Google AI helps you write a poem in the style of famous poets of the past

Google has released "Verse by Verse", a AI tool that lets you pick several poets from the past, and then write a poem with the AI suggesting lines composed in their style.

As Engadget describes it:

You can select up to three poets for inspiration, including the likes of Emily Dickinson, Walt Whitman and Edgar Allen Poe. Once you've made your choices and picked a structure for your poem, the tool will ask you to compose your first line of verse. The AI will then suggest some more options.

Verse by Verse won't lock you into using those suggestions. You can ditch or tweak them, or accept them as is. The tool is supposed to inspire you, not generate an entire poem on your behalf — though you can more or less do that too. Once you have perfected a stanza, you can add more of them to your future masterpiece.

To build the tool, Google engineers fed the system "a large collection of classic poetry." They then used each poet's own work to fine tune the AI models in an attempt to ape their writing styles. They also wanted the AI to make relevant suggestions, so, according to engineer Dave Uthus, "the system was trained to have a general semantic understanding of what lines of verse would best follow a previous line of verse. So even if you write on topics not commonly seen in classic poetry, the system will try its best to make suggestions that are relevant."

I decided to give a whirl myself.

For my dead-poet collaborators, I picked Emily Dickinson, Emma Lazarus (author of "The New Colossus", the poem about the Statue of Liberty that included the famous lines "Give me your tired, your poor, / Your huddled masses yearning to breathe free"), and Sara Teasdale (who's utterly metal and who I wrote an essay about a while ago).

You can pick whether you're writing in quatrains, couplets, or free verse, how many syllables per line, and whether the poem should rhyme or not. (I picked quatrains rhyming in ABAB format, with 10 syllables — so, mostly iambic pentameter.) I wrote this first line …

The Internet is darkest 'fore the dawn

… and then accepted, in their entirety, the next three lines the software suggested. The second one was by the AI Emily Dickson; the third line, by the Sara Teasdalebot; and the fourth, by the undead pen of Emma Lazarus, as replicated by a neural net that feasted upon a vectorized meal of her life's work.

The result was thus; I added the title after the bots were done with their dark work …

It's … not bad?

Deep-learning AI poetry is, in my experience, pretty hit or miss, with a highly bimodal distribution: The vast majority of the time it produces perfectly-fine, if unremarkable lines, metaphors, ideas, and turns-of-phrasings — and then every so often you get a burst of pure WTF, how-did-a-bot-do-that greatness. In this poem here, the lines the AI wrote here are really just serviceable; not great. To be fair, my original line is pretty meh too; not much for the AI to work off of.

I suspect Google's right, and this type of tool is a bit dull if you use it to autogenerate a whole poem — and much more interesting if you employ it as a prompt, to crack the ice of the mind and get you moving. If you were to roll up your sleeves and tweak/edit the lines the AI generates, it'd be more like the "centaur" intelligence — part machine, part humanity — that chess master Garry Kasparov pioneered back in the 90s.

After being beaten by Deep Blue, Kasparov decided human vs. AI was a rather boring competition, and it'd be more productive to cooperate with AI. So he invented a form of chess where a human using chess software is pitted against another human using chess software. Kasparov found that, once you give a human the machinelike ability to ponder tons of different moves, it produced a creatively new style of human chess. Players explored different routes they'd not have previously had time to investigate. (I wrote about this in the first chapter of my first book; a wee excerpt is here.)

So theoretically one could get the same increased possibility space out of "Verse by Verse". You rapidly iterate a ton of possible poetic gambits, then focus in on the few gems, and hone them. The act of poetic creation becomes less one of composing lines as editing them — the lines produced by the machine. This is more or less how musicians have been using new tune-generating AI. About a year ago I wrote a piece for Mother Jones about that scene, and the songwriters told me the AI tended to have a terrible signal-to-noise ratio: Tons of dreadful ideas, punctuated by some wild ones. So they ran it as as volume business. They used the AI to toss out gazillions of tunes, and their human task became curatorial, trying to sift flakes of gold from the slurry of mud. The French songwriter Benoit Carré used a Sony music-generating AI, and described it to me this way:

"You are a little bit like an artistic director or a producer, and you have a crazy musician in the room," he told me. "Most of the time it is crap," but every so often the machine kicked out a melody he would never have thought of. Carré helped write the lyrics and recorded the album with a group of meatspace musicians. It certainly wasn't push-button easy. If anything, sifting through the AI's output for useful, provocative passages was like panning for gold—probably more work than writing everything himself. But the silicon intelligence helped Carré break out of ruts. "In pop music, you know, it is always the same chords," he says, so to do something new "you have to be surprised, you have to be shaken." 

A while ago, Robin Sloan, a writer I've known since our blogging days in the early 00s, crafted his own experimental text-generating AI trained on old sci-fi magazines. He found that it was intriguing, and took your mind in usefully alien directions — but it wasn't just a free gift. It was a tool, the productive use of which required work, and required figuring out how to use it:

I should say clearly: I am absolutely 100% not talking about an editor that "writes for you," whatever that means. The world doesn't need any more dead-eyed robo-text.

The animating ideas here are augmentation; partnership; call and response.

The goal is not to make writing "easier"; it's to make it harder.

The goal is not to make the resulting text "better"; it's to make it different — weirder, with effects maybe not available by other means.

This nails it, really. I'm going to play around some more with "Verse by Verse" and see if I can get it to drag my brain into a corner I wouldn't normally go on my own.

(As as side note, it looks like the poets they used to train the AI are from before 1923 — so they're in the public domain, in the period before copyright slams down. It's too bad; I'd have loved to be able to do this with some poets from later in the 20th century.)