Composing with AI

Readings

Before students can get something out of LLMs, they really need to understand how these systems work, how critics understand them, and where they come from. Our reading list had iconic hits and some personal favorites: Stochastic Parrots, Blurry Jpeg (everyone needs a good metaphor to make the technical and unexplainable concrete), Kirschenbaum's Textpocalpyse, as well as a few of my favorites, readings from Joseph Weizenbaum about the first chatbot, ELIZA, and Alan Turing, who instigated the quest for an intelligent-seeming bot.

The magic of the “Chat” part of ChatGPT is that it takes advantage of the ELIZA effect, which names a human tendency to imagine or assign to non-human things sentience. The typical example would be the person who is cursing at their car. But a better example might be the person who is growing more and more frustrated while trying to get their Amazon Alexa or Apple Siri to understand their request. Before they know it, they are abusing Alexa (Fan 2022) and telling it to “shut up” as though it had the ability to understand the emotional force behind that phrase.

At the heart of this misapprehension, treating the bot as a person, is a fundamental misunderstanding of how LLMs operate. While the code is hidden and the precise functioning of the LLM obscured, the general notion of a bot producing content based on a predictive model is fairly straightforward. When students read “Stochastic parrots,” they develop a clearer expectation for the results of their prompts. The very accessible metaphors and analogies in these readings serve as clear handholds for students of any technical knowledge to enter the discussion. However, there are scholars who specifically address the issues of race and the exclusivity of artificial intelligence, Safiya Noble and Joy Buolamwini to name just two. Incorporating their insights on the reading list invites and empowers students to consider the elements of white supremacy (and other exclusive worldviews) that pervade both the bots and the texts they were trained on.

The Turing Test

Alan Turing, father of computing, presented the world with an imitation game, or rather two. In the first, a man tries to pass himself off as a woman. In another, analogous game, a computer tries to pass itself off as a man, for a man could hardly calculate with the speed of a computer. He offers these games as alternatives to the tar pit created by the question of machine intelligence. By the end of the century, he claimed, computers would be able to pass themselves off as humans in limited situations more than half the time. ChatGPT goes a long way to achieving that goal. But Turing’s game inspired an in-class activity.

In this exercise, students answer a reflection question in four sentences. I like choosing a question about the nature or future of machine intelligence for the echoes with Turing. Then, they generate another four-sentence answer from ChatGPT or other LLM. The students post one of the two answers in a discussion forum, and then the class tries to guess whether the text was human- or machine-generated. The results surprised me.

Even in our first run of this exercise, the students produced text that was quite difficult to identify as either human- or machine-generated. On the one hand, students were fairly sophisticated in their prompting, cuing the machine to write as a middle or high school student or to make some errors. On the other hand, and this surprised me even more, students had developed some skill at imitating the cadence and rhythm of ChatGPT, especially when lightly prompted. They had developed the ability to imitate what I call its "person-less prose," consisting of complete sentences of similar length and structure.

Thus, rather than testing the AI's performance, this exercise tested the students' skill at producing language in voices other than their own. That dexterity with word choice, that ability to recognize and imitate style is something that I would happily include in the learning outcomes for my class. Furthermore, the exercise suggested a relationship to prompting LLMs that was far more sophisticated than merely asking for text and submitting it as their own, the scenario most alarmist offer as the harbinger of the end times of composition instruction. By contrast, this exercise showed how the AI could be a central part of a discussion of what makes human writing unique and how LLMs munge their input training sets.