As far as poem titles go, I think “when the book of rain is calling the sun beautiful” is a pretty good one. It’s evocative, and has enough promise (what, exactly, is a book of rain? Why is it flattering the sun?) to make me want to read on. Admittedly, I might like that title because I wrote it. Except that I didn’t, quite. My computer did.

Recently, I’ve been writing poems with neural networks. The work of neural networks and their potential role in writing literature has already been discussed in Do Computers Write Electric Literature? Part 2. I use torch-rnn by Justin Johnson, based on previous work by Andrej Karpathy called char-rnn. torch-rnn runs in Linux and requires the Torch set of libraries, which are specifically designed for machine learning. At a basic level, this particular neural network learns connections between characters and words by studying a large text file. Once it’s done so, it can then generate new text one character at a time in an attempt to write output that’s somewhat akin to the original data.

Most of the existing work with this tool has focused on rewriting text based on publicly available datasets. Karpathy walks through examples of generating new lines of Shakespeare, Wikipedia entries, and even source code. Janelle Shane has created whimsical results such as new names for metal bands, undiscovered Pokémon and malfunctioning knock-knock jokes (you can read a great interview with her here). Yet, to my knowledge, no one has endeavored to use torch-rnn to generate something that might be considered new, legitimate literature.

I’ve been a practicing poet for about twenty years, and have published books and individual poems in literary journals. Like everyone else, I also have a festering pile of unpublished work. One day, feeling creatively stuck and like I never wanted to write another damn poem ever, I thought: What if the computer could write my poems for me?

I collated all of my work I could find into a giant text document, about 76,000 words and 435kb. I say giant, but in the world of neural networks, that size is laughably small. I forged ahead anyway, and after a lot of trial and error, I got torch-rnn to spit out things that read like “the bar is a bird / we have the speations of the stars” and I saw some surprising possibilities. After further refining of how well the neural network learned my writing, off we went, creating fifty poems.

Here’s the final stanza from “coming of age: much about the blood”:

the sunburn of the sky around

would come and laughed and wolf me.

a tired green comfort learned you off—

we’re so shadows tied to the moon.

you want to be rockets,

expecting even your exploding.

There’s so much weirdness injected by the computer that I find charming. The program clearly has no sense of tense; it manages to cram a conditional verb with a simple past one with an invented present tense verb (“wolf”) in a single line. “a tired green comfort learned you off” is a line that makes grammatical sense but is difficult to interpret literally; “you want to be rockets” throws us for its use of a plural and lack of an article.

More importantly, the program writes lines, using my own words, that I would never write. “the sunburn of the sky” is an intriguing metaphor, and I can’t help but read “we’re so shadows tied to the moon” in my best angsty-teen voice. These poems were mine, then they weren’t mine, and then, after editing (if you think that stanza is nonsensical, you should see the nonsense stanzas I cut wholesale from the output), they are finished as these unnerving “mine-not mine-but mine” constructions.



This leads us to a larger set of related questions: 1.) Who wrote these poems?, 2.) Does it matter?, and 3.) Can poems written in this way have meaning?

It’s important to state two virtually unquestioned assumptions we have when encountering a work of literature or any art. First, we assume that it was created by a human. For centuries, we’ve had no reason to suppose otherwise. And today, even when we know we are speaking with a computer (think of talking to Siri, Alexa, or Ok Google; or navigating infernal corporate phone menus before speaking to an actual representative), we know the system was still clearly designed by a person. Second, we assume that the creator of a work had some sort of intent behind its creation, however literal or abstract the idea. Humans cannot create something truly random; they compose based on their knowledge, skills, tools, intellect, and emotions, and those guy wires are impossible to escape.

The first assumption is easier to set aside, because it has been prefigured by New Criticism in the 1940s, which argues that a poem should be considered as a self-contained unit and de-emphasizes whatever circumstances, including the author’s biography, that arose to the poem’s composition. This belief was taken to its logical conclusion in 1967 by Roland Barthes in his aptly titled essay “The Death of the Author.” Barthes argues, simply, that the author is irrelevant; and for that matter, who can hope to guess precisely what the author intended? Reader-response criticism emerged in the 1960s, further emphasizing the importance of a reader’s reaction to a piece over any claims any author may have to it. Certainly you don’t have to believe that the author is irrelevant (I find it hard to separate Sylvia Plath from her biography, for example), but there is a critical tradition for doing so.

But what about intent? Even in the most abstract paintings, we know there were choices made about color, medium, form, and motion. And Language poetry, with which I believe this torch-rnn project shares some affinities, has at its core not always literal meaning but overarching goals of weirding words and creating new relationships with the reader. But this apparent void of intent, too, has literary precedent: consider Burroughs’s cut-ups. The cut-up technique is a simple machine designed to produce strange new writing with disregard to the intent or meaning of the source. In some literature, intent may not be everything.

A computer doesn’t care. torch-rnn is completely ignorant of what the words mean or how parts of speech work. All it knows is what characters can be grouped together. Given the letters th to begin a new word, torch-rnn knows to often follow that up with an e and a space to start the next word. In my particular case, torch-rnn is extremely confident that I like to use the word stars in my poetry—that word shows up in almost every poem the program generates.

So if the creator truly is creating something meaningless to it—it is agnostic as to whether it is creating poetry or new Pokémon—does that mean the resulting product is meaningless to us as readers?

In my case, it’s difficult to say, because in no way am I claiming these poems pass the Turing test. I wrote the source text; torch-rnn generates the raw material; and then I shape it through editing which is often gentle—I try to preserve as much of the output as written as possible—but sometimes severe: I’ll cut entire stanzas that don’t seem salvageable, and sometimes the entire output looks hopeless, so I simply generate a new text. Perhaps a better term, then, than computer-generated for these types of poems might be computer-collaborated. Since there always needs to be a human-created source text behind them, and, for now, a human also has to audit the results, it seems impossible for the AI to create any work independently. Some person needs to be the ghost in the machine.

As for meaning, the best way to approach these kinds of poems is to hold on to your sense and structure of the English language rather loosely. I find the end results of my collaboration with torch-rnn to be surprisingly tender. The resulting poems use my own atomic language—lake, cloud, stars. And the themes are similar to what I’d write typically write: how “you” and “i” navigate our natural world together, through love, pain, joy, death. These poems ask readers to think more deeply about the connections they make, since the logic usually isn’t as apparent as a “normal” poem. But I’m convinced these poems do have meaning. When I read them, I learn something new about the relationships people can have with each other and with the natural world, through the matrix of a pidgin language that the computer and I agree upon, even though neither of us entirely understands.

You can find five of B.J. Best’s poems here on Empty State.


B.J. Best is the author of three books and four chapbooks of poetry, most recently But Our Princess Is in Another Castle and Yes. He’s also the creator of the Arty family of Twitterbots, such as @ArtyBots and @ArtyAbstract, who invent visual art and then reinterpret it in conversation. He’s on Twitter as @bjbest60 or you can reach him at bjbest60 at Gmail.