According to the announcement, the idea was to use the tools of Large Language Models, Neural Language Models, and Machine Translators, alongside traditional tools (dictionaries, forums, conversations, and my own abilities) to translate a work published by the press into Spanish. Cita’s aim was to work with a literary translator to evaluate this approach in the context of creating open access editions of works in translation, and to make that evaluation accessible to others. After a few conversations with the editors, we agreed that the book would be Ten Days in a Mad-House, which Cita brought out in a new edition this year.
The end result would be not only the translation of Nellie Bly’s reportage, but also a series of reflections on the use of new technologies in literary translation. Cita wanted to work with a literary translator to evaluate this approach in the context of creating open access editions of works in translation, and to make that evaluation accessible to others. In this brief essay, I am sharing some of those reflections, and so only tangentially address Bly’s book, which already has a foreword by Mikita Brottman that does justice to the author’s work with greater authority than I could muster.
For the last couple of years, there has been a generalized alarm among the literary translation community. With the arrival of such programs as ChatGPT-4, Claude 3, DeepL, and others, evidence of the unethical use of the technology and a lack of understanding of its reach has endangered the economic viability of our profession already weakened by a predatory, unpredictable system of production that neither values our work nor understands its importance. Nevertheless, in parallel to the alarm, many of us began to use some of these tools, initially to figure out just what we were confronting, but also—and it would be absurd to deny this—because their employment has certain benefits.
It is important to understand that, in order to function, Large Language Models (LLMs) like ChatGPT depend on the unacknowledged and unpaid use of the works of hundreds of thousands of authors and translators who wrote the texts that “feed the algorithm,” an almost mythic formulation that stresses the monstrous nature of the machine. In short, what the program offers is language produced by creative humans, regurgitated with astonishing precision, based on the instructions it is given. Many authors have already commented on the moral issues associated with such mechanisms, and I don’t think it necessary to point an accusing finger again here: a just outcome could be achieved by seeking permission from and paying the creators and publishers of the multitude of titles used to “train” the model.
But what can be said of the result? Exactly how revolutionary is this new technology in terms of our profession? Based on my one-off experience of translating Diez días en un manicomio, I can say that the benefits are limited to speeding up the translation process while not necessarily improving it.
Firstly, the gender bias of LLMs must be taken into consideration; the great majority of the books used for training them were written by men. In the case of a text like this (Bly’s book), where the narrator-protagonist and most of the characters are women, the use of LLMs is particularly problematic. Although it is possible to give precise instructions (prompts) to counteract the gender bias to some extent, the sensitivity of the available models still leaves much to be desired when it comes to detecting and communicating the nuances of the structures of power revealed and criticized by Bly.
Added to that difficulty is the fact that the work was originally published 137 years ago. A contemporary translation faces a very particular challenge: it must make the work accessible to a present-day readership while respecting the textual markers that locate the book in a specific time frame. In other words, the book must seem simultaneously modern and dated; it must speak of our present era, while also addressing the circumstances that caused it to be written. The translation of classic books involves just that: allowing them to continue speaking to us and to resonate with modern readers as they did with those of their own epoch.
When choosing the most appropriate translation of a particular phrase or sentence, I keep in mind the readership of the text, in addition to its social function: I don’t make the same decisions when translating for a Spanish publishing house as I would for an independent Latin American publisher, or for an open access project that will be consulted by Spanish-speakers of different origins who are unfamiliar with my version of the language. At the other extreme, when translating, I am also conscious of the historical immutability of the original: I am working with a text written in 1887 and I must retain certain usages of that context, even when this may shock our contemporary political sensibilities. In this sense, I chose to translate “mad-house” as manicomio, a term that came into usage in Spanish at the end of the nineteenth century (and was popularized in the twentieth), and on other occasions occasions as asilio para alienadas (lunatic asylum) and pabellón para dementes (mental ward), all of which were commonly used in journalism at that time, both in Spain and Latin America; I also avoided the use of such present-day terminology as hospital psiquiátrico (psychiatric clinic), which corresponds to a post-positivist stage in the development of scientific thought. Those choices have political and moral implications, and it would seem to me unethical to pass responsibility for them to an algorithm trained by people whose motivation is purely financial. The utilization of the terminology of the period demonstrates how language reinforces some of the forms of oppression the author denounces: for instance, the patients aren’t women, but girls. Through a conscious alternation between the use of tú (informal) and usted (formal), I also hoped to shed light on the paternalistic attitude of the doctors toward their patients, a linguistic decision that isn’t present in the original.
A third consideration is that translation also functions as a form of cultural mediation, a negotiation between two different traditions. While I was translating Bly’s book, I couldn’t help but be reminded of Cristina Rivera Garza’s 1999 novel Nadie me verá llorar (No One Will See Me Cry; trans. Andrew Hurley), set in 1920 in the famous Mexico City insane asylum, La Castañeda. Rivera Garza later published La Castañeda. Narrativas dolientes desde el manicomio (1910-1930) (La Castañeda Insane Asylum: Narratives of Pain in Modern Mexico; trans. Laura Kanost), which reveals the archival research the novel draws on. The stories of the women who appear in both those books sounded to me like an echo of those recounted by Nellie Bly. A further element of this translational intertextuality is the fact that Nellie Bly is also the author of a book of reportage about Mexico (Six Months in Mexico, published in 1888), the country where I am writing this, and my reading of her work is necessarily marked by that interest.
As I translate Diez días en un manicomio, those interlinked readings inevitably inform my word choices. Moreover, my own sensibility and creative experience influence the way I translate certain passages, or the resonances I conjure up in the process.
During the three years I spent writing a novel about dance, I carried out research into the life and work of the German choreographer Mary Wigman, who was voluntarily admitted to a psychiatric hospital after the First World War and, based on that experience, developed a practice that led to the birth of expressionist dance in Europe. Her lover, the psychiatrist Hans Prinzhorn, was a pioneer of the study of art produced by patients suffering from mental illnesses, and his collection of “outsider art” was one of the most important of its kind in the twentieth century. As I read and translated the gestures of compassion Nellie Bly extends to the inmates of Blackwell’s Island, I couldn’t help thinking of those other experiences of internment that, via reading and study, form the corpus of my sources for the novel I had written. My translation also puts a work in dialogue with other books, but not with all of them; not with all the books that have been used to “feed” an algorithm, only with those that to me, as a creative human, seem relevant. I have yet to see an LLM that discriminates between sources on the basis of political and aesthetic sensibilities.
This said, the use of LLMs, as one of the various tools employed in the process of translation, clearly has some uses. If the translation has to be carried out on a tight deadline (as, for financial reasons, it very often is), computer models offer a number of options that are helpful when the translator gets stuck on a particular phrase. In a sense, the machine replaces the translation community that one could—ideally—consult about the translation of the problematic phrase.
Although the period of waiting for a response is reduced from hours to seconds, the replacement of the community by the machine has its dark side: in those professional conversations, the alliances forged and affinities outlined have generated a new “golden age” of translation. I have never translated so well as I did on a rooftop in Mexico City, mid-pandemic, when a group of six literary translators of various origins invited me to participate in a working session; we were all translating different texts but were able to help one another with uncertainties and passages where we were stuck. LLMs offer an alternative to that collective model, but, obviously, something is lost in order to gain speed. Communication between literary translators, including those on internet forums, also helps to bolster the material conditions of the profession. We start by asking, “How would you translate X?” and end by inquiring, “How much should I charge for this work?” As a member of the Mexican Association of Literary Translators (Ametli), I am conscious of the need for a union to defend our professional interests. When AI replaces these forums, it also disempowers their political potential.
I don’t want to go deeply into the energy costs and environmental impacts of the use of LLMs in translation—that topic could be the basis of a whole other essay—but it is impossible to ignore the fact that each ChatGPT consultation uses two glasses of water, particularly when you are translating in a city, like mine, where the effects of climate change are clearly visible.
Even so, there were moments of understanding and felicitous coordination between the algorithms and myself. After struggling with some difficult sentence for a few minutes, sometimes hours, an LLM suggestion, when viewed with a critical eye, provided the illumination for a more elegant solution than those I had thought up alone. Yes, AI can make life easier. On the other hand, I feel that when dealing with creative processes, it is important to live with the difficulty for as long as possible. Making things easier is the oft-stated aim of an industry that privileges products over processes, but it is in the process of struggling with difficulties, turning them over in one’s mind for days, that translators transcend the mechanical aspects of their work to become artists.
I will close this essay with a personal appraisal: what I learned while translating Nellie Bly remains with me; it forms part of the corpus of references I can make use of when translating other books, when writing my own, and in my relationships with my fellow human beings. Translation is slow, careful reading and there is nothing more transformative for a person’s discernment than the exercise of inhabiting the language of another. Handing over that practice to a machine implies wasting a transformative space that, in transcending the individual, impacts society as a whole: if we renounce the gesture of placing ourselves in the other’s words, we renounce much more than a mode of text production.
Mexico City, 2024
