Questions about the promise and existential threat of artificial intelligence are everywhere these days. What can A.I. do? What should it do?
As recently as last year, most of the attention devoted to A.I. (including in these pages) involved its ability to create visual artworks based on text descriptions, or prompts. But in the last few months, dizzying advancements in language-based A.I. programs have seen chatbots writing complex computer code and passing Advanced Placement exams, among other tasks. As a poet, I wondered how A.I. would do with a poetry assignment. So, I signed up for a ChatGPT account to find out.
I began with the following request:
Write a poem about the Provincelands dunes based on the structure of the poem “I Had Been a Polar Explorer” by Mark Strand. The line breaks and grammar should be the same, but the content should be different, and it should include a cormorant and be from a woman’s perspective.
The chatbot seemed confident that it could handle the assignment. But let me tell you, reader: it could not. Here’s the first iteration it produced:
I had been a wanderer
In the Provincelands dunes
Where the wind is a constant howler
And the sand a shifting moon.
I had seen the cormorant
Perched upon the dune’s crest
With its wings outspread and panting
In the endless, scorching quest.
I had felt the sun’s embrace
And the sand’s caress
And known the empty, lonely space
That only the wild can possess.
I had stood alone and small
In the vast expanse of sand
And known that I was nothing at all
In the grand and shifting land.
But still I wandered on
Through the dunes wild and free
For in that endless desert dawn
I found my own identity.
Hats off to the chatbot for knowing about hot sand and that cormorants stand with their wings outspread — but otherwise, ugh. Those rhymed quatrains are so boring! It’s nothing like Strand’s beautiful poem (which is not in rhymed quatrains), and it’s a bad poem in general, full of clichés and devoid of original specifics. So, my next attempt included some parameters to help the chatbot stick more closely to Strand’s structure: “Can you try again, but make it a 19-line poem and not use rhyming quatrains?”
The result was an even worse poem, from which I will spare you and to which I replied, “Wow. That is kind of corny and bad. Have you read Mark Strand’s poem at all?”
The chatbot answered:
I apologize if the poem I generated did not meet your expectations. As a language model, I can only generate text based on the examples and patterns I have been trained on. I do have access to Mark Strand’s poem, but my poem was not written based on that poem’s structure. I understand that it didn’t quite match up to your desired outcome. If you have any other specific requests, I’d be happy to try again.”
I felt a twinge of sympathy for the chatbot, which seemed to be trying so hard — though I was also uncomfortable with the subservient tone. It was clearly playing to my ego. I replied, “Please try again, but base your poem on Mark Strand’s poem. Use the same grammar, line length, and syntax, but change the subject, nouns, and adjectives to describe the Provincelands dunes from a woman’s perspective.”
We went back and forth like this several times. None of the generated poems were good, nor were they what I asked for. Rhyme was one thing A.I. couldn’t let go of.
I’ve got nothing against rhyming. Rhyme can be a powerful tool in poetry. Look at Eduardo Corral’s “Border Triptych” or Dylan Thomas’s “Do not go gentle into that good night.” But it wasn’t needed or requested here. I asked why the chatbot was still rhyming.
“I apologize for my previous response,” it answered. “I mistakenly returned a poem that still had rhymes. I’ll try again to create a poem with no rhyme, written in iambic pentameter and with the structure you described.” It then returned the following piece:
I had been a wanderer in Provincelands dunes
A place of shifting sands and constant skies of gray
A marsh harrier, I saw it circling high
Its wings outspread, it seemed to touch the sky
The wind, it whispered secrets in my ear
And sand, it whispered stories, ancient and sincere
I felt the lonely wildness of this land
And knew my smallness in this vast expanse
But still, I walked the dunes, free and alone
For in the endless desert dawn, I found my own
My spirit soared among the shifting sands
And in the Provincelands, my soul expands.
Better, I guess — but now there were rhyming couplets! ChatGPT was proving to be the worst poetry student ever.
I wondered if perhaps the example poem was the problem. So, I switched to Sam Hamill’s “Black Marsh Eclogue,” which I have used as a template for poetry exercises in this space before. This time, the chatbot tried to temper my expectations before proceeding further.
I understand that you want me to generate a poem based on “Black Marsh Eclogue” by Sam Hamill. However, it’s important to note that this is a very complex and layered poem with a rich use of symbolism and metaphor. It will be challenging for me to generate a poem that closely resembles it, but I will do my best to create a poem that captures the essence of the original poem.
As a poet, I’m flattered to be told that poetry is “very complex and layered” and would be difficult for a chatbot to generate. “Thank you for trying,” I typed. “Do you find poetry challenging?”
ChatGPT’s answer felt like both a promise and a threat:
As a language model, I don’t have feelings or opinions, but I can tell you that generating poetry can be challenging as it requires a high level of creativity, imagination, and understanding of literary techniques such as rhyme, meter, and imagery. As for my poems that use simple rhyme, it’s because it’s one of the easiest ways to generate poetry. Simple rhyme is the most common type of rhyme and it is easy for me to find words that rhyme with each other. It also creates a sense of musicality and rhythm in the poem which makes it more pleasant to read. I apologize if my previous responses did not meet your expectations, but I will continue to learn and improve my ability to generate poetry that meets different literary styles and techniques.
By this point, I’d seen enough. In the end, I don’t really care who (or what) writes a poem. But it’s important to me to feel that it matters to whoever creates it — that in writing the poem, the writer is grappling with how to understand something meaningful, beautiful, terrifying, or strange. However close ChatGPT and other language models get to the what of poetry, it seems they still have a long way to get to the why.
Meanwhile, just as I was finishing this article, Google announced that its new A.I. chatbot will be named “Bard.” Maybe that one will have a more poetic turn of code.