How To Break ChatGPT With Dialogue
The beginnings of an AI-proof assignment
The beginnings of an AI-proof assignment
We rarely ask students to write out dialogue in the university, even though it is one of the best ways to make a story (whether in creative or professional writing) engaging.
But it is hard … and students avoid writing dialogue like the plague, even when specifically asked to write dialogue.
This week, my students and I delved into the world of good dialogue, exploring how AI generates dialogue and ways to make it better.
As we analyzed AI-generated dialogue, we discovered that there were certain elements that make dialogue more engaging and realistic — elements that AI often lacks.
The Four Key Elements of Good Dialogue
I decided to break down the components of good dialogue into four categories. This not only made writing dialogue more accessible, but also provided a framework to help AI generate better dialogue.
These categories are:
Naturalness — Instruct AI (or students) to sound natural, according to the character’s unique traits and background. A well-developed character can greatly assist AI in generating more natural dialogue.
Subtext — Great dialogue often has hidden meaning. To help AI (or students) generate dialogue with subtext, add a sentence describing the underlying context. For example: “Subtext: Louis says he’s in love with Ed, but he actually has grown indifferent.”
Vary Sentence Length — Engaging dialogue often involves an unpredictable back and forth between characters. Inform AI (or students) how you’d like the interaction to unfold, such as: “Vary Sentence Length: Ed tends to dominate the conversation.”
Clear the Clutter — Encourage AI to show, not tell, and avoid overusing dialogue tags. Provide examples of what you want to see in the dialogue, like: “Show: Louis is getting nervous and wants to leave the room.”
This gave students more language to discuss various examples of dialogue (AI or human) and also some strategies they can use themselves when writing their own dialogues.
The Creation of a Super-Prompt
I also thought this was a great way to build a super-prompt that might help AI generate more interesting dialogue. I built one based on these principles and tested it out in class.
First, though, we generated a dialogue with a simple prompt. After having AI generate a plot for a thriller where a federal agent is taking down a Mexican crimelord named El Sapo (a student’s idea), I asked ChatGPT to write a romantic dialogue.
Here was the result.
Notice that it is pretty generic and straightforward. Everyone says what they mean, and there is no real tension or conflict. The characters also have the same voice.
So I built a super-prompt that gave AI more details that might help it create more human dialogue, incorporating the principles above. This required giving character descriptions, plot point descriptions, the principles of good dialogue, and other detailed instructions.
Having students build these themselves, or even adapt them, is a great way to help them learn and remember literary terms and strategies. One might even argue that they will learn these more using super-prompts then just sitting down and trying to write dialogue on their after reviewing the four elements.
The AI-Proof Assignment
I realized that this super-prompt could be considered an “AI-Proof” assignment. No matter how detailed the super-prompt, we discovered that AI couldn’t quite fulfill our request. For example, this generation is better, but still lacks several elements of the super-prompt.
I even tried GPT-4. It was definitely better, but not complete.
So … I could turn my super-prompt into an assignment sheet for students … and they now have an assignment they can’t use AI to fully complete. (If that is something you want).
This method of including multiple variables and elements, and then testing AI’s response, can be an interesting and engaging exercise for students, but also a way to create assignments more difficult to auto-generate.
When students have a deep understanding of what makes good dialogue, they can apply these principles to improve AI-generated dialogue or write their own dialogue. They get to choose, but either way, they learn how to analyze dialogue and apply it to their writing — the real outcome of my course.
So I’m not really sure it is important that all our assignments are AI-proof, if we ask ourselves about the true outcome.
Is the outcome really that students write good dialogue? In this case, no. The real outcome is that they can analyze how dialogue works in stories.
I’m currently writing weekly reflections on my AI & Digital Storytelling class. Sign up below for past and future issues.