ChatGPT Poetry Is Surprisingly Pretty, But It Lacks Two Things

Answering the question of whether AI can make art
Linda Caroll

Bard playing string instrument with glowing butterflies hand drawn illustration
Bard playing string instrument with glowing butterflies, illustration licensed from Deposit Photos
It started so simply. Someone sent me a poem, asked what I think. So I clicked to read it. The words were pretty enough. The poem used simile instead of metaphor but that’s whatever — could be a style choice.
But something felt weird. The poem felt off somehow. Wasn’t quite sure what it was that struck me as not right. Not right away, anyway.
On a whim, I pasted the poem into an AI detector. Boom, it scored 100%. So I pasted it into another AI detector, and another. Four in total. They all said the same thing. 100% AI. Zero chance this was written by a human.
Until that moment, I don’t think I’d ever read poetry written by AI.
Honestly? I was a little horrified. Staring at that poem, knowing AI wrote it, here’s what occurred to me. Some people will look at the pretty words and, like Kerouac standing by his truck in the dark of night watching the fireworks, they’re going to say ooh, ah, wow. Think it’s good.
It’s a lesson we humans still haven’t learned.
That pretty isn’t always good.
We haven’t learned it with people, much less with words. Look at some pretty thing, forget what good even means, too smitten by pretty.
So let’s talk about what good poetry is, okay? Then I’ll show you the two things ChatGPT poetry lacks. Things that make it pretty, but not good.
Several years ago a man named Tom Hirons wrote a stunning poem called Sometimes A Wild God. Here are the first two verses. Just a teeny snippet, because the entire poem is almost two dozen stanzas long.

Sometimes a wild god comes to the table.
He is awkward and does not know the ways
Of porcelain, of fork and mustard and silver.
His voice makes vinegar from wine.
When the wild god arrives at the door,
You will probably fear him.
He reminds you of something dark
That you might have dreamt,
Or the secret you do not wish to be shared.
— Tom Hirons, Sometimes A Wild God

The entire poem is beautiful. Stunning and I encourage you to read it. Buy the book if you really love it and want to support the work of an actual poet who’s in the trenches making art instead of typing in AI prompts.
But for now, I want to show you a comparison, okay? I asked ChatGPT to write a poem about a god coming to the table.
Here’s what it wrote:

The table is simple
bread torn, wine poured,
a fragile offering of what we have.
The light falls unevenly,
casting long shadows.
We wait for words
but the god is quiet.
There is no sermon, no decree.
Only the sound of breath
and the creak of the chair beneath divinity.
— written by ChatGPT

Pretty words, yeah? They are pretty. I admit it. Bread torn, wine poured, light casting long shadows. Makes you want to say hey, old chatty did pretty darn good for a mindless robot AI computer program.
Now let me explain the difference.
In Hiron’s piece, he says “He is awkward and does not know the ways
of porcelain, of fork and mustard and silver.”
A robot would never say that.
Not in a million years.
Do you know why? Because it’s uncommon use of language.
They’re just things on the table. Words in ChatGPT’s database. But it’s an uncommon combination of words. So it won’t make the connection.
Using words in uncommon ways is what a poet does. But using words in uncommon ways is a human ability. It is not how AI works.
AI, specifically large language models (LLMs) run through the billions of words they were “trained on” to find the words “most likely” to be correct based on probability. Not the words least likely to be strung together.
If you were to read the entire poem, all twenty-some verses, it will leave you breathless at the beauty Hirons created with words.
Like this:

The wild god reaches into a bag
Made of moles and nightingale-skin.
He pulls out a two-reeded pipe,
Raises an eyebrow
And all the birds begin to sing.
— Tom Hirons, Sometimes A Wild God

There’s no way AI can write anything like that. Ever. All those words? Common words a grade five student knows. Every one of those words is sitting in the robot’s belly. But it won’t put them together that way.
Because it can’t.
If AI was tasked with putting words together in uncommon ways, it would be a nightmare. There’d be no way to control the output without a brain.
Funny thing is, that’s the marker of a poet. Of poetry. That’s what poetry is. Uncommon use of words. It’s uncommon use of language that moves us.

Hope is the thing with feathers — Emily Dickinson
Let the soft animal of your body love what it loves. — Mary Oliver
And when he dies, cut him out in little stars — Shakespeare

Let’s talk about what AI can and can’t do for a minute, okay?
A decade ago, there were programs online called content spinners. I’d spend five hours writing a comprehensive seo guide for struggling website owners and syndicate it to a couple of sites. Then eleven schmucks would paste my work into a content spinner and voila, it would spit out a copy that passes every plagiarism test there is. Because yay synonyms.
Large language models (LLMs) like ChatGPT are kind of like that, except more advanced. Like comparing Rosie, the Jetsen’s cute little robot maid, to the Terminator. They don’t just use synonyms anymore. Basically they ate the internet and now they’re spitting out content based on probability.
That’s what AI is. Probability generation. People look at the output, say wow, they think it’s writing. It’s not writing, it’s calculating the odds of any one word appearing next to another according to all the words in its belly.
It can tell you how seo works, the benefits of meditation, how to get a stain out of cotton, or anything factual thanks to the Wikipedia pages, blog posts, dictionaries, encyclopedias and copious number of books they scraped.
What it can’t do is think of a new way to say anything.
It can’t use words in uncommon ways.
That’s the opposite of how it is programmed to work.
Simon Willison is a programmer who has written an entire explanation of how large language models (LLMs) like ChatGPT function. It’s a thirty minute read, but super informative if you’re curious how LLMs function.
He calls LLMs like ChatGPT money laundering for copyrighted data.
He’s not wrong.
In 2006, the kids in a high school class in New York were told to write to their favorite author, invite them to visit. Five kids wrote to Kurt Vonnegut.
Of all the letters that went out in that class, he was the only author to reply.
He said he doesn’t do appearances because he looks like an iguana but he told the kids to go home and write a six line poem. When the poem is done, he said, don’t tell anyone. Don’t show anyone.
Tear it into pieces and throw it in the trash bin.
He said they’ll have been rewarded just by writing poetry.
He told them they’ll have learned something about what’s inside them. Told those kids that poetry and art will make their souls grow.
Less than a year later, he was gone.
Slipped and fell. Sustained grievous injuries to his head that took his life. He died in April, 2007.
Ten years later — to the month — in April 2017, the National Institute of Health published a report saying poetry and music affect our minds. Confirming that art changes our neural pathways and psychophysiology and those changes can actually be seen with neural imaging.
Poetry is one of the oldest art forms on Earth.
It pre-dates literacy and prose.
Before people could read, before there were printed words to be read, there was poetry. The oldest known poem is over four thousand years old. Memorized and passed from parent to child until finally it was tapped in stone in cuneiform script, the oldest alphabet known to man.
Poetry isn’t made of line breaks. It’s made of uncommon word choices.
I leave you with one last sample.
I asked ChatGPT to write three lines of freestyle poetry about how it would feel to watch someone you love die.
Here’s what it wrote.

Breath falters, a rhythm breaking into stillness,
their gaze drifts to places I cannot follow,
I sit, holding what remains of forever.
— written by ChatGPT

When Dylan Thomas watched his father fighting for his life, sat at his deathbed and watched him die, here’s what he wrote.

Do not go gentle into that good night,
Old age should burn and rave at close of day;
Rage, rage against the dying of the light.
— Dylan Thomas

Two things you’ll find in that piece, in all good poetry, that you won’t find in AI generated poetry. Uncommon use of words. And emotion. Neither of which a computer program is capable of. Only humans make art.
Kurt Vonnegut letter from More Letters Of Note
ChatGPT