Machine Voice: Programmer Fiction

IN THE EARLY 1960s, a South African computer programmer was living miserably in Britain. He was working in his first job at IBM, where he had imagined doing something significant, like “translating symbolic logic and set theory into digital codes.” Instead, he spent monotonous days talking about inventories and outflows.

He soon left IBM to work for rival company ICT, where he joined the team that was developing the “supervisor” for the Atlas 2 computer housed at the University of Cambridge. This new technology, designed to allow multiple tasks to be undertaken simultaneously, was vital to using the resources of the machine more efficiently. As scholar Rebecca Roach records, the programmer’s work with this operating system was focused on “conceptualizing the design and writing the code that would instruct the computer about which jobs to implement and in what order.” While this work was more interesting, he was still dissatisfied—and worried too about the use of the technology for the British weapons program.

A reader of modernist literature, he found distraction by working on a masters thesis on Ford Madox Ford. He recounted spending his Saturdays in the Reading Room of the British Museum as well as two evenings a week when the museum closed late. However, he found the achievement of Ford’s writing to be limited, and thought the later works tailed off in quality. It was only The Good Soldier (1915)—with its concision, minimalism, and craft—that stood the test of time. Drawing on the language of computation, he declared that this one book was “probably the finest example of literary pure mathematics in English.”

The programmer’s interest in literature soon flowed back into his work with computers. His personal archives show him trying to write poems using the machines for which he was writing code. He developed programs that, once compiled and executed, drew randomly—albeit in a set order—from indexes of words he had created. The computer then printed the results. The output was a series of lines from which he would make hand selections. As Roach has shown, these experiments spanned more than a decade and several different computing architectures. The programs were written at different levels of abstraction, from machine code to FORTRAN to what is known as “pseudocode”—nonexecutable coding shorthand.

While the outcomes of these experiments did not exactly lead to great poetry, some of them were ultimately published. “Computer Poem” ended up in the University of Cape Town’s student magazine in 1963. A more ambitious work, “Hero and Bad Mother in Epic, a poem,” was featured in the first issue of the South African anti-apartheid magazine Staffrider in 1978. In this “epic,” terms like entrail, seaborn, drowsy, nude, and casino are combined and recombined in unexpected ways. A narrative of a heroic “philatelist” seeking a matriarch emerges from its odd, dreamlike phrasing, demonstrated in the poem’s opening lines:

dusk seeps up the entrail of the seaborn nude 

the vegetable sleeps in its circle 

the bedroom drowses 

the casino is swathed in tidal melancholia 

the nude awaits the hero

The programmer left the computer industry in the mid-1960s to undertake a PhD in English literature at the University of Texas at Austin. His dissertation sought to analyze Samuel Beckett’s literary style through the quantification of language, an approach known at the time as “stylostatistics.” The work had also been made possible by advances in computing, and by the Chomskyian revolution in linguistics, which in the 1950s and ’60s had proposed that there is a logical engine—much akin to an algorithm—embedded in human brains. This, Chomsky suggested, drives the production of language. The programmer’s statistical approach to literature literalized the metaphor of the logical engine; it deployed computational methods to unearth the organizing structures of Beckett’s fiction.

The project was a failure in the programmer’s eyes. He found that he had done little more than measure “an aspect of our response rather than an aspect of the text.” His stylistic analysis was, he concluded, every bit as impressionistic as other ways of reading fiction, even if it was now cloaked in a forbidding technical language. Several decades later, in an interview, he reflected on the attempt to read and write with computers. It was a “wrong turning,” he said, “a false trail both in my career and in the history of stylistics. It didn’t lead anywhere interesting.”

In the context of recent breakthroughs in generative AI, though, the attempt to make computers read and speak appears not so much foolhardy as prescient. The technology that underlies large language models (LLMs)—for example, ChatGPT—is the neural network, which is modeled on the human brain. Trained on billions of parameters, neural networks can identify patterns that are likely to be invisible to human observers. The wealth of training data, provided in particular by Common Crawl web archiving, along with sheer computing power, was unavailable in the 1970s. Instead, in what is often called an “AI winter,” there were numerous well-funded but largely unsuccessful projects that relied on rule-based “symbolic” approaches. (Contemporary approaches, by contrast, work by not knowing anything about the underlying structure—the grammar—of the training data.)

The miserable programmer whose work I have been describing is J. M. Coetzee, who across his 15 novels and three volumes of autobiography would become one of the most significant writers of the last 50 years. His poetry programs from the 1960s—all of which featured randomly generated words—were from a period in his life when he felt himself unable to write. So, too, his stylistic analyses. Their purposes were to help break him out of the cage of subjectivity and language’s determination to naturalize itself, and instead to create the possibility of producing something truly new.

His experiments in computing literature led to metaphors of programming that appear across his early work. In “The Vietnam Project” in Dusklands (1974), a supervisor named Coetzee delivers shadowy demands to his employee, Eugene Dawn. “There remains the matter of getting past Coetzee,” Dawn writes, after his manager has told him to revise his essay. The hero of the computer-generated poem, “Hero and Bad Mother in Epic” is a philatelist who exists one level above the token (the stamp). He makes his way through signs drawn at random (from indexes). The poem dramatizes the function of the computer program as it works through the many possible arrangements of words and outputs them to tape.

Well after he had left computational research, Coetzee would describe the creative process as a battle against language’s tendency to reproduce itself, meaninglessly. “Writing,” he said,

involves an interplay between the push into the future that takes you to the blank page in the first place, and a resistance. Part of that resistance is psychic, but part is also an automatism built into language: the tendency of words to call up other words, to fall into patterns that keep propagating themselves.

This “automatism,” as he describes it, is recognizable to us today as the process by which generative AI produces language. It is how large language models create sentences that feel as though they are natural.

For Coetzee, though, automatism in fiction was the stuff of pedestrian art—not least his own. In Youth (2002), he recalls how he avoided confrontations with writing as a young man living in Britain:

He cannot begin writing until the moment is right, and no matter how scrupulously he prepares himself, wiping the table clean, positioning the lamp, ruling a margin down the side of the blank page, sitting with his eyes shut, emptying his mind in readiness—in spite of all this, the words will not come to him. Or rather, many words will come, but not the right words, the sentence he will recognize at once, from its weight, from its poise and balance, as the destined one.

This is not quite writer’s block, where he simply cannot produce language. Instead, it is the condition of having many wrong words (born of automatism), but not yet the right ones, the ones not yet said.

The problem Coetzee faced was how to overcome the automatic flow of language so that he could find his own way of speaking. As his archive held at the UT Harry Ransom Center shows, his response was at least as much material as it was abstract. In an age when typewriting predominated, he composed largely by hand on lined paper. (While later he did increasingly make use of word processing, as David Attwell records, it was not until 2005’s Slow Man that the computer truly took over.) Drafting in longhand must have been inordinately time-consuming. He wrote the same sentences many times in the course of his multiple versions of a draft. Foe (1986) went through 11 handwritten versions over two and a half years before the first printout—only to then undergo further revisions.

Coetzee conceived of his writing process as its own kind of compositional art. For him, it seems that automatism cannot take over if one pays sufficient attention to the labor of writing itself. His commitment to this lifelong work of aesthetic creation was extreme. He bound together some of the pages of his drafts and put them between cardboard covers. He dated his notebooks meticulously and included reflections on his works in progress. He even made his manuscript revisions in different colored pens and included the revision dates. Process was all—and it was in process that, for him at least, the truth of the fictions began to emerge.

It is this labor that above all else gives Coetzee’s writing its sense of craftedness. As a doctoral student in his archive, I found myself puzzling over the additions and deletions of single words—a practice I undertook because Coetzee clearly had too. Why, I wondered, did he change “inherited” to “won” and then back again to “inherited” in the first sentence he wrote of Dusklands? What improvement did his revision to Disgrace (1999) make, when he modified his description of David Lurie from “someone of his age and temperament” to “a man in his fifties, divorced, alone, not particularly well-off”? I could imagine his copying over each word, each sentence, and testing it for better alternatives.

Through this laborious longhand method, where nothing came simply, Coetzee found his way toward a space where he felt he could finally speak for himself. In a remarkable passage in the notebook for Foe, he asks: “[W]hat is the whole thing about? […] The only figure I can generate anything but puppetry out of is myself. When am I going to enter?” Elsewhere, he reflects that he always writes “best out of an adversary position.” From that position he has the capacity, he feels, to unleash something not yet said, and perhaps true, if only he can find it. He is an adversary not only to the public and critics (who want his books to be a certain way), but also to the already said, to language, and ultimately to himself. He cannot enter his work in progress because he has not yet discovered how to speak: everyone (and everything) else has been too busy doing that for him. “There is a true sense in which writing is dialogic,” Coetzee once said. It is “a matter of awakening the countervoices in oneself and embarking upon speech with them.”

What was once a metaphor, or simply a dream, has now become a reality with generative AI: the machine really can generate the automatic language that comes out of our mouths. The language of society at large is newly intimate, waiting for the prompt that will accelerate through the global spectral unconscious of human expression before we make it our own. The crungus, that mythical beast that lies buried in the deep structure of the network, awaits our call—and haunts our imaginations. This is what Coetzee resisted in his writing, not least because, as he saw in apartheid South Africa, habits caught from language are what keep people in chains. His archive testifies to his work of refusal, and the results are clear in his prose. The sparseness, even aridity, of his style is a consequence of setting himself against self-replicating discourse. What makes his struggle distinctive, then, is how he resists language’s tendency to follow existing pathways—the neural network—so that he can find what he does not yet know he is looking for.

Literary composition will fundamentally change with generative AI. The impact will likely be even more significant than were other shifts, including the shift to word processing. As Matthew G. Kirschenbaum describes, a technical consultant in the early 1980s named Peter Rinearson, who had advance access to Microsoft Word, “believed that [the program] would have a profound impact on society, because—like the typewriter and the pen before it—it would change people’s relationship to language.” If this seems overstated today, it is only because we now live in that world. Kirschenbaum concludes that “[t]here is no question that the popular advent of word processing around the year 1981 was an event of the highest significance in the history of writing.” The same is true of 2022, when generative AI came to be available to anyone with an internet connection.

I grew up in the era of word processing, which already dates me to a particular time and place—before Google Docs but after WordPerfect. When I first used my father’s typewriter in his shop, sometime in the late 1990s, I delighted in thumping at the keys and pushing the carriage back to its return position. After several turns at writing sentences, I asked my father where to find the backspace key. He laughed. It was unimaginable to me then that one had to finalize the word before writing it.

The uptake of the new technology will not be uniform. Instead, multiple modes of composition will exist concurrently, from handwriting to offline word processing to AI-enabled software. This coexistence is already the case in every writer’s practice. In the writer’s studio, handwritten notes are stuck to computer monitors, pencil annotations in books lie next to the keyboard, and printouts are decorated with revisions in pen. The translation across media and versions, however, remains unique to each writer.

Software is likely to become more targeted to specific uses. Kirschenbaum ends his 2016 book by noting the “plurality of word processing platforms” that had emerged to compete with Microsoft on specific terrain, such as Scrivener and WriteRoom. This new range of software options returns us to the early era of word processing when there were many competitor products, such as WordPerfect and WordStar. Generative AI will be integrated into applications in ways that will seek to accelerate the writing process. In previous technological shifts, activities that were once performed by publishers were returned to the writer (such as copyediting). The same is likely to be true of generative AI. More structural editing and feedback, for example, will almost certainly be done by editing applications powered by LLMs.

However mystical it might sound, though, the struggle to find one’s own voice will remain as difficult as ever—if not more so. When Coetzee faced the blank page, he found that there were an infinite number of Coetzees trying to make their way into writing. The tendency of words, he said, is “to call up other words, to fall into patterns that keep propagating themselves.” The neural network forestalls the original creative act, in other words. Overcoming it, then, is at the heart of literary writing. For writers, working with generative AI is like drinking from a fire hydrant. Creativity will inevitably require inventive use of the technology. This may well include, as for Coetzee, practices of slowing down, noticing, and imagining in ways that seek to turn back the patterns of the already said.

Coetzee’s The Master of Petersburg (1994) finishes with the Dostoevsky character beginning to write. He has returned to St. Petersburg following the death of his son—a death that was very similar to the tragic death of Coetzee’s own son, Nicolas, at age 22. Dostoevsky sits alone in a room, letting into his mind his dead stepson. “Confronting [him],” Dostoevsky thinks,

is like descending into the waters of the Nile and coming face to face with something huge and cold and grey that may once have been born of woman but with the passing of ages has retreated into stone, that does not belong in his world, that will baffle and overwhelm all his powers of conception.

In this space—the nightmarish space of literary composition—“he will go naked as a babe into the jaws of hell.” It is only now that the vision he has been avoiding truly comes into sight: his son, “naked and broken and bloody, in the morgue.”

The creative act in this moment is a travesty against what can and should be thought. It requires a “descent into representations that have no place in the world.” Enjoying the perversion, even welcoming it, Dostoevsky begins to sketch the scene. The ecstatic madness of writing, the unaccountable otherworldliness of putting pen to paper, enables him to become alien even from himself: “[B]eyond the human, beyond man. There is nothing he is not capable of.” The risk inherent in writing is that it can involve taking pleasure in betrayal and the pain of others, as well as in letting the bodies of the dead rise up before one’s eyes. The experience is profoundly denaturing. After finishing the scene, Dostoevsky “picks up his hat and leaves his lodgings.” Now, though, “[h]e does not recognize the hat, has no idea whose shoes he is wearing. In fact, he recognizes nothing of himself.”

¤

Andrew Dean teaches at Deakin University in Melbourne, Australia. He is the author of Metafiction and the Postwar Novel: Foes, Ghosts and Faces in the Water (2021).

¤

Featured image: Luigi Russolo, Automobile in corsa [Dynamism of a Car], 1912–13. Le Centre Pompidou, Gift of Gift of Madame Sonia Delaunay, 1949. Photo: Jean-Claude Planchet, Centre Pompidou, MNAM-CCI, Dist. RMN-GP. centrepompidou.fr, Public domain. Accessed July 18, 2023.