Reader, I’ll be honest. When our Chief of Assessment, Alina von Davier, first suggested that I start using an artificial intelligence tool to write for the Duolingo Blog, I was surprised—and maybe even a little insulted!

As a writer, I felt no need for this new technology, and what’s more, I had serious concerns about the ramifications it could have in schools, and in the workforce. If we outsource all of our content generation to machines, will there ever be another Shakespeare? And what about me, and all the writing I’m doing here… what will happen to my work?

Sophie Wodzak, author of this post and hardworking Duo, sits at her laptop computer in a stylish brown turtleneck working on her next great blog post

Who’s afraid of ChatGPT?

Feeling threatened and insecure, I did what anyone would do. I resisted. Not only did I make no effort to explore GPT, for weeks I rolled my eyes at the mere mention of it. And believe me, here at Duolingo, it’s mentioned a lot — especially amongst my colleagues working on the Duolingo English Test.

What is GPT, exactly? Simply put, Generative Pre-trained Transformer is a language model that can produce text in almost any genre you can think of. It’s able to do this because it’s been trained on a massive amount of text data (we’re talking billions of sentences) so that it’s learned patterns and relationships to be able to write about basically any topic that’s on the internet.

There’s no denying embracing new, innovative technology has been very useful at Duolingo. From building language course content to developing items for the Duolingo English Test, AI has been instrumental in scaling our products for millions of users and test takers worldwide for years. So yes, I knew about its potential to enhance our work.

But writing blogs was something different, right? Surely this technology could never replicate my human expertise, my unique, nuanced perspective… could it?

Time for a quick chat

Finally, my curiosity got the better of me. Bracing myself, I opened a ChatGPT window and asked it to write an article about the ways we use AI at Duolingo. The text unfolded across the screen, and my heart sank—within seconds, it had generated four paragraphs that did a pretty good job of explaining Duolingo's human-in-the-loop approach to leveraging AI in our work.

A text screenshot that shows Sophie's prompt to GPT-3 and its response. Sophie asked: "Write a duolingo blog about how we use human in the loop AI to scale our test and our courses." GPT-3 responded (first paragraph): "At Duolingo, we use a combination of machine learning and human input to create and scale our language courses. Our approach, known as "human-in-the-loop AI," allows us to efficiently create and test new content while also ensuring that our courses are accurate and effective."

Pretty good—but not perfect. As I read through the draft, I realized it was too general and a little too short. For about ten minutes, I went back and forth with the chatbot. The more specific I was with my requests, the closer it got to giving me the article I had in mind, though it wasn't totally ready to publish.

But by then, I had a great starting point. As a writer, there’s nothing I find more daunting than a blank page. With some text already written, I was able to dive into the editorial process, deleting a sentence here, tweaking a paragraph there, and adding key information the bot had failed to include in its draft. In the end, it didn’t do the work for me, but it really accelerated the process, saving me hours of time.

A brave new world

I’m happy to report that my aversion to this technology has been replaced with relief, and even excitement! For years we’ve been using generative technology to solve some really big problems, like scaling the DET and making it more accessible to test takers worldwide. It’s really interesting to see it work effectively on this smaller scale, and we’re only starting to scratch the surface of what it can do.

But I also know that (at least for now) it can’t do it without us. I’ve learned that while it will write for me, it won’t think for me—that’s still up to me. From conceiving of the piece, to writing the prompt and interacting with the generated content, I was more than just a part of the loop, I was in control of it. Rather than making me redundant, this tech turbo-boosted my composing process. And guess what…

…it even helped me write this blog!