Introduction

I am probably not alone when diagnosing the increasing difficulty to write about writing technologies and their automation via digital technologies. My writing environment is now not only connected to the vast informational data that was the Internet, opening up my research questions to any possible answer online, but the document itself where I might be typing in these questions is now part of an AI-powered word processing application that shapes my thoughts as soon as they reach the page and transform into written words, sometimes before. I like to think that these writing technologies do nothing else but speed up a process that originates with me, but the lack of critical distance to assess a writing that emerges in between me and these machines blurs this certainty. Is the process of machine automation not only accelerating but shaping contemporary writing with machines? And if so, so what?

Since answering this question regarding my own process hits a little too close to home, in this essay I examine the 2023 AI-generated text Los campos electromagnéticos [The Electromagnetic Fields] by Spanish writer Jorge Carrión and the artistic collective Taller Estampa as an early example (perhaps the first)¹ of writing with GPT-2 and GPT-3 in Spanish. I read this literary experiment vis à vis philosophy of technology and literary criticism as these two fields engage the imbrication of technology and writing. Further, Carrión’s text, part essay and part creative experiment, raises the question of whether AI can be seen as an evolution of experimental writing in the tradition of André Breton and Philippe Soupault’s surrealist 1920 work, Les Champs magnétiques (The Magnetic Fields), or if it represents something entirely new, beyond the literary avant-garde.

While Carrión’s Fields initially aimed to challenge conventional notions of originality and authorship, framing AI-generated text squarely within literary theory and history, I propose that his experiment with GPT-2 and GPT-3 may be an early attempt to radicalize and accelerate the process of exosomatization observed in technological society today; that is, a way of automation that exceeds the literary experiment of automatic writing proposed by the French surrealists to become a mark of how contemporary society thinks and communicates broadly. As I explain below, Bernard Stiegler believed that our increasing reliance on the technical exteriorization of memory and cognition through technology would lead to the eventual loss of the ability to think (Stiegler, 2016), ultimately eroding the social bonds necessary to think with other humans, i.e., the basis for human politics. Thinking and memory are not writing but writing as a technology and writing technologies participate in the exteriorization of both. In a book such as Los campos, focused on challenging while championing AI as a potential mirror of the self, the implications for human writing, memory, and thinking, and the role of these in society, extend beyond the literary work at hand.

To examine so, I home in on the ramifications of Carrión’s custom GPT training, where the final textual product in Los campos is still clearly distinguishable from human-authored text. I read Carrión’s work alongside relevant surrealist and technology frameworks, but I also engage in a feedback process of writing with ChatGPT, a commercially available AI-powered bot, based on OpenAI’s GPT-4-turbo model, in an attempt to update Carrión’s 2023 practice to the almost mundane GPTs of today. As the model itself explained to me when prompted, the turbo model is an enhanced version of GPT-4, designed to be faster and more efficient than GPT-4 or its predecessors GPT-3 and GPT-2. Fittingly, I understand ‘speed’ and ‘efficiency’ to be the two main drivers behind any process of machine automation. GPT-4 is part of OpenAI’s ChatGPT products, it’s accessible to users without a premium account and, as of now (May 2025), it’s the latest version available through its commercial interface. It is also the version used by 90% of OpenAI’s 500 million weekly active users whose data, unless explicitly opting out, is used to improve future models (Palumbo, 2025). I am also writing within Microsoft Word, which is assisted by Microsoft Copilot, an AI-powered writing feature which also uses a version of OpenAI’s GPT-4o as its foundation. The writing of this text thus is both the product of a range of human prompting (e.g., ‘summarize the two main ideas in this paragraph and propose a transition to the next,’ ‘write a subtitle for this section,’ ‘in the following sentence, give me an alternative word to X’ but also, ‘hey, would you say this section was clear enough?,’ ‘I am feeling lazy, could you please turn these bullet points into a coherent paragraph in the style of my previous writing. Remember to write like a college professor, ok?’ and ‘how would you describe our writing relationship? Would you say it’s a collaboration?’) as well as some less intrusive machine nudging (i.e., grammar, syntax, and spell check). I explain this to ground my discussion around AI on both my own practice with this text you are reading, as well as a concrete LLM instantiation in the world today as the corporate version of AI that is, not an ideal version of technology with a theoretical potential of neutrality, as this could never be the case.

The GPT-4 model that lays behind this written text on a literary artwork was trained on a big chunk of stolen data from the internet (Bender and Hanna, 2025: 53), exploited labor (Regilme, 2024), and environmental harm (Hogan, 2024) that, while originating in the Global North, also affects the South in disproportionate ways like many forms of contemporary techno-colonialism (Muldoon and Boxi, 2023). This model, like any other technology and tool is inevitably traversed by, even as it traverses literary histories and corporate workforce practices, among many other (un)related things. It is also being minimally traversed by me, as I am also traversing it. As Melvin Kranzberg succinctly put it when discussing technology broadly, we could say generative AI is ‘neither good nor bad; nor is it neutral’ (1995: 5). This means that we should always analyze it in relation to the political, economic, and social context in which it emerges to explain how it is created and used, given that some concrete instantiations of technology are, in fact, worse than others.

Writing within the machines I set up to write about how these machines intervene in the writing of other contemporary and avant-garde writers before me, I wonder: Are we witnessing an early instance of corporate machines writing like humans, participating in a shared function of thinking and writing ‘collectively,’ despite (or because of) these traversals, or is this a sign that humans are beginning to write like corporate machines, losing our capacity to think with but essentially against and as others humans? In other words, am I writing with or am I writing as a mimicking machine? Who is writing with, as, or against me when writing within these AI-powered environments?

Now, while the comparison is fruitful, this is not exactly a case of ‘bot mimicry’ in Malthe Stavning Erslev’s terms, meaning ‘the practice of mimicking software that mimics humans—simply put: imitating imitative software’ (2024: 2). Rather, what I am interested in exploring here is not the act of a human consciously producing text that is meant to be read as though it were written by a bot or any other automated system, but the subconscious internalization of bot-writing. In this, I am not referring to a deliberate reverse of machine imitation, where the human would intentionally adopt the style or cadence of a machine—even if both practices may eventually serve as cultural critiques, highlighting widely held cultural conceptions or imaginaries about what bots may or may not sound or behave like. Intentional or not, however, writing like a bot today cultivates a powerful mode of ‘relating to technology’ (Erslev, 2024: 82) which goes well beyond questioning its outputs. What may this mean for my human writing and the human society I know? Who is training whom in Los campos electromagnéticos?

To ask these questions is to open this text to a longer genealogy of automatic writing, ranging from surrealist dreams of psychic automatism to the algorithmic automation of cognitive processes. Where the surrealists sought access to the unconscious, Carrión and Taller Estampa confront a different abyss; one populated not by repressed individual desire but by collective data corpora, language models, and feedback loops between (many) human and machine generated text. If André Breton and Philippe Soupault’s Les Champs magnétiques dreamed of a seamless interface between thought and writing inscription in 1920, Los campos inhabits a reality in which this interface has become programmable, scalable, and perhaps disturbingly different, if not completely indifferent, to the humans whose data was used to train the large language corporate models that hide behind it.

From Automatic Writing to Human Automation

In Breton’s 1924 surrealist manifesto he defined writing automatism as ‘automatisme psychique pur [pure psychic automatism]’ (1989: 328, translation my own), a method of transcribing thought, ‘soit verbalement, soit par écrit, soit de toute autre manière [whether verbally, in writing, or any other manner]’ (1989: 328, translation my own) unfiltered by reason, aesthetic concern, or moral censorship. This idea of automatism aimed at immediacy: thought as dictation, thought as flow, thought as a direct transmission from the unconscious to the page. The automatic writing subject would become a sort of embodied recording technology, collapsing the distinction between its experiencing of the unconscious and its recording. And yet, as Madeleine Chalmers reminds us, this surrealist automatism paradoxically required technical mediation in order to conserve the ‘immediacy’ of voice as the pure unconscious rehearsed in the recording body, entangling Breton’s ideal of immediate psychic expression with several technologies of inscription of the time: pens, paper, typewriters, and eventually tape recorders and phonographs (Chalmers, 2020). In its instantiation as a literary style, the written mediation would conform further to certain conventions; even if as an artistic proposal, what the surrealists sought was to free literature from bourgeois conventions (Lipinski et. al, 2022). While not yet digital, we see that surrealist automatism was never truly immediate; thought was always technicized, mediated by a technology that would shape its inscription in the outside world beyond each author’s body and mind. The record of the unconscious had, even in its wildest dreams, to conform to external technologies of preservation from the start. But isn’t this how all thinking, how all speech, gets preserved and thus shaped? Following Leori-Gourhan, Derrida has argued that speech both implies and depends on writing. As a recording technology, writing ‘supplements’ speech fleeting nature by leaving a traceable mark that, in turn, makes further speech possible (Derrida, 2015: 422).

In a traditional literary context, the term ‘automatism’ is primarily associated with the automatic writing of surrealism. This association is at the core of Carrión’s text, riffing its title off Breton and Soupault’s book. In modern and contemporary Anglo-American technological thought, however, and in public perception broadly, automation refers to the replacement of human workers by machines and computer systems. As Chalmers has observed, these two meanings appear contradictory at a first glance: automatic writing involves expressing human subjectivity, while automation aims to remove it from the workplace (Chalmers, 2020: 369). And yet, when Bernard Stiegler published the first volume of La Société automatique in 2015, his concept of work and automation extended far beyond merely substituting human workers with machines, focusing instead on the automation of every aspect of human life and thus repurposing altogether the literary term ‘automatism’ for the digital age. While technical exteriorization such as writing had always been for Stiegler a form of memory, ‘grammatization,’ i.e.; the exteriorization of memory in the form of written traces of anamnesis, is also seen as a framing of exteriorization by the imperialist power deployed by colonizers on the imposition of their language on colonized populations (Stiegler, 2004: 117-118). Grammatization, in this way, changes our relationship to language to achieve the control of ‘la conscience, des corps et de l’inconscient [consciousness, body and the unconscious]’ (Stiegler, 2014: 118, translation my own).

The dependance of memory and writing on artificial exteriorities makes the question of technology an irreducibly political question (Hansen, 2010: 66). When these externalities are associated with our acts of remembering (the embodied act of memory that Stiegler calls ‘anamnesis’), these artifacts and technologies facilitate and enhance memory, constituting meaningful symbolic practices and community formations. However, when these externalities are disassociated from the embodied act, ‘they advance the interests of the culture industries (Adorno and Horkheimer) and of control societies (Deleuze), which work to transform human beings into mere consumers, passive recipients of prepackaged and standardized commodities and media fluxes who have no hope of becoming producers’ (Hansen, 2010: 66). Stiegler’s writing on memory and writing preceded the rise of the sophisticated language models we have today, even if the core belief that our reliance on artificial memory aids would make us intrinsically more vulnerable to manipulation by the corporations or industries that control them is now more relevant than ever.

Framing thinking, writing and technique always in co-evolution leads Stielger to predict that what we refer today as:

[…] artificial intelligence is a continuation of the process of the exosomatization of noesis itself, such as it begins firstly with fabricating exosomatization, making things by hand, and continues with hypomnesic exosomatization, as that which makes it possible to access lived experiences of memory and imagination, which have accumulated since the origin of the play of works, as Bataille considers them, and which engender, in passing through writing, instruments of observation, calculating machines whose principles were established by Leibniz, and analogue technologies, which form the basis of the culture industries (Stiegler, 2018).

As Stiegler and other philosophers of technology and writing like Gilbert Simondon, Jacques Derrida or even N. Katherine Hayles saw it, thinking and writing have always been artificial and technified processes and, in their technological mediation and entanglement, these processes should not be seen as neutral either.

Now, returning to the surrealists and their deployment of technologies, where Breton sought to exteriorize the inner world in order to understand the unconscious, today’s machine learning architectures automate this exteriorization in ways that exceed and even bypass human reflection. Artificial intelligence algorithms do not simply record; they predict, correct, and refine according to their own grammatization logics, detached from the human anamnesic process and, in doing so, they collapse the temporal gap that once separated human thought from writing, impeding their reflection and forcing thought into a shape of their own.

AI’s grammatization no longer supposes the spontaneous trace of the unconscious. AI automation now signifies the systematic exteriorization of human thought, which Stiegler had termed ‘exosomatization,’ shaping our understanding of the world through algorithmic grammars. Following Stiegler’s understanding of grammatization, technique, and mortality, Chalmers affirms:

What we choose to remember shapes our understanding of the past … our present and future. To control methods of exteriorization is to control consciousness, suppressing critical reflection and individuals’ desire and ability to act to shape the future. In the twenty-first century, grammatisation takes the form of large-scale data gathering, its deployment in algorithms, producing an automated existence: not only in the workplace but in every aspect of our lives (2020: 375).

Before taking his own life, Stiegler warned us that the kind of hyper-automation behind AI’s grammatization risked annihilating the very conditions of thought (Stiegler, 2016). In surrendering memory, imagination, and symbolic reasoning to technical systems of this kind, we do not gain access to ourselves; we outsource ourselves to algorithmic grammatization in a radical process of mechanical remix and algorithmic homogenization of thought. If Breton imagined the writer as a medium channeling inner truths, Stiegler offers a bleaker vision: the human as terminal in a feedback loop, progressively incapable of critical difference between herself and the logic of the machine, unable to establish relationships with other different humans in turn. What part of this segment was in fact written by a bot rather than me?²

From Surrealism to AI Writings

Carrión’s GPT-2 and GPT-3 surrealist project must be situated within a lineage of experimental literature that includes both surrealist automatism and early computational poetics. In the mid-20th century, early programs like Christopher Strachey’s Love Letter Generator began to experiment with automatic textual generation. As a matter of fact, this 1950s program, developed by Strachey and Alan Turing, is conceived by some to be the first example of computational literature (Saum-Pascual, 2016). Since then, there has been a steady production of automatic generated literary works, even if the original textual generation of the 1950s and 60s gave way to the more interactive, kinetic and multimedia capacities of hypertext technologies once the internet was popularly adopted (Funkhouser, 2007: 2). Yet, automatic textual generation for literary purposes never stopped, and recent examples in the Spanish speaking world of print books that capture the output of algorithmic processes like Carrión’s range from codeworks where programming language is reinterpreted as static poetic language, like Belén García Nieto’s A 6000 metros de profundidad (2023), to small custom made programs that remix novels, like Milton Läufer’s A Noise Such a Man Might Make: A Novel (2018), where two American classic novels are broken down and remixed to create a new print text, to the training of symbolic reasoning programs like Rafael Pérez y Pérez MEXICA, a creative agent capable of ‘evaluating and making judgements about its own work’ (Pérez y Pérez, 2017), which hides behind the collection of short stories Mexica: 20 años, 20 historias (2017). None of these programs or their output, however, could be considered hallucinations from the unconscious like the surrealist work of Breton and Soupault, but rather closed systems, mostly deterministic and self-contained, yet still offering a different challenge to the idea of authorship, as well as reimagining the creative process as structured, symbolic, and rule bound.

It is tempting to read Läufer’s Markov chain method, or even Pérez y Pérez symbolic creative agent, alongside today’s automatic (artificial) generation of text, and yet I pause. The illusion vanishes when the textual output is not just surprising, as the unusual juxtaposition of concepts or structures that a close generator might offer, but completely unpredictable and untraceable, as the hallucination of an LLM making up information outside of the scope of a given prompt and apparently originating from outside any chosen data sets makes it seem as if the model were digging deeper in its black box, always overflowing with training data not provided by the programmer. Where did that new information come from?

Fossilized knowledge is hard to overcome for all, it appears. Like humans, some models are difficult to reprogram completely when their training is their essence and path to knowledge. This was obvious to me when ChatGPT inserted the work Mucho trabajo, by Pablo Katchadjian, in the first iteration of the paragraph above this one, even if I had only instructed it to reflect on Läufer’s and Pérez y Pérez’s texts. Why did the model offer another experimental Latin American male voice, and why did it settle on Katchadjian’s boutique work, Mucho trabajo, a miniature object that reduced a 250-page novel to a mere 8-page zine printed in 2.1 size font? I know that Katchadjian had experimented with technical mediation in his writing; I am familiar with El Martín Fierro ordenado alfabéticamente (2007), where the author re-orders all the verses in the canonical Argentinian text in alphabetical order, and I even have written about his Aleph engordado (2009), where he remixes and expands on Borges’s masterpiece, in a previous book (Saum-Pascual, 2018). But I had not instructed ChatGPT to reflect on any of this. It brought it up on its own, retrieving either from a fresh internet search or pulling from its training data, who is to tell? The chatbot’s interface may tell you, but why trust it? Why bring it up? I don’t know, I can’t know, and I don’t like it.

In this, Carrión’s Los campos electromagnéticos is closer to my writing exercise here than to those early digital experimental generative texts. While works like Mexica rely on symbolic reasoning and narrow, rule-based creativity, Carrión’s collaboration with GPT-2 and GPT-3 introduces a new layer of black-box opacity. The AI writing machine is not simply following instructions but predicting language based on statistical probabilities drawn from a vast archive of human text, mostly provided by Taller Estampa but not all. Could a model free itself from its archive?

Carrión explains his process working with Taller Estampa in the training of his GPT-2 model, ‘Jorge Carrión Espejo,’ to develop a tool that would mimic his literary style. To do so, Estampa trained the GPT-2 model on:

[…] tres tipos de conjuntos de datos: un gran número de textos publicados en castellano, que le permite generar lenguaje en ese idioma; un archivo de obras clásicas, lecturas generales sobre tecnología y cultura contemporáneas y libros que han sido importantes en mi formación literaria, cultural y sentimental; y mis propios artículos, ensayos y novelas [three types of data sets: a large number of texts published in Spanish, which enables it to generate language in Spanish; a collection of classic literary works, general readings on contemporary technology and culture, and books that have been important for my own literary, cultural and sentimental formation; as well as articles, essays and novels written by me] (Carrión, 2023, translation my own).

The final note does not specify if Taller Estampa initialized the transformer from scratch and trained it completely on a different dataset, erasing GTP-2’s pretrained knowledge of WebText and using only its architecture, or if they used the pretrained model and later fine-tuned it with the kind of text’s that shaped Carrión’s literary style. However, in this note, Carrión reflects on a particular segment produced by his mirror GPT-2: ‘Las inteligencias artificiales han empezado a ser percibidas como seres inteligentes y nos han dado un nombre: Visions. Soy incapaz de detectar su origen en mis libros [Artificial intelligences have begun to be perceived as intelligent begins, and they have given us a name: Visions. I am unable to detect its origin in my books]’ (2023: 129, translation my own). Puzzled, unable to track the source of this ‘Visions’ in his texts, he wonders where the English term might have come from. Carrión doesn’t reach any conclusion and doesn’t dwell too much on the technical implications of this visionary revelation. I, on the other hand, do dwell: Was this term somehow buried in the training data set that included generalists works on science and technology? Or was it an inedible print of the pre-training of the original GPT-2, as I (and perhaps you) would assume? The answer is important and, even if unresolved, it should make us wonder about the relation between an artificial intelligence’s model architecture and its training data, and about our capacity to ever truly re-shape an informational technology structured on data, whose ontological essence is this data, and perhaps not only its infrastructure since, without data, this structure is simply not an AI. Data is not intelligent, but the ‘intelligence’ that we ascribe to the model comes from its uncanny ability to predict text, and this transformer’s intelligence capacity was granted on its original 40GB of training text. GPT-2’s 1.5 billion parameters were just knobs inside a model, becoming only ‘intelligent’ when self-adjusted through their encounter with the world. A learning algorithm cannot learn without data that a priori originates outside of itself, neither can an attention mechanism pay attention to itself without making a cut from itself and the object of this, an optimizer necessarily needs data to be optimized, even if what ends up being optimized is the algorithm, and so on. In this case, the intelligent machine becomes intelligent in its intermingling of structure and world (of data).

Now, if ‘Visions’ was never part of the model’s re-training corpus, was Jorge Carrión Espejo bringing back some old data from its pre-retrained past? This feels like the only technical possibility. Was the re-trained model unable to let go of its first encounter with the world? What does that mean for any interaction we have now with any model explicitly pretrained in any way? What is the pact I am making when writing together with ChatGPT-4? How is my memory, my writing and my thinking being repurposed in my interaction with this machine of external memories that do not begin with, cannot begin with, and do not belong, and cannot belong to, me? Anything I write will always be tainted with the model’s violent past, a past that, while maybe shared, is not mine (please go back to paragraph 5 in the introduction).

Training the Mirror: Jorge Carrión Espejo

‘Carrión’s custom language model offers a provocative twist on the concept of the literary mirror’ was ChatGPT’s last interference in this discussion before I decided that our ‘collaboration’ was over. ‘The machine does not merely reflect Carrión,’ it insisted, ‘it extends and possibly distorts him,’ and feeling the distortion that was overcoming me, our very different intents when writing this text, I decided to end our co-generation, for the generated mirror does not offer a faithful image (how could it?) but a simulated voice capable of producing poems, reflections, even critical essays, even in a so-called Carrión-style. Given the deep entanglement between thinking, writing, memory and technology, this generation (here and in Los campos) was no simple co-authorship; it is a speculative entanglement of identity with predatory machines.

Los campos electromagnéticos is divided into sections that foreground this instability. Part one, titled ‘Introducción: Teorías y prácticas de la escritura artificial,’ theorizes the implications of writing with machines and presents the reader with Jorge Carrión’s (human) written voice and ideas. This text serves as an introduction to the literary experiments of GPT-3 and GPT-2 that follow; explicitly aligning the text with its surrealist precedents, together with other examples of computational literature, here termed ‘literature predictiva [predictive literature]’ (Carrión, 2023: 25).

Part two, ‘Los campos electromagnéticos’ pays direct homage to Breton and Soupault’s Champs, echoing surrealist spontaneity through the mechanical randomness of AI outputs. This section, explicitly penned by GPT-3, presents a combination of short stories and poems generated from the prompting of OpenAI’s Davinci model, the largest and most powerful pretrained model available at the time. Following Breton and Soupault’s process, they prompted the model with a version of the titles or epigraphs from Les Champs magnétiques: ‘Si el primer poema de los franceses se titula ‘El espejo sin azogue’ y el último ‘las máscaras y el calor coloreado,’ hicimos que el programa escribiera a partir de temas o títulos como ‘El espejo de píxeles’ o ‘Las máscaras de los avatares’ [If the first poem by the French poets is titled ‘The Mirror without Silvering’ and the last one ‘The Masks and the Colored Heat,’ we made the program write texts following prompts like ‘The Pixel Mirror’ or ‘Avatars’ Masks’]’ (Carrión, 2023: 124, translation my own). Given the pre-trained nature of the Davinci model, Carrión explains that his prompting would sometimes produce English responses—‘esa lógica colonial [that colonial logic]’ (Carrión, 2023: 128, translation my own)—adding a new layer of complexity to the textual generation presented in the final version of the book, where GPT-3’s inferences were then parsed through Google Translate, ‘esa familia de algoritmos [that algorithmic family]’ (Carrión, 2023: 128, translation my own). Leaving aside the question of Carrion’s automatic translation choice, which I am avoiding to engage with in this essay altogether by leaving all sources in their original languages (or, at least, in the languages I have encountered them in, and then adding my own human translation), the resistance and re-affirmation of the pretrained model float to the top in this second section of the text. To me, it is the least interesting of the book but, perhaps, the closest to the way most of us interact with the latest version of OpenAI’s transformers. The dialogic nature of this section recalls both Breton and Soupaul’s surrealist method based on conversation and prompting, as well as the chat structure in ChatGPT’s latest versions. It also is closest to my own relation to the bot when writing this text, until, of course, that relationship also ended, and I dumped the bot.

Part three, ‘Una escritura artificial de las prácticas y las teorías [An artificial writing about practices and theories]’ is generated entirely by Jorge Carrión Espejo (GPT-2) and presents a text that reflects on the model’s own artificial subjectivity. As explained earlier, this is the model trained on Carrión’s chosen data, configured with specific hyperparameters to balance creativity, coherence and stylistic novelty, aiming to ‘crear un programa que escribiera como yo [to create a program that wrote like me]’ (2023: 129, translation my own). While the full training data sets and overall process is still a mystery to me as reflected upon earlier, the parameters are not. Right under Jorge Carrion Espejo (GPT-2), they are made explicit: ‘temperatura: 1.0; repetition penalty: 2.0; Top_p: 1.0; Top_k: 50; Sample: True; Num beams: 2’. In other words, Jorge Carrión Espejo’s temperature was set to 1.0, which allows for a moderate level of randomness in word choice (nothing too deterministic nor too crazy) thereby enabling a certain creative variation without losing textual coherence. A repetition penalty of 2.0 was applied to discourage the model from producing redundant phrases, a common issue in neural text generation that, I am guessing, the fleshed Carrión would also try to avoid. The Top_p value, set to 1.0, and Top_k, set to 50, together define the scope of possible next-word choices: while Top_p allows sampling from the full probability distribution (ensuring no arbitrary cutoff), Top_k limits the selection to the 50 most likely next words, thus ensuring fluency without excessive restriction. The model was also configured to sample, rather than always choosing the most probable word, promoting diversity within its style, and it employed 2 beams in generation, to maintain a degree of structural control and explore alternative phrasing paths before finalizing the output. Together, these settings reflect a deliberate calibration of Carrión’s mirror, aimed at producing somewhat surprising language while trying to preserve the formal integrity of the text. Isn’t this what any good writer would aim for? Isn’t this what Jorge Carrión should sound like?

Glitches, Hallucinations, and the Trace of the Machine

Yes, and yet. When I first started working with computational literature, what most clearly distinguished machine-generated writing from human writing was not its fluency but its errors; its many funky glitches, intriguing mistakes, provocative non sequiturs… those things we may now call ‘hallucinations’. These were not failures per se, but symptomatic signs of the machine’s othering logic, and for that they were invigorating. They allowed for a glimpse into the Other, a breach into the otherwise fixed structures of my thinking through language and writing with technologies. Like Rosa Menkman, I found glitches can be exciting and revelatory: interruptions that expose the politics of code and rupture the veneer of seamless functionality. A glitch can be ‘a source for new patterns, anti-patterns and new possibilities that often exist on the border or membrane (of for instance language) […] For a moment I am shocked, lost and in awe, asking myself what this other utterance is, how was it created?’(Menkman, 2011: 4-5). While mostly polishing the glitches in his text, Carrión’s refusal to completely edit out these ‘failures’ is a critical aesthetic and political decision. Jhoerson Yagmour Figuera believes that these glitches eventually ‘redirigen al lector, alejándolo del discurso tradicional que rodea a la tecnología y disipando la idea de perfección lineal corporativa de la técnica [redirect the reader, drawing them away from the discourse that traditionally surrounds technology and dispelling the idea of perfect linear corporate technique]’ (2023: 27, translation my own).

To me, and until very recently, just like the typo had traced the human hand behind the typing, syntactical glitches and other exciting hallucinations had become the aesthetic markers of a nonhuman authorial presence. And that’s what made them interesting, the realization that there were others involved in the production of literary texts, briefly freeing literature from the oppression of the human author’s language and writing traps. If like the typo, Breton’s automatic writing produced typographic slippages that revealed the unconscious, Carrión’s machine glitches reveal the opacity of the AI algorithm. These are not messages from the collective unconscious but echoes of a different kind of machine logic; one structured by statistical inference, shaped by datasets, and operationalized by a model we no longer can understand. In this way, the machine does not dream; it calculates its own logic. And this very difference gives Los campos its unsettling power. It is not that the GPT-2 and GPT-3 misunderstand language, but that they operate in a different regime of meaning-making altogether. Their ‘errors’ are the seams of the system, traces of a logic that both overlaps with and exceeds human intention in their own different ways. For Los campos electromagnéticos to work as source for this revelation, it has to be read as such: as an essay on a praxis about writing with machines, this is, an essay in its etymological sense in Latin: as trial, attempt, endeavor; as trying and never quite succeeding to produce text in any given literary genre. It is also, in Carrión’s own explanation: ‘un ensayo en el doble sentido de la palabra: un experimento abierto, que anuncia la puesta en escena futura que sí será cerrada, definitiva, en forma de libro; y un ejercicio de pensamiento creativo y en colaboración con humanos y con no-humanos [an essay in the double sense of the word: an open experiment, that heralds its future as the closed, definite, form of the book; and a creative thinking excercise, undertaken in collaboration with humans and non-humans alike]’ (2023: 42).

Realizing this made me understand the issue with my own writing with ChatGPT-4. Even if the Carrión that pens the introduction as a human sounds a little bit too close to the following text produced by the machine, both styles were still discernible. The literary success of GPT-2 and GPT-3 resided in their weirdness, even if fleeting, or perhaps even because of our realization that this weirdness was about to disappear from future models. Concluding his introduction to the text, the human Carrión predicts that:

[…] los algoritmos ya escriben mejor que nosotros, pero en su propio lenguaje, en su exclusivo idioma. Nosotros solo tenemos acceso a una adaptación, a una traducción, una suerte de versión balbuceante o infantil, aunque consideremos que es la tecnología la que está en su infancia. El laboratorio tiene forma de parque de juegos: ¿pero quién está jugando con quién? [algorithms already write better than we do, but in their own language, in their exclusive tongue. We only have access to an adaptation, a translation, a kind of babbling or child version of this language, even if we tend to think of this technology as if still being in its infancy. The lab has taken the form of a playground: but who is playing with whom?] (2023: 47).

What happens when the models sound like humans, and not just because of their increased sophistication in their mimicking of our writing, but because we end up writing like them? Had GPT-4 turbo achieved total human likeness or had I simply been re-trained to sound like a bot myself? I know it didn’t get Stiegler, but it surely sounded like a confident human who did. Am I unable to tell the difference between the two anymore? Am I that confident human now?

As I mentioned earlier in this essay, Carrión’s GPT-powered writing enacts what Stiegler calls ‘exosomatization’: the transfer of memory, thinking, and thus subjectivity to external, technical aids. In the case of algorithmic aids, the transfer is not neutral, since the outsourcing involves an adaptation into the technologies and their algorithmic shapes. As I experienced in the case of me attempting to write this essay with commercially available generative AI, when writing is outsourced to predictive models, the human writer becomes a sort of curator of a corporate product rather than creator, an editor rather than a point of origin (or originality, for that matter). Writing could no longer be conceived as the privileged site of interiority, or even of its technical mediation, but a distributed process across humans, nonhuman agents and corporate products, where the latter are outpacing the first.

This shift reconfigures the ontology of writing itself. Language becomes code; text becomes data; data becomes product; the author becomes a trainer, until no more. As Carrión suggests, we may be approaching a moment in which machines not only co-write with us but ‘incluso más allá de nosotros, como máquinas de escribir autónomas… con o sin nuestra ayuda [even beyond ourselves, as autonomous writing machines … with or without our help]’ (2023: 19, translation my own). This is a transformation of the writing act, an act that has been dislodged from human writing, and which perhaps should then receive a different name altogether, though who am I to try to name such a thing—others might, others have.³ What was once seen as a human technology (for what are we if not technological beings) has now become an emergent property of human-machine-corporate interaction, ready to leave us behind.

Conclusion: Toward an AI Prosthetic Culture

Is this sad? I don’t know. Probably yes. And yet, Yuk Hui proposes a reframing of AI not as a threat but as a prosthesis, a technological extension that can ‘assist us in realizing human potential’ (2023a). If we were to stop anthropomorphizing machines, dwelling on debates around their capacity to think, for example, and instead cultivate a ‘culture of prosthesis’ (2023a), we could perhaps recalibrate our relations with these AI models. Hui proposes we stop reducing machines to either being in competition with us or reading them as simple patterns of consumption, to rather understand them instead as part of ‘our current technical reality and its relation to diverse human realities, so that this technical reality can be integrated with them to maintain and reproduce biodiversity, noodiversity, and technodiversity’ (2023a). Hui understands noodiversity as human’s different ways of thinking (noo coming from the Greek nous, as in ‘mind’ or ‘intellect’), which ‘has been maintained by technodiversity, a variety of ways of understanding and constructing technology in different cultures, emerging from the locality and always in exchange with other localities’ (2023b). Hui believes that understanding the current technical reality in such a way would allow us to liberate machines from the many apocalyptic views that have emerged around AI, (2023a) but to do so we must continue to contextualize what concrete instantiations of generative AI we are discussing each time, looking closely at its entanglements in the world. As I mentioned earlier, and echoing Kranzberg again, technology is never neutral and thus any experiment with ethical ways to live with machines must understand the unethical ways in which commercial AI is deployed in the world today. I want to say yes to Hui’s technodiversity claims, but this involves accepting that this diversity is the corporative kind, and that noodiversity then involves a cleareyed vision of our thinking with that corporate logic as well.

Once the playing field is leveled, we may reclaim writing as a collective (or even collaborative) act, and not just with other humans but even perhaps with other forms of intelligence currently at play in our market driven ecosystem—in which, of course, we live and sometimes, even, play. This is a vision that resonates with Carrión’s final provocation regarding automatic textual generation and surrealism:

[…] ya no se trata de transformar en literatura una dimensión de la psique; sino de invitar definitivamente a nuestros exocerebros, a nuestros aliados tecnológicos, a nuestras inteligencias artificiales y compañeras a participar en el viejo arte de contar historias y desarrollar ideas y construir belleza [it is no longer a matter of transforming a dimension of the psyche into literature, but rather of inviting our exobrains, our technological allies, our artificial intelligences, and partners to participate in the old art of storytelling, developing ideas and constructing beauty once and for all] (2023: 19, translation my own).

Carrión echoes the idea of working with an external organ, the brain as an exteriority, an ‘exobrain,’ that extends the thinking process by becoming a partner in the act of thought. Contextualizing how that exobrain was made and why matters a great deal.

So, who is training whom as we think and write with machines? The answer is recursive. We train the machines, who reflect us back to ourselves, shaping our thoughts in return. In the current manifestation of corporate-run AI, the development and training of these machines reflects back a particular kind of humanity and transforms us accordingly—we live in the world we live. And yet. Does it have to be this way? As long as we don’t fail to interrogate who is training whom each time, whose patterns are being reproduced, whose desires are being encoded in the many loops, we may continue to put forward a vision of collaboration with machines where these provide for our augmentation rather than our subjugation, but one has to remain pragmatically cautious and suspicious. In any given culture of prosthesis, as Hui would have it, being critically aware may help us to short-circuit the loop and start again from a position of technodiversity that establishes a different pact between us, our writing, and our intelligent writing machines, corporate and not.

Funding

This work was partially supported by the Research Council of Norway through its Centers of Excellence scheme, project number 332643 (Center for Digital Narrative), and its SAMKUL project scheme, project number 335129 (Extending Digital Narrative).

Competing Interests

The author has no competing interests to declare.

AI Declaration Statement

I acknowledge the use of the multimodal large language model OpenAI GPT-40 via ChatGPT and Microsoft’s Copilot in Word, throughout the writing of this essay, and for the purpose of critically engaging the act of writing with generative artificial intelligence explored, which is the focus of the essay. I used GenAI in the prompts specified in the essay as such, as well as some less intrusive machine nudging (i.e., grammar, syntax, and spell check). All these instances are critically explored, and never offered without contextualization; the output from these prompts being used in my submission to demonstrate changes in writing within GenAI environments. I take full responsibility for the content of all AI generated outputs used in my research.