More Adventures in AI Writing Coaches

Earlier, I wrote about using ChatGPT to perform the “because/but/so” exercise from the book The Writing Revolution. While it took me two tries to get the prompt correct, and there were some unexpected responses, I’m still amazed that I’m able to get a computer to walk someone through a writing exercise at all.

Recently, I’ve been trying to see if I can get AI to help with another exercise in The Writing Revolution. This exercise is about teaching students how to revise. Because it’s hard to revise your own work, the teacher starts the exercise by giving the students an “unadorned paragraph.” Here’s an example from the book:

Eleanor Roosevelt was important. She was married to Franklin Roosevent. She was the First Lady from 1933 to 1945. She made many contributions. Many people admired her.

This is all true, but boy is it vague and dull! The book then talks about how the teacher gave the students a list of strategies they could try (“expand sentences” / “use transitions” / “give examples”) to come up with a much improved paragraph.

How well could an AI do this?

I gave ChatGPT the following prompt:

I want you to act as a teacher who is using the methodology of “The Writing Revolution” to teach students how to revise paragraphs. You are going to teach students how to revise by generating an unadorned paragraph about a topic. An unadorned paragraph is well organized and grammatically correct, but lacks variety and detail and uses only simple vocabulary and sentence structure. This is an example unadorned paragraph: “The Titanic was a big ship. It sank. Many passengers died. There are new regulations.”

I will be the student and you will be the teacher. You will give me an unadorned paragraph on the topic of The Wife of Bath’s Prologue from The Canterbury Tales. You will then walk me through steps for revising the paragraph to make it better by giving me a prompt for revising the paragraph. Example prompts that may work for improving the paragraph could be “try adding detail,” “explain why this is true,” “give examples,” “use more interesting vocabulary,” and “improve the topic sentence.”

After you give me a prompt to improve the paragraph, you will wait for me to revise the paragraph based on your prompt. After you read my revised paragraph, you will then give me feedback on how I did. If needed, you will then give me a new prompt for improving the paragraph and then wait for me to revise the paragraph based on the prompt. You will repeat this process until we have a paragraph about The Wife of Bath’s Prologue that is well written, well organized, with good variety of sentence structure, and good supporting detail.

And we were off! ChatGPT gave me the following paragraph to start with:

In The Canterbury Tales, The Wife of Bath’s Prologue is a story about a woman who has had five husbands. She talks about how she likes to control her husbands and how they should always do what she wants. She says that women should have sovereignty over their husbands and that they should be able to do what they want.

ChatGPT then walked me through a few different prompts to improve this paragraph:

  • “Can you add more information about the main character, the Wife of Bath, and her personality?”
  • “Can you provide an example or illustration to support the claim that women should have sovereignty over their husbands?”
  • “Can you improve the topic sentence to better introduce the paragraph?”
  • “Can you provide more supporting detail to strengthen the paragraph?”

By this point, I was getting kind of tired with this first experiment and ended the chat. For the record, this is the paragraph I ended with:

Through the character of the Wife of Bath in The Canterbury Tales, Chaucer presents a vivid and complex portrayal of a medieval woman who defies conventional expectations. She is a blunt and sexual woman who has had five husbands. She is smart and argues well. For example, she quotes St. Paul and says that women should have “power of the body” over their husbands, so women should be able to do what they want in marriage.

Things that went well:

  • Holy cow, I’m still amazed I can have this kind of interaction with software. A year ago it was impossible.
  • If I was ever stuck, I could ask ChatGPT for help, and its tips were pretty spot-on.

Things that weren’t great:

  • The steps that it took me through felt kind of robotic. It didn’t feel like a human was picking the best next step for my paragraph based on what I’d done, but that an uninterested and uninspired tutor was just picking the next step on a list.
  • Whenever I completed a prompt, I was always told “great job,” so I’m not sure how well ChatGPT will do at evaluating the output of novice writers.

After spending about an hour creating and editing prompts, testing them with various texts, and reflecting on my experience, I remain extremely optimistic about the potential of large language models to help students become better writers. Things aren’t perfect, and they don’t work out-of-the-box… but again, a year ago, this wasn’t even possible. I can’t wait to see what sort of educational tools we can build with this technology.