Bot Poetry: Engineering Randomness

BY STAFF WRITER BEN GUZOVSKY ‘24

Source: Computer Generated Art by Jakub Cech

How do you teach art to a machine? Great writers have the ability to find the “right words” and to use those words to put together an evocative and meaningful phrase. At first glance, this skill is very hard to automate. Computer scientists have been taking a surprisingly effective approach for decades: randomness. After building a grammatical base for a machine to work with, they use some element of random word generation. Whether drawing from a dictionary or from published writing, that strategy seems to work — computers are getting better and better at fooling people into believing their writing is human-made.

There’s a growing movement in the literary world to embrace what is computer generated as genuine art. Just a month ago, Lillian-Yvonne Bertram’s Travesty Generator, a poetry collection that utilized Python and Java to write certain lines, was longlisted for the National Book Award. It’s one of many examples where humans have collaborated with machines to create something they couldn’t have alone. With collaboration comes questions of authorship — questions we can only explore with a better understanding of how these machines work.

Machine learning poetry programs are complicated and, more importantly, proprietary. We can only guess how they work. To try it out, I did some coding on my own — not with the hopes of winning a National Book Award, but to see what this trend was all about.

I started with haikus. Fewer words mean less complexity for a COS 126 student to deal with, and they’re just as instructive as longer writing. You can illustrate all the principles of computer-generated poetry in 17 syllables or less. First, a quick refresher on the modern English language haiku: there are no rules. The 5–7–5 form has been discarded as a relic of bad translation. In Japanese, that flow of syllables is much more natural. Haikus don’t even have to be three lines anymore. For example, examine Francine Banwarth’s poem, a recent winner of the Haiku Society of America’s Museum of Haiku Literature award:

off to on I disappear into the visible

Using around 300 of the most commonly used words in published haikus, I created a random haiku generator. The program wrote each line of the haiku using a “mold” — a predetermined order of words, like “noun, verb, adverb” or “conjugation article, noun, verb.” Each mold was drawn from published haikus as well. Then, I generated ten thousand haikus.

Of the couple hundred I read, 30% were logical and grammatically correct, 10% could pass for human-made, and a handful had literary merit. Here are two of my favorites:

stone blossoms between hills

cloudy river disappears

blue daydreams

moment —

everything swimming leisurely

a leafless world

I couldn’t resist tweaking a couple of the haikus that were almost evocative, but had one word out of place. In the following composition, I changed only a single word:

one snow —

through the window

a withered plum tree

I’ve saved the best for last. The true brilliance of the program is not its ability to create haikus, but the way it randomly puts words together. Here are some phrases it invented:

a stars’ gathering

withered rage

world-moons

Here, the benefits of collaborating with a machine start to become visible. Whereas the goal of a computer scientist is to fool people into thinking these haikus are human-made, the goal of a poet is to write the best poetry. Since creating the program, I’ve used all of these phrases, these little computer-generated ideas, in my own writing.

With a completely customizable input for this program, there is limitless room for exploration. By changing the word bank and the molds, the program will generate prose or other forms of writing as well. Computer-aided writing is the future. Make this modern thesaurus your own.

I have a sense that this is cheating. The “author” isn’t technically the one writing. The chess world has a similar problem, as computers are close to solving the game but human players continue to compete for millions in prize money every year. Machines break chess because unlike humans, they have no preconceived notions of how one should play a position, no years of habit and teaching that they are afraid to deviate from. Chess engines see an idea, and then players have to struggle to figure out why such a strange move is the best one.

Writing isn’t a competition. When a computer puts words together, it doesn’t know if they sound good to a human reader. It is up to the author to decide if the computer-generated phrase is worth using, if it needs to be changed slightly, or if it doesn’t work at all. Unlike chess, where there is always a right move, writing maintains enough ambiguity for computers to be useful, but not dominant.

These authorial choices, the slight changes that make computer-generated poetry palatable, take agency away from the machine. There are few real questions of authorship to answer here. It doesn’t matter who writes the poetry — or, better yet, what machines help the writer — if the poetry is beautiful. The author is just a name on a page, accompanying art.

If an artist writes their own program, all is well. When another artist uses that program for their own art, even with slight modifications and a completely new, randomly generated output, is that ethically acceptable? Writing has always been a cumulative process, with one author building off of another’s ideas. But these aren’t implied connotations and allusions, or even structural choices. Anyone can write a sonnet without citing Shakespeare. Computer programs are someone else’s tools. Anyone can use tools like a thesaurus to aid with writing, but as more and more resources become available to authors, where will we draw the line?

We already have humans to write poetry; let’s write something that humans can’t. As nice as that idea sounds, even that phrase suggests an element of plagiarism, of taking something from a machine. Computer-aided writing is full of as much possibility as it is pitfalls.