real question underneath the project

The idea was simple to state and hard to execute. Traditional eLearning feedback is pre-written. You select an answer from a list, and you receive a canned response. It works, but it cannot react to what a specific learner actually wrote. It cannot meet them where they are.

Could generative AI close that gap? Could it read a learner's own words and respond with feedback that is immediate, personalized, and genuinely useful?

We didn't just want to build the module. We wanted to test it. A formal academic study wasn't quite the right fit for either of us, since we both specialize in corporate work, but a data-driven case study was. We would define what success looked like and measure against it.

First, we needed a topic.

Why suicide prevention

At work, I had recently collaborated with a subject matter expert on a suicide prevention module for our annual safety and compliance training. The subject was already on my mind, and it raised exactly the question worth investigating: would a sensitive, high-stakes topic benefit from AI-generated feedback, or would the risk of the AI saying the wrong thing outweigh the value?

That tension is the whole point. Suicide prevention carries real consequences. If an AI tool gave feedback that was cold, inaccurate, or contrary to best practice, a learner trying to prevent death by suicide in a friend or family member could inadvertently do or say the wrong thing at exactly the wrong moment. If we could make AI feedback work safely here, the approach could work almost anywhere.

We later added a second, lower-stakes module on leadership styles as a comparison, so we could see whether AI-generated feedback only worked for low-risk content or could hold up for a sensitive topic too. More on that in Part 2.

Collaboration, done differently

I'll be honest: my history with collaboration, like a lot of overachievers, has not always been good. Most of us have had the group project we ended up doing alone.

This was not that. Early on, I asked Jennifer directly how we'd divide the work and what each of us owned. She made that easy to ask. We landed on a clear split: she would lead the technical integration, and I would own the scenario, the instructional design, and the writing of whatever we ultimately published. As the project took shape, we spent more and more time brainstorming and questioning each other, and it kept getting sharper because of it.

That's where it started. In Part 2, I'll get into the hard part: designing the study, building two versions of the module, and discovering that the prompt was going to be the most difficult thing of all.

Part 2: Designing the study and building the modules.