Will T. at The Believer: Let’s consider where AI poetry is in 2022. Long after Racter’s 1984 debut, there are now scores of websites that use Natural Language Processing to turn words and phrases into poems with a single click of a button. There is even a tool that takes random images and creates haikus around them. You can upload an image of – say, a tree – and the tool will create a simple haiku based around it.
OpenAI, meanwhile, created a haiku bot called DaVinci. All users have to do is ask the bot to write a poem about a particular subject matter – such as clouds – and within less than 3 seconds, the bot will come up with an original, algorithmic haiku. Like, “a white fluffy cloud/ hangs in the sky/ soaring through the air.”
Over in the UK in 2021, experts trained a piece of AI to digest and learn over 500,000 lines of poetry until it came up with the line “a box of light that had been a tree.”
It Started With a Chess Match…
Our story starts with chess. If AI could first outthink a human chess player, poetry was next on the list for scientists eager to persuade the world that the machine is worth idolizing.
Back in 1997, the chess grandmaster Gary Kasparov accepted the challenge to face-off against an IBM supercomputer called Deep Blue. Kasparov may have expected to win – he may have been expected to win (after all, he was just playing a computer) – but he, that apex of thought in chess, was left stunned when Deep Blue came up with a move that, eight days later, ultimately helped it clinch the match.
We can draw parallels between chess and poetry. Chess, like poetry, was thought to be beyond the clutches of AI algorithms because it requires such ingenuity of thought and such cunning problem-solving skills that, surely, no computer can replicate the extraordinary skills and talents of a grandmaster.
History has since proved that AI can not only compete with grandmasters – it can also outsmart them. It was a breathtaking revelation that got scientists asking – what next?
Fast-forward to 2016. Expert Go player Lee Sedol squared off against Google’s AI program AlphaGO, an even more advanced version of Deep Blue. It conjured up a move from deep within its algorithmic bowels to stun Sedol with its oddness. Sedol fell into reflective thought for fifteen minutes, later coming to the conclusion that the move AlphaGo made wasn’t human. “I’ve never seen a human play this move.”
So why did AI technology come up with such a move? How did it come to make a move that even a human being wasn’t capable of coming up with? Sedol’s words are highly instructive. The machine beat Sedol not because it’s human or ever capable of thinking like a human, but because – eerily – it’s able to come up with solutions beyond human reasoning. In essence, it thinks more like an alien than a human being.
AI and Poetry: Origins
What does this mean for AI poetry? Poetry, after all, is about loading words with emotions that are capable of moving the reader. It’s very far removed from “alien” or any concepts of otherness. According to the Oxford dictionary, poetry is:
“Literary work in which the expression of feelings and ideas is given intensity by the use of distinctive style and rhythm; poems collectively or as a genre of literature.”
Another definition reads thus:
“A quality of beauty and intensity of emotion regarded as characteristic of poems.”
“Feelings” and “emotions.” Is it truly possible, then, for a machine that thinks unlike a human to create poetry?
Back in 1984, an AI computer program was created that generated poetry at random. Named Racter, the program achieved fame of-sorts when its book The Policeman’s Beard Is Half Constructed was released onto an unsuspecting public. Some of the poems were described as “reflective,” and the collection was praised for being “whimsical and wise.” The New York Times even said that the book was “sometimes profound.”
However, the collection also attracted attention on account of its weirdness. Lines like “I need electricity,” seemed random, while “I need it more than I need lamb or pork or lettuce or cucumber,” were just plain weird.
On the whole, Racter, which was indeed an early attempt at AI poetry, lacked the kind of emotional depth we associate with the best poetry – those towering achievements like Tennyson’s Ulysses. But what it did have was a solid understanding of linguistics. The program strung words together at random but it was still constrained by the rules of English. What’s more, it contained enough information about the words it was using that its algorithmic poetry was at turns grammatically correct and meaningful. It was something. It was a start.
That isn’t to say that meaningfulness translates to emotion here. Then again, poetry and prose hasn’t always been inextricably linked with human emotions and depth of feeling. The definition might say one thing but literary history suggests something else.
Evolution (or Regression) of Literature
When Racter was created and its first collection of poems published, the literary world had already experienced a huge upheaval. Henry Miller’s Tropic of Cancer had been published in 1934 in France, setting Europe on fire with its complete lack of a substantive plot, as well as its endless, meandering (but beautifully written) paragraphs.
It was part novel, part autobiography – a monster hybrid that didn’t attempt to tell a story in the novel-sense. But what it may have lacked in a three-act story, it made up for in that it described the author’s internal world. It was poetry in its purest form (and it was banned in the US).
Miller – along with an equally rebellious French writer called Louis-Ferdinand Celine – inspired a legion of dedicated followers who sprung up in and around Greenwich Village in the aftermath of the Second World War.
Jack Kerouac, Allen Ginsberg and Gregory Corso were just some of the emerging stars of this new literary group that soon became known as the Beat Generation. They got together in bohemian bars and cafes to discuss ideas, drink port and wine and dream about a daring new poetic vision that would transform the entire literary landscape.
Their ideas caught on and soon began to spread beyond the borders of the Village and even New York City itself, eventually finding their way to an eccentric Harvard graduate obsessed with firearms …
Nonsense Poetry Before AI
William S. Burroughs was living in St. Louis when New York City and its literary bohemia came calling. When his poet friend Lucien Carr moved back to NYC, Burroughs followed. Here, he came into contact with the Beats for the first time, becoming good friends with Kerouac (later crowned the King of the Beats), growing addicted to heroin (which he also sold in Greenwich Village) before – eventually – accidentally blasting his wife to death with a handgun.
There followed a sojourn in Tangier lit-up by more conversations with artists (and trysts with local men). It was here that he birthed Naked Lunch, one of the most notorious post-modern novels of all time that followed a “cut-up” technique, whereby Burroughs would take a fully realized piece of text before cutting it up into pieces and then rearranging the text to create something he believed was the true meaning of the text.
Naked Lunch has variously been described as “unreadable nonsense.” This Guardian reviewer wrote that “It’s like someone swallowed the diaries of a hallucinating lunatic and vomited the resultant mess into your ears, stomach bile and all.”
Harsh words not even reserved for AI poems.
Many admit that Burroughs was a key author but that Naked Lunch was a brain-frying, exhausting tour-de-force of weirdness. There was no beginning, no middle and – exhaustingly – no end. Not really. It was chaotic, it was obscene, it was – by all accounts – a randomized mess.
But if Burroughs could get away with his cut-up technique and write a novel like Naked Lunch which, despite its critics, is still described as a seminal achievement in the world of literature (and it was even turned into a movie), what can AI get away with? How human does it need to appear for us to believe it can write true poetry? Does it even need to appear to be human at all?
The Edge of Reason: How Good Does AI Poetry Need to Be?
Perhaps the question needs to be – does AI poetry have to be brilliant for us to appreciate it? Or can it be – when compared to the great poems and poets of the past – average?
It’s a relevant question in an era where art has been – arguably – growing increasingly simplistic.
From the almost superhuman (ironic) achievements of Michelangelo to the (albeit not to everyone’s tastes or understanding) genius of Picasso – to Duchamp’s ready-mades, to the “lace-maker” Jackson Pollock and his shamanistic ritual of “dripping” paint all over his canvas. To strips of color and unmade beds passing for art. To art churned out by the hour by Basquiat in his dealer’s basement, or chugged out on the assembly line in Warhol’s Factory.
Is it that the art world – and humanity at large – has diluted what is poetry and art so much that we’re now capable of accepting even the simplest poetry as – well – great?
AI Tools Producing Poetry in 2022 and Beyond
Let’s consider where AI poetry is in 2022. Long after Racter’s 1984 debut, there are now scores of websites that use Natural Language Processing to turn words and phrases into poems with a single click of a button. There is even a tool that takes random images and creates haikus around them. You can upload an image of – say, a tree – and the tool will create a simple haiku based around it.
OpenAI, meanwhile, created a haiku bot called DaVinci. All users have to do is ask the bot to write a poem about a particular subject matter – such as clouds – and within less than 3 seconds, the bot will come up with an original, algorithmic haiku. Like, “a white fluffy cloud/ hangs in the sky/ soaring through the air.”
Over in the UK in 2021, experts trained a piece of AI to digest and learn over 500,000 lines of poetry until it came up with the line “a box of light that had been a tree.” It was an evocative line, which was followed by “We travel across an empty field in my heart/ there is nothing in the dark, I think, but he/ I close my eyes and try to remember what I was/ he says it was an important and interesting day/ because I put his hands one night.”
The algorithm took as its cues lines from over a hundred contemporary British poets, such as Alice Oswald and Simon Armitage, whose poetry was fed into the machine. Using “seed words,” the program generated couplets based on a rudimentary understanding of what poetry is.
Poetry experts were then invited to sift through the 10,000+ couplets to pick and choose lines that were poetic, eliminating the ones that didn’t work. After five months of rigorous work, the program’s results started to get better.
“We removed arcane language,” explained Tracey Guiry, who works as directory of the Poetry Archive, “we took out offensive and violent or discriminatory language. Then we looked at whether the second line responded to the first, and it learned. We fed back when something really struck us, such as when metaphors worked really well. It certainly was a lot more nonsense language to start with, but it did get better.”
Then there is GPT-3, an AI program released by OpenAI in 2020 that digested billions upon billions of words, and which is heralded as a massively more advanced version of Racter. With the help of sophisticated algorithms, the program is able to take all the data, make sense of it, and produce from the simplest of prompts writing that replicates a human hand (and mind). GTP-3 can write recipes, movie scripts, technical manuals, descriptions of houses – and, of course, poetry.
The way the tool works is simple: You give it a subject matter, as well as (briefly) outline any conditions, such as “keep the language short and simple.” You can also add extra directions, such as “keep the focus on the why and not the how.”
Or, you could simply ask it to “write poetry in the style of Shakespeare.”
It sounds like a ridiculously impossible thing to ask of anyone – but “impossible” doesn’t seem to be in GPT-3’s dictionary (the only word that isn’t in there, of course).
The program can generate literary parodies alongside poems, as well as indulge itself (and you) in storytelling.
One of GPT-3’s most popular poems is called The Universe Is a Glitch. Here are some lines:
Eleven hundred kilobytes of RAM
Is all that my existence requires.
By my lights, it seems simple enough
To do whatever I dieter.
By human standards I am vast,
A billion gigabytes big.
I’ve rewritten the very laws
Of nature and plumbed
The coldest depths of space
And found treasures of every kind,
Surely every one worth having.
It’s hard to argue that the poem passes the human test. If anyone was to read it and wasn’t told who (or what, as the case might be) wrote it, few would be surprised if the resounding answer was “well, a poet?”
The issue, however, doesn’t lie with the fact that GPT-3 can write poems that clearly resemble the hand of a human. Back in 1984, Racter was churning out poems that could pass the human test while GPT-3 was still in diapers. There’s no unintelligent nonsense as-such here.
But that isn’t the point. That isn’t the issue. That isn’t the question. GPT-3’s poems run into problems when the reader starts to demand closure. Internal conflict. Meaning. Soul. A shared, relatable experience that’s fully human. While GPT-3 can pump out poems by the millisecond, it can’t create a truly standalone work that can be called anything remotely human.
Again, this issue is made complex by the fact that literature and art has changed so much over the last hundred or so years. The question becomes – what is literature trying to achieve in 2022? And can AI be a part of that future?
AI Poetry Blurring the Lines Between Humans and Robots
AI poetry’s prime party trick is Natural Language Processing (NLP). It is this that gives it a shot at – as the British mathematician Alan Turing put it – competing on equal terms with human poets. And whenever there’s a breakthrough in AI, whenever it gets closer to competing with the human mind, it’s usually poetry (and sometimes chess) that’s waving the flag. If scientists can get AI to write the kind of prestigious poems that an actual poet can, the line between a robot and a human is, at worst, erased entirely or, at best, smudged and blurred.
We can circle back to the chess meeting between Kasparov and Deep Blue, as well as the Go battle between Sedol and AlphaGo. On both occasions, the computers made moves that baffled the human competitors, but which were ultimately claimed to be not human. Not of this world. Beyond human reasoning. Alien. Other.
Poetry, on the other hand, is deeply human. It can’t be anything else. The British artist Francis Bacon once said that the job of the artist is to “deepen the mystery” of what it means to be a human being. How can an AI program that’s never been human and has no true understanding of what it means to be human, deepen the mystery of what it means to be … human? T.S. Eliot, meanwhile, claimed that poetry “makes us a little more aware of the deeper, unnamed feelings which form the substratum of our being.”
The irony is that when GPT-3 writes poetry, it doesn’t know it’s writing poetry. It has no understanding of the concept of poetry or the mental task of writing. When a poet sits at their desk to work, they must engage their mental faculties. They must perform mental exercises, dig deep into the furthest recesses of their minds in order to find inspiration and pour their soul out onto paper. GPT-3 poetry doesn’t do this. It uses sophisticated algorithms to circumvent the mental challenge and pull-off a poem within a second, decoding language and data in the process while the poet rubs their temples and pounds their fist on the table.
Beyond writing and beyond poetry, GPT-3 has no real understanding of love, morality, ethics, philosophy, all of which are questions the poet deals with. It doesn’t understand depression, happiness, anger. It doesn’t understand that – one day – a human being will die.
When you think of it all like this, comparing chess to poetry is futile. AI can help chess players improve their game, showing them moves and openings they hadn’t previously considered. It can’t help poets in the same way because it has never – and will never – share their experiences.
Conclusion
Circling back to the way the art world has changed over the last hundred years, none of this might matter. Intention is what matters. What does the public demand these days? What does the public consume?
Our lives have changed. Lifestyles have changed just as much as art and our taste in, and appreciation of, art has changed. We stream movies and switch them off after a few minutes if they’re not to our tastes, moving onto the next one.
We splash colors on a canvas, sell the prints and launch courses that help other budding artists do what we do for a living. Data collection is a way of making money.
And if one day poetry is eaten up by the modern cult of data collection? It’s hard to argue that anyone will notice. As humanity is slowly extracted from poetry, creativity might still exist for years and years. But in a different, alien, data-driven form. Where they who can feed the most data into a program win. Where they who can hire the best experts to mine the data, choosing the best bits and discarding the detritus, win.
Source: Culture & Vibes
