There’s a man I know, a friend, Toros Garabaldian, and in all the years that we have known each other, he has always been, shall I say, dissatisfied with reality. When we were children, he read every volume of every fantasy novel he could find at the library — whether or not they were appropriate for his age — and this led to a period of severe disappointment at the lack of magic inhabiting our own world. Long after I moved away from the grease-stained streets of the neighborhood where we played, we still kept in touch, about our newest pursuits and fading dreams. Lately, I haven’t heard anything from him. The last communication I received was about a project he was excited to discuss.

Unlike Pierre Menard’s efforts to reproduce a book that already existed, Toros sought to create a computer program which would produce unique texts in the style of Jorge Luis Borges. You know the kind: a second- or thirdhand tale about a man stumbling upon a supernatural mystery or an imaginary disagreement between two gauchos or a critic describing a book that doesn’t exist. Toros had named it A Borgesian Contraption; it was a bit pulp for my liking, but in his imagination, it must have sounded very Escherian, a plain and simple label that belied the insanity of such an invention. Toros didn’t merely want to reproduce the stories which Borges penned; he reasoned that if he could devise a machine which had the capacity to produce texts “beyond” those of a human’s best efforts, feelings of despair and humiliation would surely spread like an infectious disease. Instead, he wanted to invite the reader to collaborate on the text generation. He worked out a way to turn the reader from a spectator to a participator. In his e-mail to me — which I cannot provide here fully, for two reasons, the first being that it would be disrespectful to his privacy, and the second that Toros always insisted on writing to me in our native Armenian, so reproducing his message directly would be unfair to the vast majority of readers — he was insistent on rebuking Gérard Genette several times, specifically his viewpoints on the (im)possibilities of imitating text. According to Genette, both a text and its style are distinct components, which can undergo processes of transformation and imitation. Both the text and the style of a parent hypotext can be distorted through a comedic, satirical, or serious transformation, resulting in a new hypertext. But Genette determined that a text could not be imitated in a hypertext, only its style, in three ways: by pastiche (for the purpose of entertainment), in caricature (for the sake of derision), or through forgery (for perpetuating antecedent literature). At the end of his explanation, Toros quoted directly from Genette: “It is impossible to imitate a text…because it is too easy, hence insignificant.”¹

For Toros, only half this statement was true. “It is impossible to imitate a text,” he wrote to me, “because the totality of possible imitations exceeds human patience.” In other words, there are many ways, near infinite, in which a text can be imitated, which makes the task too difficult for someone to definitively complete. When we talk of transformation, we mean to take a text and modify it in some way so as to produce something inspired by the original but altogether new. When we talk of imitation, we mean to take a text and generalize its style, reducing it to a genre whose predecessor we can identify. But we can imitate a text, Toros argued, if we stop thinking about it in terms of its literal words. “Think of Theseus and his ship,” he wrote. “If one rotten plank is removed and replaced with fresh timber, do we not still call the ship his? And if, over the years, every plank is replaced, is the ship not still the same?” Toros believed that there was another word already available to capture the notion of imitating a text: translation, or the act of finding words from one language that closely matched another. Genette himself “devotes barely four pages” on the subject and considers translation to simply be a subform of textual transformation.

“But how can this be?” Toros demanded. “Earlier in his argument, Genette explains that textual transposition can be considered as the ‘lifting’ of one element of a text (character, plot, etc.) into another; thus, Mann’s Faustus from Goethe, Joyce’s Bloom from Homer. Yet, in translation, the underlying signifier in the source text is changed while its meaning remains. And, when we are engaged in a reading of a translation, each printed word is a shadow of all the words that could have been. Genette states — rightly! — that ‘every act of translation affects the meaning of the translated text.’ It stands to reason, then, that translation is not a strict transference (any word in one language cannot simply be substituted for one in another language), but neither is it distortion. A translation exists as a unique form of imitative hypertext, one which, in the act of crossing languages, both is and is not its own hypotext.² Because there are so few direct equivalencies between languages, and so many ways to approximate one sentence for another, the possibilities for imitation appear unbounded (to a very distant limit). A perro, for example, is a dog, but it’s also a canine, a hound, a pooch, and a mutt. Are these not all imitations (of varying efficacy)?”

That is what Toros strove to prove by creating A Borgesian Contraption. Whether he ever completed his vision or is still laboring on it, I don’t know. In his e-mail, he attached the source code without any explanation for how it works, although he did give me permission to post it online as a sort of draft, which I’ve done. I’ll try to convey my understanding of it.

It should no longer feel like outlandish science fiction to imagine a computer program which can “predict” your next action and subconscious desires. Over the last thirty years, we’ve seen such systems integrated (willingly or not) across our lives: in shopping advertisements that follow us on each website we visit, from prompts to watch a new movie based on our streaming habits, and even through recommendations for who we might best become friends with. All of the activities you perform online are ingested and analyzed into a sort of recorded history; based on your past choices, these programs simply suggest your most likely outcome. Generative text systems are not so different. Given a substantial volume of text to study (or train on, in machine learning parlance), if a program is given a phrase, it can predict the words that should follow, based on their probability of previous uses. For example, after prompting a system which has only trained on children's fairy tales with the phrase “Once upon a,” it’s exceedingly likely for it to continue the sentence with the word “time.” With this understanding in mind, I can see the connection to translation which Toros is attempting to conjure. If a certain well-known phrase of, say, Tolstoy, is translated from Russian into English, subsequent translations of other Russian works which harken back to Tolstoy may need to repeat that same translation, in order to evoke the same reference for the English reader.

A Borgesian Contraption is a system which has been fed nearly all the short stories of Borges (in English). Its “brain” — the underlying software which reads, parses, evaluates, and responds to prompts — is able to take a starting sentence and continue crafting a story that resembles something Borges might have written. Although one could have a computer generate a story on its own, this was not Toros’ goal: “To have a machine produce text endlessly is not in and of itself interesting. It could, of course, generate texts which contain every possible permutation of English words — but then these would cease to be stories and become Babelian gibberish. A Borgesian Contraption begins to write a text of its own, but it is intentionally halted, in order for a human to intervene and provide his or her own opinion on how the text should continue. It is the collaboration between a machine/author and its user/reader which is what makes the project special. The work-in-progress is a machine translation in the loosest sense, an interpretation of Borges whereby a machine decides, with capricious neutrality, what word to place next, tempered by the aesthetic sensibilities of a reader. Such a union is, I believe, vital for understanding our relationship with literature in the future: ‘If we would like to believe in our own creativity or freedom, the challenge is to think it together with the machine.’”

Why Borges? Most obviously, the author’s interest in the infinite and the permeability of readings pair well with the project’s generative qualities. (Toros certainly believed his imaginary books “mesh into a unifying metaphor of the complex and ambiguous relationships between writing and reading: relationships…that are quite evidently the very soul of hypertextual activity.”) Still, it would be the wrong question to ask. For Toros, it was Borges’ translator, Norman Thomas di Giovanni, who interested him more. A Borgesian Contraption is not trained on all of Borges’ translated works, but only those which came from di Giovanni, who, although not Borges’ first English translator, was certainly his most favored and most prolific.³ Many of the translations were done in collaboration with the author, and Borges even acknowledged that di Giovanni’s texts surpassed his own. “Therefore,” Toros concluded, “when we are reading Borges in English, whose words are they? And when a computer produces texts after reading all of di Giovanni’s translations of Borges, whose words are those? Is a computer producing new works through a process of distortion known only to it? Or is it translating a copy of the text into an imitative pattern of harmonic ones and zeroes?” More to the point, Toros preferred to work with an author whose opinions on translation were not ambiguous: “translation occupies for Borges a place of honour, a locus where experiments with writing can be made and new ways properly tested.” The most intimate reason Toros rationalized the union of Borges and di Giovanni was due to the fact that di Giovanni’s texts have now fallen out of copyright, and Borges, having long since passed into the other world, is unable to dispute with publishers who want to produce new versions of his texts (presumably for profit over fidelity). “In another decade,” he wrote, “di Giovanni’s works may be a distant memory, as more and more translations are produced, each claiming a ‘superior truth’ over the previous edition. (This sequence, itself Borgesian, is ironic, were it not so tragic.) As readers, what does it mean to witness the reconstruction of a ‘stable’ translation that has been blessed by its author?”

Can a text be truly imitated? I’m not sure. Toros mentioned Theseus; I’ll draw upon Heraclitus: one cannot read the same text twice, because it is different every time. This makes the question moot, I suppose, not because an imitation is too easy or too hard but because it’s simply impossible. I would agree that categorizing translation as a textual transformation is an inadequate designation. If a translation were a distortion, then it would conclude that there can be varying degrees of distortion, implying “good/accurate” and “bad/inaccurate” versions of target texts. In my opinion, the only types of bad translations which can exist would be the ones in which a translator intentionally manipulates the source text for the sake of censorship. Even honest misinterpretations wouldn’t be considered “bad,” since the translator can justify their reading as they understood it. But I don’t wholly agree with Toros’ idea that translations operate as imitations either, as that implies a broader unity between multiple versions of the same text. Certainly, texts can be infinitely generated, but is it correct to call any random sequence of words in one text that happens to inadvertently match another sequence in another text an imitation? I suppose it all depends on what we mean by imitation. From my understanding of Genette, it’s to mean a text that is heavily inspired by another to the point of direct equivalence; for Toros, it means the potential equivalences of all possible variances of a text through translation. An imitation can describe an accurate copy of something: a mimic, a duplicate. But it can also mean a derivative, possibly inferior, object: imitating a sound, imitation leather. Perhaps the imitation of a text is not in its writing but in its reading, as a text only truly exists once it is read. The difference here, I think, is that A Borgesian Contraption relies too heavily on the idea of the infinitude of writing a text, and not enough on the plurality of a text’s readings.

A text’s existence is distinct from its multitude of readings and interpretations, of which translation is just one. Every translation, by definition, can only ever be an approximation. I believe that the process of translation is one which can never be completed, or at least, can only ever be completed imperfectly, a mutable half-truth whispered by a perjurer to a listener eager for the fabrication. My favorite way to understand how a translation is produced and interpreted comes from Kate Briggs: “when it comes to writing and reading translations the question…of what was really said or written, gets suspended, slightly. The translator asks me to agree to its suspension. To suspend, or to suspend even further, my disbelief.” The translator attempts, through their translation, to imitate a reading of a source text and its style; the reader knows that the target text is not something which can be considered authentic. To read A Thousand and One Nights or The Odyssey in languages other than their original is to willfully participate in a conspiracy, a duplicitous secret which both the translator and reader share silently. The phrase “traduttore, traditore” comes to mind, which has been translated from Italian as “translators, traitors” — but a translator’s treason is only acknowledged when it is read.