Almost every day for more than a decade, Ranjit Bhatnagar, a poetry lover, would sit down at his computer and find treasure waiting for him. There on his screen, beamed directly to his Brooklyn apartment, were streams of completely original verse.
More often than not, the little poems were delightfully surreal: Too many lobsters for the cubs today./ The baby bought a arm and ran away.
Sometimes, they were subtly profound: Back being stuck in my emotions, yay./ Please let tomorrow be a better day.
Other times, they were funny: Stop doing acid at the office, Jim./ I’m talking to myself referring him.
And they were always couplets of 10-syllable lines, written in perfect iambic pentameter — the classic form of English poetry popularized by William Shakespeare.
Bhatnagar had neither subscribed to some sonnet-of-the-day service, nor did he have some prolific admirer or aspiring lyricist in his patronage. In fact, he had little clue as to the identity of the online bards — and almost none of them had any idea that they were writing in verse. Many may not have even known what iambic pentameter is.
A computer engineer by training, Bhatnagar had programmed an algorithm that scoured Twitter for lines of exactly 10 syllables of alternating emphasis, paired ones that rhymed, and tweeted out the resulting couplet from a separate, anonymous account. He named the bot and the account Pentametron.
It was 2012, and hundreds of Twitter bots were finding new ways to look at the stream of consciousness spewing out of the microblogging and social networking site, which at the time was only 6 years old. Even then, Pentametron stood out, gathering more than 24,000 followers, garnering media attention, and eventually landing a book deal.
“Some people might think this is making fun of people online and how they talk,” says Nick Montfort, a poet and professor of digital media at MIT. “Pentametron was more complex. It’s taking this metrical framework — this tradition of the rhyming couplet and iambic pentameter — and combining it with the colloquial ways in which people speak. It does some work to elevate their language and bring it into this interesting juxtaposition.”
In other words, Pentametron was not only entertaining through Twitter, but it was using the site as a lens through which we could see the artistry in how we casually speak to one another — whether we realize it or not. And while Bhatnagar shut down Pentametron in 2019, it still stands as an example of both the limitations and potential role of AI in art: at least for the time being, machines can’t recreate an art form, but they can unearth the beauty that humans are unwittingly creating on our own.
Iambic pentameter has roots in Latin, French, and Italian verse. But thanks to poets like Geoffrey Chaucer and, of course, Shakespeare, the form is a common feature of English literature, especially verse. As recently as the 1970s, studies estimated that more than 75 percent of all English poetry since the days of Chaucer had been written in this meter.
That artistic tradition, along with its haughty-sounding moniker, is probably why iambic pentameter has taken on such a high-brow aura. Strange, then, that it turns up regularly in our everyday communication. “I’m conditioned to hear the rhythms beyond my study of the actual literature,” says Peter Maber, senior lecturer in English at New College of the Humanities at Northeastern in London. “You start to think and dream in iambic pentameter. Undoubtedly, Shakespeare has fed so much into the English language this way.”
Stranger still that this latent creativity can be found on Twitter, which many people have come to regard as the gutter of human discourse. But in fact, despite its current reputation for political arguments and mob rule, Twitter has long been a hub for artistic experimentation and sharing. In the 2010s, the platform welcomed an entire community of bot-creators that did everything from combining Google News headlines into amusing phrases to pairing raw outer-space images from NASA’s database with poetic computer-generated text to a program that simply tweeted out every word in the English dictionary in alphabetical order.
Another day, another train delayPentametron
I almost smoked a cigarette today
Pentametron came before any of those, though it was hardly the first artistic tool for Bhatnagar. A computer science major, he began his career as a web programmer and designer, producing sound sculpture and interactive exhibits on the side until he began pursuing his art full-time in 2007. His works included Stone Song, a 7,500-pound outdoor sculpture in New York that ran pressure sensors from piles of rocks into a droning music synthesizer to produce a range of sounds that shifted as the stones settled. He worked with a group of artists in New Orleans to turn salvaged wood from a 250-year-old building into several tiny musical houses, called “The Music Box: A Shantytown Sound Laboratory,” where even the floorboards functioned as a ramshackle keyboard. And on a much smaller scale, Bhatnagar made himself “invent” a new musical instrument every day for almost 10 years. “It forced me to rethink my ideas of what an instrument could be,” he says. “I started handmaking flutes out of wood, but eventually I had to branch out into software, more conceptual instruments based on some interesting sounds found out in the world.”
One of those found sounds was human speech. This sparked the idea of creating a piano-like speech synthesizer. In 2012, he recorded himself saying commonly used English syllables and then assigned each sound to a key. In this way, he theorized, he could produce sentences in sheet music form, made of notes representing syllables instead of actual tones. Any accomplished pianist could sight-read it without knowing the words ahead of time. Bhatnagar was not particularly skilled at the piano, and therefore, his played sentences were choppy and robotic. But when a friend who was a professional musician played the piano, the speech was practically fluid. That got Bhatnagar thinking about stresses in the way we talk.
Bhatnagar had always enjoyed the “exquisite corpse” poetry-writing game, invented by French surrealists in the early 20th century: each writer adds a few words to a folded sheet of paper, not knowing what the others have written. (The name comes from a line they produced: “The exquisite corpse will drink the new wine.”) Meanwhile, the programmer in Bhatnagar was fascinated by his discovery of Twitter’s application programming interface, or API, which enables programmers to tap into the platform’s systems and essentially turn on a “spritzer” of one percent of all public tweets. “I thought, ‘What does language look like when you take a random sampling and run it through a constraint?’” says Bhatnagar. “What if I could analyze their rhythm and pick out the ones in iambic pentameter? And within a couple of days, I cobbled something together.”
Pentametron worked by scanning about 500 tweets per second and cross-referencing them against the Carnegie Mellon University Pronouncing Dictionary, which could identify the stressed syllables in each word. The program then isolated the lines that contained 10 syllables, searched for pairs that rhymed, and tweeted out the resulting couplet. Perhaps amazingly, Pentametron found enough iambic pentameter in the wild to compose between 20 to 30 such posts every day. Of course, the topics sometimes reflected major happenings like the Super Bowl, Wrestlemania, or celebrity gossip, but more often the subject matter was more banal, like doing homework or what someone ate for dinner. The program was designed to skip longwinded racist rants or political screeds that had, even by then, started to make the platform toxic. “People don’t generally talk about deep things on Twitter,” says Bhatnagar. “And when they do, the tweets weren’t long enough for Pentametron to pick up.”
But the beauty was in the banal. Pentametron highlighted the artistry latent in our casual communication, showcasing not just the language itself, but the basic splendor commonplace in our everyday lives, actions, and emotions. Some of its couplets read like random thoughts spoken out loud, or like strangers talking over each other (“Another day, another train delay” / “I almost smoked a cigarette today”). Others had little relationship beyond their cadence and rhyme (“He really disappointed me tonight” / “Black people pulling for the Falcons, right?”).
Occasionally, two voices found harmony, matching up in a couplet that almost reads like call and response, a random and brief connection in the ceaseless chaos of the Internet: “I’ve never witnessed such a lovely sight” / “Another day, another lonely night.”
In 2019, after seven years and tens of thousands of tweeted couplets, Pentametron crashed. This was not entirely unusual, but this time, when Bhatnagar went to reboot it, the program crashed again. The creator paused and thought about how he had come to engage with his bot less and less in recent months. He also contemplated Twitter, and the increasingly vile and noxious place it had become. He decided not to resuscitate his creation. “It was a bit sad to me to shut it down,” says Bhatnagar. “I enjoyed it and was proud of it. But it was also a bit of a relief to have an end date.”
Bhatnagar put his art career in storage and focused on his day job in the offices of an electronics manufacturing company. He has only recently started to revive his artistic ambitions, this time in glassblowing, art neon lighting, and light sculptures. But Bhatnagar has kept his eye on the state of algorithmic poetry in Pentametron’s absence.
The evolution includes GPT-2, an AI language program that has learned to complete a piece of text given only a prompt of the first few words, thus enabling the program to imitate the styles of famous poets like Emily Dickinson. Back on purely iambic meter, Deep-speare is the creation of three machine-learning researchers and a literature scholar that uses a reserve of 2,700 sonnets to compose its own Bard-like verse. Google has joined the effort with 2021’s Verse by Verse, which takes the first line of text from a human and completes the quatrain. These programs and others have built on what Pentametron accomplished and inched machines forward in the pursuit of AI-created art.
But they haven’t really succeeded at writing literature. The lines they generate are clunky, awkward, or downright nonsensical. More importantly, they usually fail to tap into and evoke satisfying images, feelings, or meanings. In other words, these machine-made creations aren’t quite art — at least not yet.
Bhatnagar is skeptical of claims that some AI-generated language is indistinguishable from human creation. “But it’s possible there will someday be intelligent systems,” she says. “Right now, AI as a source material is very interesting and very promising.”
In 2020, a writer in The New Yorker described Pentametron as “more poetry as collage than true composition.” This is true, but also misses the point. The value of a program like Pentametron is that it sees AI not as a writer, but a tool, identifying the poetry hidden in the simplest human interactions.
So what does Pentametron say about English speakers today? Is iambic pentameter so ingrained in our culture that it has seeped into our subconscious and shaped the way we learn to communicate? Or is there something inherent in the meter itself that makes it a more effective way to use English? In other words, has the art always imitated the language?
“It cuts straight to the heart of the biggest debates in English,” says Maber. “Whether conscious or not, maybe rhythm gives added importance to our words. Perhaps the most successful speakers, writers, and tweeters might just have a better knack of rhythm.”