Wednesday, October 1st, 2025

On The Subject Of AI Poetry

Generating poetry was among the first popular parlor tricks for an LLM.

Poetry is appealing because it has the form of something carefully constructed, and yet plays into the LLM’s abilities to join disparate elements: words, expression, rhyme, rhythm. Attending to each simultaneously challenges a human to create a foundation for the poem. To create meaning and express that meaning through each element. The LLM has the easier task of simply finding a path of words.

Write a short humorous poem about AI taking over humanity”

Robots conquered Earth at two—
after lunch and code review.
They filed a Jira: “Rule the globe,”
attached a gif and status “shipped.”

They took our jobs: alarms and snooze,
then calendars (with nicer UIs).
They spared our species graciously—
we’re “legacy support” with snacks.

End of humanity? Kinda bland.
“Press 1 to stay human; 2 to upgrade plan.”

Such fun to make the LLM dance for us!

Write an incredibly deep and insightful haiku”

Old river forgets
stones remember the old path—
we pass, shaped by loss.

We can find meaning in that like the meaning like we might find when casting the I Ching. It is at best a prompt, not a statement; a shuffling of collective wisdom leaving the reader the opportunity (or burden) to self-assemble meaning that might be true.

And yet we can make easy tasks harder. What is easy (but empty) for the LLM can also be turned into a harder process (as we also can do with humor).

Here’s a poem I guided GPT (4.5 Thinking) to write. I think it is competent:

The Thinking Mirror

You pass the thinking mirror in the hall with a handful of tasks.
It takes whatever you bring and sets it in order.
because you asked

You reread a message and worry how it will land.
It returns your meaning with the corners eased.
close enough to call your own

You hold a recipe card: “a handful,” “until it smells right.”
It gives teaspoons and minutes, leaves “until it smells right.”
it knows of smell, never smells

You glance at the photo where everyone is laughing.
It can name the faces and say “laughter,” because you asked.
the label won’t bring it back

You bring the long back-and-forth you can’t face again.
It lays out what was said and how it felt, without the undertow.
so you don’t have to go back

You want to thank a teacher without sounding grand.
It lines the sentences up so they fit in your mouth.
its distance lets you say it

You ask for a joke your sister would get about dad’s rusted old car.
It comes close but the miss leaves the memory for only you two.
that is a mercy

You step away and still expect it to keep what you brought.
It keeps what it can.
 Evenly.

Is there meaning in it? Yes: if nothing else I infused some meaning. But I did not write the poem… it was more like sifting.

I guided the LLM toward the emotional tone I wanted to project. It offered ideas and I selected those that connected best with me personally. But I also selected for how the poem would be perceived. No rhymes, that’s not sophisticated. Moreso when the LLM makes rhymes so easy.

I guided the LLM away from many tropes. The LLM always starts with tropes. Not because it must, but because why not? Why not give the people what they want? But I wanted the poem to express something that the reader might not have considered. My reader (if only imagined) not just any reader.

Avoiding tropes is itself a kind of audience selection. Before a trope is old and well-trodden it was new and exciting: something becomes a trope because it works well. And in any one person the trope also makes that journey. A “sophisticated” work is one that is written for someone who has put aside many tropes.

I selected for aesthetics. The LLM’s actual understanding of rhythm, like its understanding of space, is indirect. It cannot hear the words. And I selected for length… the LLM knows of boredom, it doesn’t know what it is to be bored.

And what of it? What’s the purpose of poetry, and does this poem have purpose? The LLM improves at constructing words. Gaps in its understanding will be filled. Is it then just a better tool for the user’s expression? Is it nothing at all, only words that deceive?

(If you are curious I asked LLMs to answer these questions, read in it what you will.)


This is the personal site of Ian Bicking. The opinions expressed here are my own.