Can computers replace poets? I’ll share my thoughts, as a published poet, with a few ChatGPT-generated poems. ChatGPT, the AI-driven chat app that generates machine responses from human-provided prompts, has gone viral the past week (to the point where the free service is frequently unavailable due to server overload). We’ve recently updated this article at the bottom with additional sponsored content from Originality.AI (an AI and plagiarism content detector).
Google’s Chief Decision Scientist Cassie Kozyrkov describes ChatGPT this way: “ChatGPT — a fascinating new application of GANs (Generative Adversarial Networks) … a type of neural network that uses two competing networks — a generator and a discriminator — to create realistic-looking outputs. The generator creates fake outputs, and the discriminator tries to tell the difference between the fake outputs and real-world data. Through this back-and-forth process, the GAN is able to produce outputs that are indistinguishable from real data.” (read more…)
Using ChatGPT is simple – you type in a prompt in plain English and ChatGPT generates a response. Below are the first few examples I generated (this isn’t cherry-picking the best or worst examples). No edits were made.
Let’s start with a poetic form just about everyone has tried. Haiku typically follows a 17-syllable form (in English, three lines following this syllabic pattern: 5-7-5). Rhyming is not required. Very short, deceptively simple.
ChatGPT Prompt: Write a haiku about a Devon Rex cat.
A soft and fluffy ball A playful Devon Rex cat Purring in my lap.
ChatGPT adds extra syllables in the first and second lines. A reasonable first draft and if ChatGPT were capable of revision and editing I’d recommend removing “cat” in the second line, and “and” in the first line to tighten it up and align with the syllabic form (unless, of course, ChatGPT added syllables intentionally – because breaking rules in poetry with intention is encouraged).
ChatGPT Prompt: Write a sonnet inspired by a father holding his baby for the first time.
In his arms, he holds a miracle A tiny life, so new and so pure His heart swells with love and adoration As he gazes upon his little one The world around them fades away As they share this moment so divine The father's love is a guiding light A beacon of hope and protection His gentle touch, a soothing balm As he cradles his child close to his heart In this moment, he is complete A father, at the very start And as he looks into his baby's eyes He knows that nothing will ever be the same For in this moment, he has found his purpose A lifetime of love, now his to claim.
Asking ChatGPT to write a sonnet tests the limits of what is possible with AI-generated poetry. Beyond the idea, the technique of writing a sonnet is challenging; as Pulitzer Prize Finalist poet A.E. Stallings advised when I interviewed her on the Viewless Wings Poetry Podcast about rhyming in particular:
“When rhyme sounds jingly and jangly, and that’s all you’re noticing in the poem, that’s often to do with the syntax of the poem. If you have a poem that’s rhyming ‘cat’ and ‘hat’ and ‘mat’, which are all monosyllabic nouns, it means those monosyllabic nouns are ending up at the end of every end-stopped line. Instead, think about rhyming across parts of speech, ‘love’ and ‘above’ for instance, is a good rhyme. Then you’re going to have more interesting syntax going across the lines. If the rhymes are falling on different kinds of words that’s often a much more subtle effect.”Excerpt from my interview with poet A.E. Stallings
If anything, ChatGPT wrote a broken sonnet, but was the break from any known form of sonnet intentional? Was the lack of syllabic structure, rhyming pattern, and ignoring other elements of form an intentional choice required to convey the meaning of the poem? These are things poets labor over when revising and editing a poem.
In contrast, here is a Shakespearean sonnet I wrote which recently won 2nd Place in the Oprelle Oxbow Poetry Contest 2022. Ms. Stallings advice on rhyming was ringing in my ears (and she graciously provided notes on a draft of the poem, including suggesting I substitute “rooster” with “cockerel”, resulting in an extra syllable to be squeezed out of two lines).
A sonnet for a Neapolitan cockerel After Woman in Neapolitan Costume, ca. 1635 by Massimo Stanzione, Legion of Honor Museum (San Francisco) You won’t linger on delicate details painted in oil, the lace as light as air stitched to velvet, purple on red. Nor pale face—androgynous. Neither can compare to the cockerel, calm and cradled firm with red comb brilliant and speckled feathers, spindly legs snuggly tied so claws can’t squirm. Companion? Fighter? The two, together, make quite the pair, now ready to impress Napoli’s finest at the viceroy’s ball where princes will puff, exquisitely dressed, eyes on the hunt, making maiden catcalls. Still the cockerel, cool and sure, on guard for his princess in the festooned courtyard. copyright 2022 James C. Morehead
I worked hard during the revision and editing process to mix up the rhymes, to ensure they didn’t “scream out” their presence (e.g., avoiding always ending a line with a rhyme – lines that wrap-around).
Finally, I asked ChatGPT for a poem without any implied form. Arguably, this ChatGPT poem is the best of the three examples, as an early first draft.
ChatGPT Prompt: Write a poem about a haunted submarine.
The ocean is dark and deep A ghostly ship doth silently creep Through the waves, a silent stalker The haunted submarine of Captain Walker Once a mighty vessel of the sea Now a haunt, for all to see Creaking and groaning in the night A chilling sight, a terrible fright No living soul dares come near For fear of the ghosts that live in fear In the depths, they haunt and roam The haunted submarine, forever alone.
The ChatGPT poems included above are reasonable middle school (perhaps early high school) poems, and teachers will be hard-pressed to spot the difference between human and machine. Unlike direct copies of existing web content, these AI-generated poems can’t easily be recreated given the nature of the machine learning models used to generate the poems.
For AI models like ChatGPT to evolve will require simulating the strategies human poets use during the revision and editing process: reciting a poem to friends and family (or into a mirror) to assess the “sound” of the poem and to gauge the audience’s response; getting critical feedback from other poets (and parsing through contradictory feedback since nothing is distinctly “right” or “wrong”), sleeping on a poem for days/weeks/months, drawing on instincts built up from years of personal experience.
Detecting AI and Plagiarized Content
The additional information provided below is sponsored content from Originality.AI.
It is one thing for us to be able to tell the difference but I tested a tool that attempts to automatically tell when any written text (including a poem) is original (human-created) or AI-created.
Using several of the poems included above below are the results from Originality.AI which correctly identified each of the three poems (two AI-generated and one created by me). You can read more about Originality’s AI content detection accuracy here.
Other than as a curiosity, however, do we really need machine-generated poetry when humanity generates so many wonderful poets already? (For more on AI and poetry read: Playing with Found Poetry from Björk and the Limits of ChatGPT.)