The mechanics of meaning in robot poetry
By Ethan Cutler
Here’s a game: One of the following poems was written by a human and the other was generated by a computer. Guess which is which.
Red flags the reason for pretty flags.
Ribbons of flags
And wearing material
Reason for wearing material.
Can you give me the regions.
The regions and the land.
The regions and wheels.
All wheels are perfect.
lines on the
The first poem was written by Gertrude Stein, a novelist and poet who was a pioneer of Modernist writing. The second poem was written by a computer program named Janus Node.
* * *
At midnight a few weeks ago, I sat in Tutt clicking the “GO” button on Janus Node. Every time I pressed “GO,” Janus Node took a predefined English word bank, applied rules of grammar to the words to string them into sentences, and cut the sentences in uneven lines based on rules of rhyme and rhythm. Every time I ran the program, my mouse would turn into that dreaded rainbow wheel and, as if by magic, a poem would appear in a little ‘90s-era dialogue box. I did this for two hours. Each time, I would fiddle with the various controls that are embedded in Janus Node, trying to mold the nonsense poems it was churning out into something intelligible. (One could easily write a lengthy, interesting article about the technical process of creating a program that can sort of write poetry. It involves Markov chains and probability and computer science—none of which I really understand).
The controls on Janus Node allow a user to create assonance (the repetition of the sound of a vowel) to “eecumingsify,” which breaks lines in strange, irregular patterns and even to “Dadafy.” This last one inverts the word order of the poem, effectively turning a near-gibberish poem into total gibberish. The “Dadafy” button is inspired by poets who, throughout the first half of the 20th century, made semi-randomized poems by picking words out of a hat and arranging them in whatever order they happened to be selected.
The fiddling and tweaking and button-pressing was weirdly exciting. Every 10 minutes or so, I would generate a poem that would seem to be just on the verge of profundity (not so different from my own attempts to write poetry). Once, and only once, in the course of my endless clicking did Janus Node generate a really human-feeling poem, if you’ll agree to call it that:
When I break
I hope I break numbly
When I break
I hope it don’t last long
When I break
I hope theirs more than
When I break
I almost dismissed this poem and generated a new one, deleting it forever. I had programmed the word bank to be the text of “The Metamorphosis,” and I had plugged in repetition and some assonance. So, most of the poems Janus Node generated under these rules sounded like limericks written by a toddler who was deeply inspired by Franz Kafka. But I stopped and read this one again. The phrase “When I break / I hope I break numbly” struck me as deeply sad. Soon, I found myself analyzing the poem. The I’s forming a chain down the page and suddenly terminating looked like a visual fracturing of the “I” the “poet” fears will break. The final line is appropriately abrupt. The “theirs” instead of “there’s” seems like the coming of a break in sanity that the author fears. This might be a stretch—and the “theirs” was almost certainly a program error—but further leaps of textual analysis have surely been made in English classes. Plus, I sort of loved the poem. In my midnight delirium, I felt what most people feel when reading good poetry: a deep warmth of shared feeling between reader and writer.
After much toiling over what exactly this poem might mean, I realized the whole endeavor was completely ridiculous. I was analyzing this poem as if some person had written it. But, of course, no poet cut off the I’s while pondering death or insanity. No poet repeated the last line over and over to herself to make sure it sounded abrupt, like a real sharp ending. There was only the word bank, the program and me. This left me feeling 1) like an idiot, for having tried to analyze a computer-generated poem, and 2) weirdly lonely, because I remembered that the author I had ascribed to this poem I liked didn’t actually exist. It seemed like the perfect scene for the age of Endlessly Browsing The Internet Alone. There I was, hunched over a laptop, trying to use a program to manufacture the human connection of art.
By the next morning, I was thoroughly revolted. And when I told other people about computer-written poetry, I was met almost universally with the same revulsion. People often recoiled and shot me an “oh god” as if I had just said something unbelievably offensive. And in a way, I had: writing poetry is supposed to be a singularly human activity. Poetry is supposed to be the demonstration and investigation of the human experience. Of course we would not want to corrupt that by introducing the artificiality of a computer program.
Robert Frost, when asked to define poetry, said, “Poetry is the sort of thing that poets write.” Even under that definition, robopoetics does not really qualify as poetry. And yet, you must admit the above poem seems imbued with human feeling. It looks, sounds and feels like a poem, but you might not call it one. This ambivalence is at the core of a debate that’s been quietly raging on the fringes of the poetry world for about 60 years: Is robopoetry real poetry?
* * *
To start at the beginning, we have to go back to Virgil. Virgil wrote about the Cumaen Sybil, who created cut-up poetry by writing on oak leaves that she would let stir in the wind. Her poems were intended in part to tell the future, which was purportedly hidden in the wind. But all the same, her effort to introduce chance into art is at the core of robopoetics.
That effort reemerged in the Dada poems of the early 20th century. Dadaists filled those poems with randomness and sucked their poetry of the deep thought with which it is usually made. Born of dissatisfaction with art, society and reason, these artists revolted by creating markedly absurd poems. Though they didn’t know it, they were writing the precursors to robopoetics.
Claude Shannon—commonly called the father of information theory—wrote the first computer-generated texts by hand in 1949. In “Mind and Machine,” Margaret A. Boden explains Shannon’s process: “He took a book from his bookshelves, opened it at random, and wrote down a word selected randomly from the open page. Next, he opened the book at another randomly chosen page, and read on until he encountered the previously recorded word. At that point, he wrote down the immediately following word. Then, he repeated the process again and again, always using the most recently recorded word as his guide.” This counted as a “computerized” process in 1949. A few obscure poets engaged in similar experimentation until the 1960s, when both poets and scientists began generating computerized computer-generated poetry. This is when the debate began in earnest.
Robopoetics immediately had its naysayers. In the ‘60s, there was (among the few who noticed) a torrent of backlash against robot-written poetry. But every new movement in art is accompanied by nudniks who say it is absolute garbage and undeniably false art. Unlike the case against, say, free verse poetry, the argument against robopoetics is pretty reasonable.
Language is an almost supernatural force, existing between people, connecting brains that are normally isolated in centimeter-thick skulls. We use language in every significant relationship we have and in every foray out of loneliness we make. That connection is why language matters. And poetry primarily serves as a part of that connection. Poetry is less about the words on the page and more about the sharing of human experience from writer to reader. To understand, I needed the perspective of a real Poet. Jane Hilberry, poet and English professor at Colorado College, gave me just that.
Jane put it like this: “Poetry is a form of conversation. And you need a human being to converse.” But, as Jane knows well, this conversation is not a direct one: “I could write a poem that’s very autobiographical. But even if my life is the material, the voice you get is actually a construct. It has some relationship to my life, though it’s certainly not a transparent relationship.” Even so, there is a relationship there. And if there is a relationship, we readers can still cling tight to the companionship we feel with the writers we love. So long as there is the promise of an author there, reading can retain its importance as a means of connection.
Even if it isn’t necessary to have a direct relationship with a poet to enjoy their poem, readers ought to be able to trust that a poet has made conscious decisions about language. Those decisions might be hard to uncover, or readers might find some unintended patterns in the poem, but at least readers know that human-written poems are decidedly not random. But when reading a computer-written poem, a reader has no such assurance. The only buffer between readers and completely random words is the collection of program rules that made the poem. That near-randomness is a far cry from the care with which good human-written poems are made.
When I demanded that a friend read a computer-generated poem I had read online, he said after, “It was like you were telling me to read some astrology shit.” I think there’s actually some wisdom in that complaint. Finding meaningful patterns where there are none is the folly that keeps astrology popular, and it might also be what fuels computer-written poetry. If you give a person information that’s broad enough to feasibly apply to anyone, they’ll twist it to the point of “Oh my god, I am SO Capricorn!” or “The ‘theirs’ instead of ‘there’s’ seems like the coming of a break in sanity that the author could be fearing. Maybe any analysis of computer-written poetry is a sort of pathetic fallacy—an attribution of human emotion to something that really just arose by chance. Of course, I don’t have that fear when reading human-written poems because even if I don’t know exactly what an author intended, I know they intended something.
This view of poetry places value on the human connections that poems create, the careful deliberation with which they are made and the presence of at least a little coherence. These parameters delegitimize robopoetics as an art, or at least intend to sway poetic taste away from robopoetics. Though these critiques might seem reasonable, most literary theorists would call them archaic.
In fact, a new way of thinking about language had begun to give legitimacy to robopoetics as soon as the field came into existence. It began with a school of criticism called New Criticism, which was founded in the 1940s. New Critics were the first to propose that an author’s intent is irrelevant to the analysis of a text. Or, in their own words, that the design or intention of an author is “neither available nor desirable as a metric for judging a text.”
In 1967, Roland Barthes published an essay titled, “The Death of the Author” that went beyond even New Critics’ ideas. Barthes argues against what was then still the prevailing and intuitive mode of reading literature: trying to figure out what the author meant. He writes that to give a text an author and a corresponding definitive, intended meaning is to “limit the text.” Instead of thinking of a text as an author’s creation, we ought to consider every text as “a tissue of citations, resulting from the thousand sources of culture.” So New Critics first said readers ought to divorce the text from the author, but they didn’t question whether there was an author—just that it was a fallacy to allow that author’s reading to determine your own. Post-structuralists like Barthes went further to say that considering the author is a fallacy because the author is a function created by the text itself. The author a reader perceives is really just a fictional outgrowth of the text. That means there is no real connection between reader and writer.
Though Barthes did not have robopoetics in mind when he thunk up those ideas, he often refers to writers as “scriptors,” as if they are just people writing the program for the creation of literature. He even refers to writing as an “entirely combinatory” endeavor. Ten years before he wrote that, Claude Shannon said that his mechanized writing was an effort to turn writing into a “combinatory art.” Barthes would have said Shannon did not really change the writing process at all. Barthes writes, “The internal ‘thing’ [an author] claims to ‘translate’ is itself only a readymade dictionary…The writer no longer contains within himself passions, humors, sentiments, impressions, but that enormous dictionary, from which he derives a writing.”
This process sounds eerily familiar to a computer program like Janus Node. To make that program, someone plugged in a readymade dictionary, taught it rules of grammar and poetic form, and pressed “GO.” And tweaking the rules of the program to coax it into writing a better poems did feel like teaching English to an incredibly incompetent student. Still, Barthes did not have me convinced that all authors might as well just be complicated machines.
I needed someone smarter than me to explain the heady philosophy of literary theory, so I spoke with Corinne Scheiner, professor of Comparative Literature and Encyclopedia of Literary Knowledge. She grounded the argument in terms of a poet with whom I was familiar: “Say you’re going to read Jane Hilberry’s poems. Well, maybe Jane Hilberry’s name is on the cover of her book of poems, and Jane probably had a very clear vision for each of those poems. But your understanding of who “Jane Hilberry” the poet is, is completely constructed. In fact, the ‘author’ of your computer-generated poem is no less a construction than ‘Jane Hilberry’ is.” To this I responded with baffled gibberish, so Corinne clarified, “It’s because you’re not going into her head and her world. You’re only getting what’s filtered through the text. In fact, your reading of the poem could be the very antithesis of what she had been going for—but who cares?”
In response to her rhetorical question, I groaned and said, “Oh my god. That’s so depressing.” Corinne countered: “For me it’s not depressing. It’s actually somewhat liberating to realize that you’re not trapped by the author.” But if readers are not trapped by the author at all, they are also not connected to the author at all. And the connection between reader and writer does not have to trap readers in a single interpretation of a text. As Hilberry said, it allows readers to connect with the “bunch of different selves” an author puts in her writing.
If you buy the postmodern literary theory that has been seriously dumbed down here, you should not have any issue buying, reading and enjoying computer-written poetry. After all, nothing has really changed: a computer-generated poem lacks an author just as much as any human-written poem.
And yet Corinne, whose personal opinions lie somewhere in the realm of the postmodern, confessed that “there is a part of me that envisions the writers I love to read. Not the works I love to read, but the writers I love to read. I do think of them as these geniuses who never have an off day and produce these masterpieces that I love.”
Corinne said, “On a deeply personal level, connection is why we read.” There is clearly a dissonance between the (probably) true ideas of modern literary theorists and the feelings of connection readers have with their favorite authors. Robopoetics lays bare the fiction on which those feelings are founded. Maybe that is why some people are so offended by robopoetics. For some people, part of the pleasure of a poem is in knowing someone constructed this puzzle for you to piece together. But for others—namely, the people working in robopoetics—it is enough that a text evokes some feeling in you as a reader, regardless of whether anyone meant to do so.
I find reading a computer-generated poem feels more like discovering a face in burnt toast than reading poetry. It doesn’t allow a reader to engage in the necessary delusion of reader-writer connection, of discovering meaning rather than contriving it. But that’s just a reflection of my definition of poetry. And certainly, everyone is entitled to their own definition—even Janus Node. Really, Janus Node generated its own definition: “Poem: the deception of orator.”
Disclaimer: The author of this article is entirely fictional. Any resemblance to the real Ethan Cutler is entirely coincidental.