How would an AI illustrate a famous poem?

In my Linkfest earlier this week, I wrote about the new crop of AI “generative” models, in which you provide a few words and see it translated it into an image.

Often, these AIs create pictures that are fascinatingly alien. You get images composed with a strange blend of literalness and hallucination. The AIs try earnestly to depict what you described, but they filter it through an inhumanly vectorized form of machine perception that has no clue how the world actually works.

This is why I agree with what Alberto Romero argued, which is that generative AIs are powerful new tools for art, possibly on par with the impact of the camera back in the 19th century. Artists are learning how to use AI in art — a new type of paintbrush, on a new type of canvas.

It got me thinking:

Hmmm, what could I use a generative AI to do?

Well, one of my favorite genres is poetry, and one of my favorite poets is Emily Dickinson. (Above!)

Rather like a generative AI, Dickinson could be gorgeously hallucinatory and slightly alien. Her metaphors were always plugged directly into the voltage of human existence, but they’d often describe life from some oblique, bank-shot angle. “Tell the truth,” she wrote, “but tell it slant.” She never beat around the bush. She’d begin a poem by slapping you with a starkly forceful metaphor, like “My Life had stood — a Loaded Gun”.

This made me think that Emily Dickinson + a generative AI would be a pretty cool combination.

My favorite poem of hers is “Because I could not stop for Death”. It’s profound, weird, dark, witty, and crammed with soul-piercing imagery. You can read it in full here; it’s short.

So I fed it, couplet by couplet, into the “dream” generative-AI tool, to see how it would illustrate the poem.

Dream lets you pick from several different visual styles to guide the AI’s drawing. I chose three — “Steampunk”, “Synthwave”, and “Fantasy Art” — to illustrate each couplet, so we’d get three different attempts at visualizing each couplet. Those three styles seemed pretty appropriate because I’ve always thought that collections of Dickinson’s poetry should be illustrated like novels from the golden age of sci-fi.

Here’s what the AI drew, showing all three pictures for each couplet …

Because I could not stop for Death —
He kindly stopped for me —

In each picture, the AI seems to have picked up on “Death” as looming figure. The first one looks like an alien totem as Death; the other two look like murky Star-Wars-style menaces, but I kind of dig them.

That said, the AI isn’t really channeling the dry wit — and whoa-dude intensity — of this opening line. Dickinson is making a really stirring point about how death arrives unexpected and unbidden in the midst of life. But the AI is just illustrating things in a very literal fashion. Hey, the line mentions death! Let’s get a Death figure in here! Moody stuff, bruh.

The Carriage held but just Ourselves —
And Immortality.

Carriages, certainly! With a nicely desolate, steampunk kind of feel. The middle one looks like two people, one riding in a sort of sidecar? I like it.

We slowly drove — He knew no haste
And I had put away

I think the AI seized on the word “drove” here, which is why these first two illustrate cars, or car-like things. The third image is intriguing, though. I think it’s an attempt to show the inside of a vehicle, and picks up on the “we” in “we slowly drove” — with what looks like two unsettling, but evocative, humanoid figures.

Again, though. the AI is generally pretty literal. “Driving == automobiles, amirite?”

My labor and my leisure too,
For His Civility —

These ones kind of made me giggle. I think the AI has picked up on “leisure”, hence the figures sitting in recline. I dig it.

We passed the School, where Children strove
At Recess — in the Ring —

Definitely schools and children here. The “steampunk” style on the far left seems to maybe have the kids at desks? “Fantasy art” on the far right is more clearly outdoor recess, with even the hint of a ring.

My fave is the middle “synthwave” style. The school looks like it’s being hit by eldritch lightning, and …what in hell is going on with the kid? Are they in a three-legged robot pod or something?

We passed the Fields of Gazing Grain —
We passed the Setting Sun —

These pretty clearly show fields of grain and a setting sun. The second and third image processors clearly share some aesthetics, since they both hit upon the big floating vertical lines too, a sort of sky-born echo of the grain.

Again, though, nothing of the poetic dimensions of the language really finds it way into the images. These fields of grain aren’t “gazing”.

(To be fair, I don’t know how you’d illustrate “gazing” grain. I don’t even have a clear visual image of what that looks like in my head; I do not render the poem as a series of comic-book images, mentally, as I read it or contemplate it.

Indeed, this is part of what makes this AI-image-generating process so literarily strange. It makes me reconsider what precisely the act of reading a poem does inside my mind. Me, when I read the lines “gazing grain”, it’s certainly evocative. It makes the fields into an active character — theybear witness to me as I pass by, with Death, inside the coach.

But personally, I do not translate that into a pictorial image in my mind’s eye. I don’t visualize the gazing grain. Though maybe you do?)

Or rather — He passed us —
The Dews drew quivering and chill —

Lol, I have no idea what’s going on here! A spider-legged airship on the left, some sort of Bladerunner outtake in the middle, and … a massive dread platform rising over yonder bucolic hill?

TBH this may actually be a great illustration of the inside of Emily Dickinson’s consciousness, she was a serious weirdo.

For only Gossamer, my Gown —
My Tippet — only Tulle —

I love these dramatic, Victorian red/purple gowns!

That fibrous, almost fly-wing’s-like hatching: It seems like it’s picking up on “gossamer”, or possibly the archaic word “tulle”, which is “a sheer often stiffened silk, rayon, or nylon net”. (A “tippet” is a “shoulder cape of fur or cloth often with hanging ends”.)

We paused before a House that seemed
A Swelling of the Ground —

A house! But the metaphor at hand — the arrival at one’s own gravesite — appears to have utterly sailed over the AI’s head. This thing just does not grasp metaphor.

Or maybe … I’m the one being too literal here? The creepy exploded nature of that house on the far right, half floating in the sky, is something to behold. Maybe it’s channeling the idea of death in some super-subtle fashion? Maybe I’m taking the images too literally?

But I doubt it. I think it’s just drawing “a house”, in that sort of gently unhinged style of a GAN.

The Roof was scarcely visible —
The Cornice — in the Ground —

Man, this is getting depressing.

Since then — ’tis Centuries — and yet
Feels shorter than the Day

I like these huge buildings, but I can’t quite figure out why the AI thought that was the best way to illustrate the yawning stretch of time. Maybe because they buildings are … monumental? They loom like time?

I can’t detect anything in these images that reflects the witty conceit in this couplet — i.e. that, once we’re dead, each century feels brief as a day.

And now we come to the final lines …

I first surmised the Horses’ Heads
Were toward Eternity —

Heh, I figured it would nail the horses.

Weird horses! But weird a great way, I think. The ones in the middle really evoke the unsettlingly gorgeous specter of infinity.

All told, I think the AI did a cool but not terribly interesting job of illustrating the poem.

The main problem is the AI’s bluntly literal nature. It doesn’t seem to have any ability to grok the metaphors at hand.

To be fair, metaphor is a deeply nuanced part of human communication. The literary critic Northrop Frye used to say that metaphor was one of the most uncannily powerful mental tricks we possess. A metaphor suggests that two completely different objects are the same thing — i.e. that when you write “the river snaked through the forest” you are saying, an a sense, that the river is a snake. The aesthetic and cognitive force of metaphor is thus always kind of hallucinogenic. It creates meaning via association, not via logic. And it does so in a curiously brutish, almost bullying fashion. “That one thing you’re thinking about? It is also a different thing.”

So it probably isn’t surprising that it’s hard for these AI tools — which are, after all, a glued-together kluge of several sub-systems — to grasp the complexities of metaphor. After all, plenty of humans have trouble navigating metaphor in poetry. It’s part of why they find poetry — our literary form with the highest metaphor-per-utterance ratio — annoying. “Can’t you just say what you mean?”

Still, I enjoyed this experiment. I’ll be interested to see how newer forms of generative AI handle the fabric of poetry.