Aaron Meskin and Lindsey Harding
University of Georgia
This assignment asks students to use generative AI (viz., ChatGPT) to help write poems that they judge to be good and/or successful. Students were provided with easily available online resources about prompt engineering and the evaluation of poetry. They were encouraged to think creatively about their prompts and to revise them to generate multiple drafts of their poems. Then they were required to reflect and report on what they had done. Finally, students were invited to share their poems and accounts of their process in class. We have used the assignment in two educational contexts—a seminar for first-year students and a one-off lifelong learning course for adults over the age of 50.
Learning Goals
Original Assignment Context
The assignment was originally designed for a First Year Odyssey Seminar at the University of Georgia. First Year Odyssey seminars are small-size (typically 15-18 students) one-credit classes that are required for first-year UGA students. The FYO program began in 2011 as the University's Quality Enhancement Plan linked to reaffirmation of its accreditation. Over 300 seminars are taught per year on a variety of topics chosen by instructors. They are designed to “engage first-year students in the academic culture of the University” and “promote meaningful academic dialogue between students and faculty in a small class setting that encourages reflective thinking and learning to learn” (See FYOS 1101 Course Description).
This particular FYOS seminar focused on Artificial Intelligence, Technology, and Art. The seminar description was as follows:
This seminar will focus on how artificial intelligence and related technologies are changing our relationship to art. We will explore the question of whether AI systems such as DALL-E and Midjourney can make art on their own and, hence, count as creative artists, or whether they should only ever be seen as tools that are used by humans to make art. We will also explore how we should evaluate and interpret the art that is made using such systems. And we will look at the way that machine-learning recommender systems are affecting the way that we experience and engage with art and music. Are recommender systems problematic in any way, or are they just a good way to get you to listen to things you might like? We'll read a lot of fun articles and do some experimental art-making of our own.
Materials Needed
Time Frame: Short homework assignment plus debrief or one long class session (approximately 75 minutes)
Overview
Introduction
Early academic discourse about generative AI, especially ChatGPT, was largely negative and anxious. We were interested in taking a more positive approach and exploring the potential usefulness of generative AI tools in a creative setting. More specifically, we were interested in how ChatGPT might be used as a tool to facilitate or support creativity. If creativity is, to a first approximation, the intentional and non-mechanical production of new and valuable things (Boden, 2007; Gaut, 2010), it should be possible to use generative AI creatively. The hope was that this assignment would get students to explore that possibility. Other goals included encouraging reflection on prompt engineering, highlighting the significance of revision in the creative process, and calling attention to the limitations of generative AI.
Initial Experiment
In the FYOS seminar, students completed the assignment roughly midway through the semester. By that time, they had been introduced to uses of generative AI within both fine and popular art, had explored philosophical theories of creativity, and discussed the specific possibilities of computer and AI creativity.
Students were strongly encouraged, but not strictly required, to present their work in class. Most students took the opportunity to do so. Presentations generated good discussion, as students noted ChatGPT’s limitations and were amused and engaged by their peers’ poetic goals and prompts.
One student wrote very expressively about her experience using ChatGPT to write a poem about the feeling of fall. This was her initial prompt, which shows an attention to detail and the student’s use of a simile to describe the poem’s desired tone:
Write me a poem that gives me a visual of how people feel in the fall. From the leaves changing colors, to Starbucks new fall flavors, pumpkin-scented candles, and more. When I read the poem I want to feel like I’m listening to a love letter to fall.
The student described her experience of using ChatGPT as follows:
Some adjustments I asked Chat GTP to make were to add more vivid details and imagery to set the scene and make the poem more graphic. Next I wanted to change the style of the poem so I asked it to make a haiku then realized that was too short for my liking, and I wanted to read more about Fall and all that comes with it. So, I settled on having Chat GTP create a sonnet. I also wanted the poem to have descriptive vocabulary in order to move the reader.
Here is the third stanza of the final version of her poem:
In cozy scarves and sweaters, we embrace,
The crispness of the air, the leaves' soft grace.
With pumpkin-scented candles glowing bright,
We find our way through autumn's gentle light.
Another student took a distinctive “distanced” approach to the use of ChatGPT. His aim “was to contribute as little to the substance of the poem as possible, in order for it to be as much of GPT’s words and, more importantly, ideas as possible.” Although he chose the theme, “the substance of the poem would be all on GPT.” To do this, he had ChatGPT generate twenty descriptors for two settings that would be compared in the poem. He directed ChatGPT to end each line of the poem with one of the descriptors it had generated. His process and the prompts he used evince a creative take on prompt engineering reflective of his experimental approach.
Here is the final version of the poem he generated:
“Nexus”
Amidst the heart, where boldness calls,
Aspirations reach for the skies, a shimmering world,
A symphony of glass and steel, geometry,
From the heights of a starry night.
But in canvas, a profound grace persists,
Carved by time, a tranquil solitude profound,
Layers etched in stone, their timeless embrace,
Where silence's echoes are softly found.
In the nexus of dreams, where worlds entwine,
Their beauty aligns, a shared splendor.
WIth respect to evaluation, the first student was fairly positive about the poem:
I would say Chat GTP came close to creating a successful poem. One with sound, rhythm, rhyme and format. Poems often feature intense feelings which to me was in this poem. It brought out my feelings of love and appreciation towards fall which I would say goes along with what makes a poem, a poem. It’s no man made poem, but I think this was one of the good ones Chat GTP has made.
The second student was a bit more mixed:
Overall, I’m not in love with the final poem it produced, but I also don’t hate it. I think it's a pretty average poem, which is great for chatGPT considering what it normally writes.
Second Experiment
In the lifelong learning course for adults over the age of 50, students were asked to complete this assignment as part of a seminar on generative AI and creative writing. During the class, we looked at how AI could be used as a writing tool and how AI appeared in creative writing. Since the class only met once, asking students to experiment with generative AI to write poetry seemed more feasible than working with longer forms of creative writing.
After creating accounts with ChatGPT, the class discussed their prior knowledge, experience, and perceptions of generative AI. Then students prompted ChatGPT to write a personal jingle to introduce themselves to the class before we moved into the poetry activity. Following a brief introduction to poetic virtues and prompt writing, the class engaged in a process of writing and rewriting prompts and generating poems until they had a version they wanted to share with the class. Afterward, we discussed students’ experiences and their take-aways from interacting with ChatGPT.
One student attempted to generate a poem in the style of Emily Dickinson, though she acknowledged that ChatGPT was not as successful at emulating Dickinson as she had hoped. Here is the opening stanza from the third iteration of her poem.
In twilight's hush, with sun's retreat,
The squirrels scamper, quick, discreet.
In shadows deep, they dig and paw,
Their mission clear, their aim I saw.
Another student prompted ChatGPT to write a personal poem about her life experiences and roles—an older woman, mother, wife, English teacher, and novelist. In response, this is the poem she shared:
In twilight years, a woman wise,
with tender heart and searching eyes.
An English teacher, mother, wife,
Each role woven into her life.
In golden years, her pen takes flight,
Novels born from depths of night.
With wisdom's grace, she crafts anew,
A legacy for ages, true and true.
Comparing Student Engagement
Both groups were eager to experiment with using generative AI to write a poem. For the adult learners, their prompts were rather simple and general, especially compared to the more imaginative and specific prompts employed by the undergraduate students. In addition, the adult learners seemed keen to use ChatGPT to recreate a recognizable poem (i.e., in the style of a famous poet, like Emily Dickinson) or generate a poem about a familiar topic (e.g., coffee, themselves). The undergraduate students, meanwhile, approached prompt engineering creatively to generate something fresh (i.e., the poem ChatGPT wanted to write), to explore a new perspective (i.e., a visual rendering of fall feelings), or to “challenge” ChatGPT. Here, we see these young adults consider the innovative potential of generative AI and the older adults adopt a more conservative approach, using the tool to replicate human expression digitally.
Estimate: 75 minutes (during and/or outside class)
Students were given the following instructions and resources. In the first-year seminar these were provided at a class session, and the assignment—along with a brief presentation—was due the next session (viz., one week later). In the lifelong learning class, the instructions were provided at the beginning of the class session, and students were asked to be prepared to share the poem with the class by the end. When introducing this assignment, instructors may want to emphasize how prompt engineering itself presents an opportunity for creative thinking and writing.
Instructions
I want you to try to write a (good and/or successful) poem using ChatGPT. In order to do this, you will have to:
(a) plan your poem
(b) think about and revise prompts for ChatGPT
(c) do multiple drafts (i.e., get ChatGPT to do multiple versions, ask for changes, etc)
(d) select the best one
(e) write up a brief (one page) report on what you did and why
(f) be prepared to share your poem with the class and say a few words about how you got it
(g) submit the report along with a document that shows how you used ChatGPT
Resources
Students were provided with a set of resources to think about poetry and prompt writing. While there are countless videos and materials available online to address these topics, the following offer information and insights to get students started.
What makes a good and/or successful poem:
Various poetic virtues:
Tips for writing effective prompts:
Prompts to guide reflection
The following prompts were shared with students to encourage them to think intentionally and specifically about their goals and choices for this assignment, which they were asked to write about in a brief report and also share with the class.
Adaptations for Adult Learners
Presentation
A Google Slides presentation was used to structure the class session and walk learners through the series of activities and discussions in a step-by-step manner.
Initial survey on prior use and knowledge on generative AI
Given the one-off nature of this class, the following questions offered a quick way to prime students for engaging with generative AI and also get a sense of their prior knowledge and experiences.
Warm-up activity
This brief initial activity built on the survey above and offered a way for everyone in the class to introduce themselves and meet their peers.
Prompt ChatGPT to write a personal jingle to help introduce you to us. Be sure to consider including the following information in your prompt:
Then, after we shared our jingles, we moved into the poetry assignment as outlined above (see Instructions). Together, we watched the YouTube video on what makes a good poem and discussed the poetic virtues listed above (see Resources). After reading through one of the websites on prompt-writing listed above (see Resources), students were set loose with ChatGPT to iterate between prompts and poems while the instructor circulated. After students shared the poem versions they were happy with, the class reflected on the experience.
Reflection prompts
The following prompts, slightly revised from those listed above, facilitated a debriefing discussion:
Boden MA. (2007). Creativity in a nutshell. Think, 5(15): 83-96. https://doi.org/10.1017/S147717560000230X
Gaut B. (2010). The philosophy of creativity. Philosophy Compass, 5: 1034-1046. https://doi.org/10.1111/j.1747-9991.2010.00351.x