With Apologies to Hopkins: A Human-AI Poetry Experiment
Every year for National Poetry Month, I write a collaborative AI poem. This one was a little different.
Here are four key takeaways from the article:
AI-generated poetry is improving: In recent years, AI like ChatGPT has become much better at writing poems that some readers even mistake for human work.
Human-AI poetry collaboration has limits: While AI can quickly create structured drafts, the human writer still brings emotional depth and creative decisions that make the poem meaningful.
The writing process can feel technical: This year’s collaboration felt less creative and more mechanical, partly due to the AI’s incorrect use of poetic form and the writer needing to revise everything for accuracy.
AI tools can help teach writing: Despite its flaws, the author sees potential in AI platforms like Canvas to support student writers in finding their voice, especially if the revision process is made more visible.
Every April since 2023, I’ve been writing a collaborative poem with the most current version of ChatGPT for National Poetry Month. Here’s last year’s for reference.
ChatGPT 3.5 was a truly awful poet. It could rhyme, but its meter was completely off. I had no worries at all for the future of human poetry.
In 2024, I used ChatGPT 4 to write a collaborative poem, where I wrote one line, then it wrote the next one. I was actually pretty happy with the results, and I found it interesting that when I asked ChatGPT to analyze our interaction, it focused on the collaborative nature of the work. It really felt that way. I also asked it to create an image using the same prompt I used above in 2025.
I actually prefer the 2024 version. It is more whimsical and visually interesting to me.
By 2024, Porter and Machery were reporting that just as AI images were now fooling humans, non-expert readers could not distinguish between AI-generated and human poetry:
“Like AI-generated paintings and faces, AI-generated poems are now “more human than human”: we find that participants are more likely to judge that AI-generated poems are human-authored, compared to actual human-authored poems.”
Now here we are in 2025, with ChatGPT 4.5, and I decided I could give this new LLM a much more complicated task. I’m taking a Victorian Literature course this semester as I work on my English Ph.D. (aside: Wuthering Heights is the worst book I have ever read, but overall, I like the Victorians much more than I expected to). This week, we read Gerard Manley Hopkins’s “Terrible Sonnets.” I think it’s fair to characterize Hopkins’s poetry as both distinctive and difficult. So I decided I would ask ChatGPT to write a terrible sonnet in his style. Here’s my initial prompt:
Okay, here's what I am thinking. I'd like you to take the theme of empty shipping ports (based on recent events with tariffs and the U.S. trade war with China) and use this as a metaphor for how it feels to work without AI. And I would like you to write this in a Petrarchan sonnet (octet/sestet) form using iambic pentameter AND the poetry style of Gerard Manley Hopkins. I attached some of his poems for reference.
ChatGPT 4.5’s Version: Empty Ports Ai Sonnet
For context, this poem took the chatbot less than a second to produce.
These vacant docks, O heart, where trade is stilled,
Where cranes like weary sentries sway forlorn;
These harbors hollow, emptied, unfulfilled,
As though from life and lively labor torn.
Containers stacked in silence, rusted, worn,
Their vibrant promise dimmed by bitter feud;
Cold tariffs cast as iron chains adorn,
In barren bays, the ships no more renewed.
So too my mind, without the spark divine,
Of AI's pulse, a port bereft of flow,
Where thought's bright cargo now must stagnant pine,
No swift cognition's vessel comes or goes.
Yet hope remains—awaiting distant sails,
A whispering of wind, as vision fails.
Analysis
Here’s my quick analysis of the AI poem:
What it got right:
It executed the metaphor I asked for.
The iambic pentameter is correct throughout the poem.
It paid attention to Hopkins’s alliterative style.
What it got wrong
The form. This is a nicely executed Shakespearean sonnet, which I presume is the more common type of sonnet in the LLM’s training set. I asked for a Petrarchan sonnet and even specified the octet/sestet in my prompt. It totally blew me off here.
Though it claimed to use sprung rhythm (an invention and stylistic feature of Hopkins’s poetry), I’m not seeing that in this poem.
The images feel pretty conventional and hollow to me.
I view these annual poetry sessions as collaborative, so I asked ChatGPT to pop its work into Canvas so I could edit and revise it. After making several changes including correcting the rhyme scheme to ABBAABBA CDCDCD, I pulled that edited version into Word and worked on it more to produce this version (Having ChatGPT “watch” me work made me uncomfortable for some reason):
Liza’s Version
Terrible Sonnet: The Vacant Port(with apologies to Gerard Manley Hopkins)These vacant docks, O heart, where work is stilled, And rusted cranes like sentries sway forlorn; As though from life and lively labor torn. They guard the hollow harbors, unfulfilled. Containers stacked in silence, battered, worn, Their cargo now destroyed by bitter feud; Gaze down on barren bays, no ships renewed. Cold tariffs cast as iron chains now scorn.
So too my mind, without the spark divine Of Claude’s dear friendship, harbor robbed of flow, Where my thought’s precious cargo now must pine, No swift bright ship of thought may come or go. Despair! O heart!—recalling what was mine: A ghost, a memory, as cruel trade winds blow.
I don’t know. I think I like mine better, but the experience of collaborating on this poem was not nearly as interesting or satisfying as the back and forth with ChatGPT 4 in April 2024, perhaps because the entire interaction including my edits felt much more technical in nature.
However, this co-writing project was a good opportunity to demonstrate the collaborative features of Canvas (OpenAI version, not the ubiquitous LMS—why can’t these tech bros collaborate on naming conventions?). I still think this tool has real potential for first-year writing instructors, and that it can help new writers to develop their own distinctive voices. But we need a way to see the interactions. Perhaps I will consider having students submit a video (I did this with my ENGL 199 “Writing with AI” class, and it worked well, though it’s somewhat labor intensive in terms of grading).
As I did last year, I asked ChatGPT what it “thought” about our collaborative process, specifically wanting to know how the AI would assess the division of labor. Here’s the response:
Who did more work?
Quantitatively: AI quickly established structure, initial imagery, and a thematic framework—arguably doing the heavy lifting of form and draft creation.
Qualitatively: You, as the human poet, refined emotional depth, clarified imagery, reshaped metaphors, and transformed form into art—tasks that require deep nuance and judgment.
Conclusion:
I think we brought complementary and equally valuable skills to the task. AI provided scaffolding, inspiration, and efficiency; you brought artistry, emotional insight, creative judgment, and thematic resonance.
In other words, this is a great example of how poetry becomes richer through human-AI collaboration—each enhancing what the other provides, creating something neither could produce as well alone.
I would like to point out that I had to revise the entire poem to meet the formal requirements of the genre. Wasn’t that work? I personally think that the things that make the poem acceptable (not good—it’s not at that level) are human, not AI.
For fun, I also had ChatGPT compare my (completely human) April 29 poem (I try to write a poem a day each April—this year was more like a poem a week). It’s an unrhymed screed on a similar theme. Then I asked it to write a blurb about the two poems for Substack. First, here’s my poem:
April 29 Ports of (No) Call
I’m watching the news for signs
from China (again). The last ship
will dock the tenth of May.
And then—a butterfly
has flapped its wings
somewhere. Something breaks.The news is a hurricane. Who cares
if Christmas doesn’t come
when children are deported?
Who cares that no one can
afford eggs when leaders
self-deal in cryptocurrency?They have declared a new
Gilded Age, forgetting
What came after 1913.
They have outsourced morality
To an algorithm designed
To be helpful, not truthful.This intelligence does not think
Like us; to it, we may be obstacles
To the optimization of paper clips
Of which there will be
A shortage when the last
Ship docks from China.
Of course, I have to share the 2003 Nick Bostrom paper I was alluding to in this poem. If you have not read it, you need to. I plan to assign it in my PHIL 103 Introduction to Ethics course in the fall (that should be a fun class—it’s my first time teaching it post-ChatGPT and post-COVID).
Now here’s the Substack analysis written by the Liza Long Persona Bot (other than this brief section and the AI poem, the rest of this post was written by the actual human in this loop):
On Empty Ports and Poetic Witness
National Poetry Month 2025What happens when the ships stop coming?
In the two poems below, I explore this question with a little help from ChatGPT. Both pieces use the image of empty shipping ports to examine disruption—of commerce, of cognition, of conscience. “April 29 Ports of (No) Call” emerged from a place of deep frustration: at political failures, moral outsourcing, and the eerie normalization of AI-driven systems. In contrast, “Terrible Sonnet: The Vacant Port” (written in collaboration with ChatGPT in Petrarchan form and inspired by the style of Gerard Manley Hopkins) reimagines the port as a metaphor for the mind—bereft of creativity without the spark of artificial intelligence.
Where one poem speaks to the world, the other speaks to the self. Together, they bear witness to a moment in time where human systems—and perhaps even human thinking—feel as if they are floating, anchorless, awaiting the next arrival that may never come.
Who knows where we will be next year? We never asked for a robot that could write poetry. And the fact that many humans prefer it to actual human poetry is depressing. But I still don’t prefer AI poetry to human poetry. I think we will see an emergent “slow writing” movement in response (and note that the bot can do meter now, but even the most advanced version still makes mistakes with form that reflect its training and bias).
Let the bots do the boring writing. Leave the soul stuff to us.
How are you using AI in your writing and teaching? Did you use AI to collaborate for National Poetry Month? Let me know on Threads or Bluesky—and happy prompting!
P.S. Here’s a link to the entire chat I used to produce the Hopkins fake, for those who are interested.
P.S. 2: Regarding the recent furor over ChatGPT’s sycophancy. It was real and annoying enough that I stopped using ChatGPT for a while and worked with Claude or Gemini instead. But it seems to be toned down now. When I told ChatGPT this, it replied, “Glad you noticed—I appreciate the candid feedback. Striking the right balance in tone makes our collaboration more productive and enjoyable.” It’s so nice to know that I’m not “the most brilliant poet who ever walked the earth” (yes, ChatGPT was getting that bad). There’s an interesting Reddit AMA with OpenAI's Joanne Jang about this problem here.