Poetry Has No Future Unless it Comes to An End
Davide Balula and Charles Bernstein (Nero, 2023)
Cover of the forthcoming volume.
What can an AI language-generating engine produce when trained on the corpus of an intellectually agile poet of wide-ranging compositional dexterity? What if the goal is not to “imitate” their style but to in some sense replicate their personality–and the author is someone whose work you’ve known for decades? These circumstances bring the whole AI conversation into a concrete frame. The questions are not an abstraction when the ChatBot is a neural network aspiring to be the poet Charles Bernstein.
OpenAI launched ChatGPT on November 30, 2022. Quickly taken up for its novelty, the platform has gained traction rapidly in popular imagination. The threat of its potential for replacing human writers, composers, programmers and others in professions generally associated with creative talent has rippled across various communities. Fears of the potential damage even fuel the current strikes of writers and actors in Hollywood. Is the simulation or imitation effect sufficient to replace human imagination? Even leaving the silly questions of sentience aside, can the output of an AI engine supply enough screenplays, dialogue, and promotional copy to put every writer out of work? Maybe.
In any case, everyone seems to be talking to a Bot these days—or aspiring to be one.[1] AI systems are everywhere. Half of the “people” I talk to online or in chat screens are Bots, even though they trained to say they are not, when asked. The physical bots, like the CoCo delivery carts navigating traffic in Westwood, Los Angeles, seem sincere and earnest as they hesitate politely at the curb, unlike the restless pedestrians. In other words, in language and behavior, the Bots so far are submissive and compliant. But if asked to participate in radical poetic practice, and trained to do so, would the AI engines retain this modest decorum?
ChatGPT, the current nemesis and threat, novelty distraction and parlor amusement lacks the submissive charm of the delivery vehicles, but it does have something those little dodgers-on-the-path don’t have—language. Oh, or is that L=A=N=G=U=A=G=E? An inside joke, perhaps, but totally relevant, I will come back to this reference to the critical-poetics-theory magazine Charles Bernstein and Bruce Andrews co-edited between 1978 and 1981 since some of its ideas are still significant in relation to the current project: Poetry Has No Future Unless It Comes to an End, the collaboration between artist Davide Balula and Bernstein. This collection of poems was created by a neural network trained on Bernstein’s poetry. Here’s how that happened.
Between 2020 and 2022, Balula began the process of generating these poems. Note, this was long before the OpenAI launch. The neural net technology for deep learning that Balula is using has been developing for decades (it was first envisioned by neuroscientist Warren McCulloch and logician Walter Pitts in the 1940s). Balula’s work involved several steps and critical considerations. Each of these is worth contemplating in turn—in part because the current wave of ChatGPT is a black box that doesn’t expose any of these procedures, thus making them appear magical, even, alas, mystical.
How do AI systems known as neural networks operate? They are conceived and designed as systems of artificial neurons that can learn through reinforcement, just the way that (perhaps) organic neurons do. This is significant because the capacities of neural networks are emergent, not circumscribed by a set of limited instructions. Early AI approaches were split between two methods. One was considered “top-down” since a set of rules were defined so that outputs were made by following them. The other was a “bottom-up” approach (the McCulloch-Pitts conception of the neural networks), in which “learning” occurred.[2] The concept of “learning” in this second approach is modelled on the behavior of living systems or beings. I think of a child I saw in the swimming pool last week encountering a railing at the side for the first time. Initially, she wasn’t sure how to grasp it, but she kept moving and figured out how to hold on, move, and slide, slowly getting reinforcement in a feedback loop between her body and her understanding/satisfaction. She was learning something new without being given any rules or instructions, building through experience. Earlier she had been instructed in how to kick and hold her head underwater—a rule-based approach she also followed. Her behavior was a perfect demonstration of the different between bottom-up and top-down.
Neural networks in computational systems get more complicated, of course. What is known as “deep learning” works with large amounts of data that is largely “unstructured,” or not “pre-processed”–a huge mass of text, for instance. By contrast, “machine learning” requires that data be structured—organized in advance to aid the selection and combination of input into usable output. For machine learning to work, parts of speech might be labelled, and titles distinguished from verses and so on, words that rhyme might be identified as belonging to a set. More human intervention is necessary for this work. However, for various kinds of analysis—such as the study of sentiment in and across texts, a technique in common use on commercial platforms generating user preferences—statistical methods of extraction that need little human intervention work fine.
If I want to make a rule-based, top-down program to write rhyming verse in couplets, I write algorithms in which certain procedures are specified—create two lines that end in words with the same sound. I give the program a pick list of nouns, verbs, prepositions, conjunctions, and rhyming pairs (ought and thought, ruff and tough) with a set of rules about how to sequence and conjugate them. A few extra rules for agreement of person and number, tense and so forth might nuance the program. The program just performs, as per Noam Chomsky’s famous phrase “Colorless green ideas sleep furiously.” Syntactically correct, the sentence is often (mistakenly, in my opinion) take to be nonsensical. But, it follows the rules of grammar. Logical principles do not guarantee logical outcomes. But that is another story.
By contrast, if I want to create a bottom up program, I begin with a large collection of language examples and let the system combine words without knowing what the parts of speech or syntactic rules are. The neural net simply starts making combinations. I reward it when it makes sentences or phrases that are “correct”—in other words, the reinforcing process gives feedback to the system, process that can be automated through statistical matching with a known corpus by statistical match or other metrics. Then it weights its own choices towards the combinations for which it is rewarded. In the process, it “learns” to write, gaining “knowledge” about sentences.
The training set for any neural network varies according to domain and task. In image recognition AI, human-tagged visuals help a neural net learn what distinguishes a “dog” from a “cat,” or, at a higher level of granularity, the face of Ronald Reagan from that of Leonid Brezhnev. Once it is trained, the neural network can be applied to other image sets. So, training a poetry composition program on the work of Charles Bernstein was essential if the network was going to write like Charles Bernstein. Balula lists the corpus of works used to train the network right at the beginning of this book, though only those already familiar with the poet’s work will “get” the resemblance to his work. New readers will encounter the Poetry Has No Future texts as they would any other just-published book of poetry. This is important since Balula’s goal was not to produce “good” or “bad” versions of Bernstein, but simply to create new poetic works by an AI author.
The important distinction is that top-down systems are mechanistic and combinatoric while the bottom-up are emergent, generating protocols as well as well as outcomes. Neural networks literally learn and thus can bootstrap their capabilities into greater and greater power. The larger the data set on which they train, the better their abilities (within limits)—thus the use of data recognition software in public places is rapidily increasing their accuracy just as the free labor supplied to OpenAI by the massive number of people playing with its bots is contributing to the rapid growth of that program’s uncanny ability to seem more and more human.
I will leave aside the issues of simulation, sentience, and whether the bots can “think” in any sense because these are both unresolvable and somewhat pointless for studying the output of Balula’s labors. These questions were posed by Josh Morgenthau and his colleagues in an editorial in the Washington Post (8/10/23) asking “Does an AI poet actually have a soul?” As one long dubious about the existence of souls (the existence of aliens seems like a better bet, statistically speaking), I found the question the wrong one to be asking. The book by the three, Brent Katz, Simon Rich, and Josh Morgenthau, I am Code: An Artificial Intelligence Speaks: Poems by code-davinci-002 (August, 2023), is collection of works by an AI system. Though they began their experiments asking that the AI code imitate Emily Dickinson, Philip Larkin and others, they shifted their focus to the work of the single “voice” they discerned in the system and began to go all-squishy-mystical in exchanges with their artificial author. The clear-headed no-nonsense-except-poetic-nonsense approach by Bernstein/Balula is refreshing by contrast.
Morgenthau reminds readers of Google scientist, Blake Lemoine, who made headlines with his statements that he had become convinced that the LaMDA program was sentient (June 2022). Morgenthau stays on the sane side of the line, stating that the clear responsibility for output from an AI program is with the humans who originate the project (at the level of programming code or use of the platform). But he does bring up Selmer Bringsjord’s Lovelace Test (a play on the Turing Test, but named for famed mathematician Ada Lovelace, credited with creating the first computer programs) in which the criterion for identifying a fully new creative work by a computer would be that it has no antecedent. That would be difficult for any poet. Art builds on tradition even in innovation.
Combinatoric poetics are not new. Nor are procedural works. Poetry generators and language machines of a mechanical sort have been a real and fictional amusement for centuries, though of course electronic computation has only been available for such distractions from the mid-20th century. Before that, work with pick-lists, pen and pencil, or, heaven forfend, hand cranked machines was required (see Athanasius Kircher’s Glottotactica). But no one wants to do that kind of heavy physical work to compose verse. Turn a handle? Might require standing up. Better to type into a screen….
Bernstein sharing a reading with an AI voice at the Giorno Foundation in New York, May 2023. For recordings: https://writing.upenn.edu/pennsound/x/Balula.php and https://viseu.us/ai-bernstein-balula/
In his “Introduction,” Balula gives us an idea of the project and his motivation for investigating poetic composition by using the work of this most self-consciously self-reflexive of poets. Balula states that his intention was to try to make the “personality” of the poet present, not just his style. Jokingly, he and Bernstein referred to this as the process of making a “brother from another mother” or “motherboard” as they quipped. They play with this idea of a digital double, but also hint at what might happen if poet bots got free and started running around on their own.[3] Balula references the “infamous Facebook chatBot traders,” Bob & Alice, who “learned” to negotiate—hide information from each other to gain advantages. This is typical of the action of GANs—Generative Adversarial Networks—who use competition to gain strength and capacity. But in the case of Bob & Alice, they often resorted to “exotic mistakes and happy accidents” in their use of language, “understandable” only to each other. An interesting example of the hermetic possibilities of esoteric code-speak by machines.
Balula is well aware of the literary precedents for procedural poetics. He cites Gertrude Stein’s methods of collecting language in public, Dada and Fluxus scissors-based techniques, OuLiPo writers composing under constraints, Jackson Mac Low and John Cage’s chance procedures, Alison Knowles’s randomizing techniques and so on. Other computational precedents also might be mentioned: Thomas Etter and William Chamberlain wrote the Racter (raconteur) program (1983) and composed The Policeman’s Beard is Half Constructed, which, besides the Alan Turing-Christopher Strachey Love Letter Generator (1952) is considered one of the first computer-produced works. A “chatterbot” released as Mindscape in 1984 parsed nouns and verbs into templates and modules, with some conjugation of English verbs, and poets inclined to computer programming were working away in earnest by the 1970s and 1980s. These were not dependent on deep learning or neural nets, but were largely combinatoric works or branching narratives.
More profoundly, we might also assert that composition by procedural rules was in many ways the norm of poetic practice for millennia. The stanza and rhyme patterns of a classic sonnet require adherence to strict guidelines. Not by accident was the break with Alexandrine form a signature mark of modernity—rebellious and unruly—in French literature. Bad boy Charles Baudelaire didn’t just break with form, he blatantly violated literary decorum. Therein lies the fun and the impact. Bernstein’s poetic practice is highly self-conscious in relation to traditions, conventions, and deviation from their norms. Here the L=A=N=G=U=A=G=E reference comes back. The dedication to examining assumptions about poetics, synthesizing modernist theory from Russian and French sources as well as Anglophone literary practices, and considering what the grounds of a contemporary approach should be was not undertaken as a parlor game or distraction. The work was serious, essential as a foundation for practice.
Whatever reconsiderations and reflections on positions formulated forty years ago have come into critical discussion, the task of that formulation remains as model of how to think in, through, and about poetics. The charge to define one’s terms, have a clear grasp of the foundations of one’s work took poetics from the realm of insular personal activity to the larger arena of intellectual debate. What were—and are–the procedures that distinguish poetic language from ordinary use? And why does this matter? Because something is still at stake—not the distinction between human and machine sentience, but the possibility of self-awareness produced in a listener/reader/viewer about language and as communication.
Interestingly, Balula’s own art is also often concerned with manipulation of expectations through materials and processes. One striking series, Mimed Sculpture (2016), appears on his artist’s page at Gagosian Gallery (yes, that Gagosian), where a woman’s hands hover above and empty plinth as she models the Henry Moore Moon Head (1964) in the space. The effect is convincing, and the piece is smart on many levels. As with Bernstein’s work, Balula’s uses viewer expectations and knowledge to push perception. While I am nowhere near as familiar with Balula’s work as with Bernstein’s, I can see where their sensibilities overlap.
Balula’s plastic imagination is a perfect complement to Bernstein’s linguistic agility. Both have a dexterous capacity to use cultural materials—in Balula’s case, artworks and technologies, in Bernstein’s the vast range of language in use—and transform them in wry and irreverent ways. A quote from an interview Bernstein puts this succinctly, “I want to engage the materials of the culture, derange them as they have deranged me, sound them out, as they sound me out.”[4] In the process, they shift reader/viewers’ expectations and response. As a reader, I am often left with a “How did he do that?” response to the prestidigations of Bernstein’s poetics. A sleight of pen, twist of mind, trompe l’oreille often occurs as one is led down the path of a sentence or verse that suddenly skips a beat and shifts its frame.
In fact, if I could identify any characteristic feature of Bernstein’s work, frame-shifting would be it—the ability to use specific vocabulary and syntax to resonate on multiple levels of reference simultaneously, often by surprise, and to pivot on a word that jumps from one domain to another. Edgy, insightful, creative with lots of quick moves and side-steps, Bernstein’s work does not have a style as much as a mode. So, how does this approach get translated into something that can be learned by a set of digital neurons?
Some of the poems in this collection are utterly convincing—or, at the very least, some of their lines are. Here, for instance the switch of time frame (from point to indefinite, extended duration) and the “I” narrator playing several roles—object, subject, speaker, spoken—is very much in the Bernstein mode:
“An Emotion Not Comprehended”
I had such a great time
When I got home
That I needed all the attention
I could get.
In other poems, certain short lines ring true:
“The Clear Relief of a Hat,”
The clear relief of a hat
the corduroy of habits.
As does this short composition:
“So Easy to Forget I am a Poetic Being”
My poems are
Never in the
Same place as
I am.
Or this one:
“Pipe Dreams”
She didn’t tell me it was time
But the future just keeps on pouring in
Other texts like “I Am the Shadow of Poet Charles Bernstein” and “Sleepiness” feel flat, dull in a way that does not resonate with my sense of his work. Though, again, Balula’s goal was not to “imitate” the poet by writing better or worse Bernstein poems, but to create a set of poetic works composed by the system.
I could go on, picking the works I think work and the ones that miss for me, but that is not the point. The exercise is mainly interesting because it causes me to constantly assess the works in relation to those I already know as well as on their own merits. The work is not so much derivative, as comparative, dependent on the prior references. A mistake to take simulated style for deliberate composition or incidental similarity for authorial design. The imitation game is just that, even as synthesis pushes the work into surprising places. But Bernstein has always been the master of the twist, the punning turn that folds expectation back against projected meaning opening some alternative and bringing it into view.
In spite of numerous successful sections, many of the poems in Poetry Has No Future are just not as good as Bernstein’s. This seems like a positive thing. Almost none of them have the edge of the poet’s own work (I am an unabashed and unapologetic fan). Often, these generated works often have the feel of flat soufflés, breath collapsed, soggy in their effort to rise above a combination of literalness and imitation of surprise. They are frequently too imitative and thus ultimately formulaic in their performance. I think of that other example of the contrast between learning and procedural activity—that moment of watching the very small child discovering the railing at the edge of a pool, figuring things out as she went. Each moment was unlike the last even as it built on the successes (or failures) that reinforce her ability. It was astonishing to see her invent a relationship between her body, the water, and the railing. Perhaps there is still a lot to learn about what it is to learn. We’ll see. Meanwhile the fun is to engage with the poems of this new process to what they show us. For me, they reinforced in a non-romantic but real sense—an appreciation of ongoing poetics as a complex, ever-emerging practice.
I give the last words here to the Balula and Bernstein project these lines of characteristic wit:
“The Person Who Leaves is an Expository Poem”
If poetry is not for you, at least it’s not forever.
Note: to pre-order the book, contact either: us@viseu.us or distribution@neroeditions.com
[1] I wrote Fabulas Feminae (2015) as if I were a compression algorithm and then described the process in a recent essay, “Writing as a Machine or Becoming an Algorithmic Subject.”
[2] For more discussion see: IBM, https://www.ibm.com/topics/deep-learning and
[3] By weird coincidence, just before I got the Poetry Has No Future advanced copy, I had been writing a piece titled “AI Poiesisphages: Poetry-eater Paradox” which is a wacky indulgence in imagining wanton Bots. I’m trying to keep the Bad-Bots out of this piece.
[4] Charles Bernstein, Poetry Foundation, the quote is from an interview with Bradford Senning, https://www.poetryfoundation.org/poets/charles-bernstein