AI Character Building; or Who is Not Here?

Submitter: Kyle Booten, U of Connecticut, Storrs

——————————————————

The experiment:

At their best, classrooms can serve as “dialogic zones” (Freedman & Delp 2007) where perspectives collide through conversation. But what if key perspectives are missing? In this activity, based on my own work in AI-assisted writing (Booten forthcoming), students engage in a conversation about a significant topic—e.g., “Will AI make universities as we know them obsolete?” (though it could be anything). After an initial discussion phase, the instructor prompts students to take stock of what perspectives have been missing from the conversation and then asks them to fabricate an “AI character” to fill this gap. The instructor shares the interface of an LLM (such as GPT-3.5) on a screen. Fabrication of this “character” is accomplished through prompt engineering—for instance, “Respond in the perspective of an anarchist.” In practice, such unspecific prompts lead to boring, expected output. Repeatedly generating text in response to the question under discussion, the class iteratively revises and nudges to prompt by making it more specific and, perhaps, stranger. Revision stops once the AI character offers an interesting opinion, a perspective that produces surprise. The class conversation continues in response to this interesting opinion.

Booten, Kyle. Salon des Fantômes. Inside the Castle (forthcoming).

Freedman, Sarah Warshauer, and Verda K. Delp. “Conceptualizing a Whole-Class Learning Space: A Grand Dialogic Zone.” Research in the Teaching of English (2007): 259-268.

Results:

When I deployed this activity in a first-year writing classroom, I noticed that several students who had rarely contributed voluntarily to classroom conversations eagerly and energetically participated in the design of the AI character. In short, this activity seemed to be perceived as a “fun” one, and, as the students together designed their own character (a sort of curmudgeonly, grandfather with agrarian hobbies and Luddite sensibilities), they spontaneously engaged in back and forth debate about characterological details. Still, though students were interested in making a character, they seemed initially skeptical that this character could say anything of interest to them—and, indeed, the initially boring utterances of the character validated their skepticism. I had to encourage them to keep refining the character, eventually adding to the prompt myself to push it in a more interesting direction. Next time, I would scaffold this activity with a list of prompting tactics that (in my experience) can lead to more interesting outputs. On a more practical level, it is clumsy to have a spoken conversation and then to pause to feed part of this conversation to the text-based LLM interface. Next time I might experiment with using “speech-to-text” and “text-to-speech” models so that we could speak aloud to our AI Character and be spoken to in turn.

Relevant resources: https://www.kylebooten.me/salon.html

Contact: kylebooten.me