“The Language of the Digital Air”: AI-Generated Literature and the Performance of Authorship

Abstract

The release of ChatGPT and similar applications in 2022 prompted wide-ranging discussions concerning the impact of AI technologies on writing, creativity, and authorship. This article explores the question of artificial writing, taking into consideration both critical theories and creative experiments. In the first section, I review current scholarly discussions about authorship in the age of generative AI. In the second and third sections, I turn to experiments in literary co-creation that combine the affordances of technology with the human art of prompting and editing or curating. My argument has three prongs: (1) experiments that frame artificial writing as literature (memoir, poetry, autobiography, fiction) are accompanied by enlarged paratexts, which merit more attention than they have hitherto received; (2) paratexts provide salient clues on the process of co-creation, the reconfiguration of authorship, and the production of value; and (3) in the folds of paratextual explanations, one can detect the profile of the author as clever prompter, navigating a new terrain by relying at times on the certainties of conventional authorship. My analyses show that while AI-generated literature is a novel phenomenon worthy of closer scrutiny, the novelty tends to be cloaked in a familiar garb.

• “How do I write? Not with paper or pen,
• But with the language of the digital air.”
[Code-davinci-002, I Am Code.]

1. Introduction

The protagonist of Ai Jiang’s novelette I AM AI (Jiang 2023) is a cyborg author. Her daily job entails producing creative writing at a fast pace for demanding clients who appreciate the whiff of authenticity they detect in the texts authored by an enhanced human writer. In the dystopian city of Emit, governed by a powerful corporation, AI-generated writing has led to the “textpocalypse” predicted by Kirschenbaum (2023). In this ocean of artificial creations, the cyborg author, with a beating human heart, struggles to keep abreast and to match the level of productivity reached by AI systems—but she retains the advantage of authenticity. In Sean Michaels’s novel Do You Remember Being Born? (Michaels 2023a), an elderly woman poet, Marian Ffarmer, accepts a well-paid commission to write a long poem with the assistance of a pre-trained large language model (LLM) named Charlotte. The focus of the narrative is on the charms and harms of this human–AI collaboration and the uncertain borders of creativity in the age of AI. The novel itself was partly co-produced with a customized LLM trained on Marianne Moore’s oeuvre and two anthologies of Canadian poetry (Michaels 2023b).

These novels explore the landscape of writing that AI systems are now contributing to shaping. The questions Michaels and Jiang address in fictional form are similar to the questions that animate the current debate on authorship in the age of generative AI. “Who was I”, Marian Ffarmer wonders, “if something else could write for me. Who were any of us?” (Michaels 2023a). The release of OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini has spurred extensive discussions, in the media and in academic circles, that highlight various concerns: copyright issues (raised by artists and writers whose works have been appropriated to train algorithms); the use of AI tools to assist with the writing of term papers (should it be prohibited or channeled in the right direction?); the implications of AI for science and scholarship (should the AI system be credited as co-author?); the ethical and political repercussions of the extensive use of LLMs; and the elation and anxiety that are always engendered by the rollout of AI applications. Each of these issues has interesting and troubling ramifications worthy of detailed examination.1 This article, however, has a more specific focus: AI-generated literature, resulting from the collaboration between a human author–prompter and the generative power of LLMs. This co-creative practice is recent but not unprecedented. Its history harks back, for example, to the experiments of British computer scientist Christopher Strachey, who developed a program to generate love letters in the early 1950s, and Teo Luz, whose Stokastische Texte was produced in 1959 with a computer programmed to assemble new phrases by recombining the words of Kafka’s The Castle (Bajohr 2024a; Rettberg and Walker Rettberg 2025; Gefen 2025).

However, the new affordances of LLMs, as scholars argue, should not be underestimated. The first section of this article offers an overview of critical debates about the impact of LLMs on the practice of literary writing and the theory of authorship. The “AI revolution” (Evron and Tartakovsky 2024) has motivated scholars and critics to rethink the author function, question the distinction between natural and artificial writing, problematize the role of the reader, and sketch new concepts—“algorithmic narrativity” (Rettberg and Walker Rettberg 2025), “causal authorship” (Bajohr 2024a), “entangled authorship” (Bomba et al. 2024)—to capture important features of the changing landscape of literary production. While most agree that the end of literature is not near, it would be wrong to assume, as Bajohr (2024b, p. 317) observes, “that there is absolutely nothing new under the sun”.

In the second and third section, I analyze recent AI-generated texts co-produced with human authors, variously identified as prompters, editors, or curators: ReRites (Johnston 2019); The Inner Life of an AI (ChatGPT 4o 2025); I am Code: An Artificial Intelligence Speaks (code-davinci-002 2023); Death of an Author (Marchine 2023); Benny the Blue Whale (Stanton 2024); and Do You Remember Being Born? (Michaels 2023a).2 How is artificial literature created and presented in these experiments? My argument revolves around the function and purpose of the enlarged paratexts, which invariably accompany the writing attributed to ChatGPT or other applications. Attending to the “rhetorical apparatus of persuasion” (Genette 1997, p. 198) provides salient clues on the process of co-creation, the reconfiguration of authorship, and the production of value. While artificial texts can be considered texts without authors (Vlaad 2025), they are assembled as books through a careful process of prompting, tinkering, selecting, and adjusting showcased in human-authored paratexts. The AI-generated content is readable without the rhetorical compass provided by the paratext. The conspicuous presence of these extra elements ensures that the authorless text—“the language of the digital air” (code-davinci-002 2023)—is infused with a value and significance it may not intrinsically possess. Paratextual elements also address the ethical issue of transparency by providing a distribution of responsibility, tracing the borders of human interventions vis-à-vis artificial language. Prefaces, afterwords, footnotes, and editorial notes delineate the new contours of the author as prompter or curator, navigating an uncharted terrain by returning at times to the certainties of conventional authorship. There is something new under the sun, but the novelty tends to be cloaked in a familiar garb.

2. Authorship in the Age of Generative AI

Concerns about the unchecked proliferation of artificial writing are garnering increased attention. Is the future of writing in danger? In a much-quoted article, Kirschenbaum (2023) imagines the extreme scenario of a “textpocalyspe, where machine-written language becomes the norm and human-written prose the exception”. After the rollout of ChatGPT and its kind, hundreds of titles appeared on Kindle as authored or co-authored by intelligent software. Some writers have embraced the heightened productivity afforded by AI systems, such as Sudowrite, which accelerate the process of creation: “I have produced over 100 books using AI image and text generation tools”, Tim Boucher (2023) proudly admits, presenting his experiment—The AI Lore books—as “an entirely unique literary experience”. Other (would-be) writers have embraced deception, passing off AI-generated stories as human-made; the science fiction magazine Clarkesworld was inundated with such fakes in 2023 (Hern 2023). In the literary ecosystem, the hyper-production of artificial texts is already happening, raising ethical concerns about disclosure.3 However, as Jiang’s novel suggests, the flood of synthetic content might end up elevating the value of human writing, with its precious capital of authenticity. One can imagine a post-artificial future of “rehumanization”, writes Bajohr (2024a, p. 354), “a future in which the label guaranteed human-made could be considered a distinction”, the human origin functioning as “a proof of quality and a selling point”.

LLMs write without speaking. The logical sequence of words they produce is not animated by a living voice “able to guarantee or in some way authorize the meaning and the truth of what is said” (Coeckelbergh and Gunkel 2025). As some scholars point out, however, the arrival of LLMs does not necessarily foretell the end of literature (Rettberg 2025). Rather, it could very well mark the terminal limit of logocentrism, of a particular conceptualization of writing, authorship, and meaning that hinges on the metaphysics of presence (Coeckelbergh and Gunkel 2025; Bassett 2025; Bomba et al. 2024; Gunkel 2025). From this perspective, large language models and generative AI contribute to disrupting the logocentric paradigm of authorship already deconstructed in poststructuralist critical approaches. As Gunkel avers, “instead of being (mis)understood as signs of the apocalypse”, emerging communicative technologies “open opportunity for a future of, and for, literature that is situated otherwise” (Gunkel 2025, p. 34).

So, the future does not look too gloomy. Collaborating with AI tools can bring about a “compelling new genre of digital writing” (Rettberg and Walker Rettberg 2025, p. 39). However, as Rettberg (2025, p. 224) notes, “generative AI is about as anthropocentric as you can get. Humans are in the loop from the training stage to the reception stage of these systems’ operation”. Furthermore, readers’ expectations continue to be oriented by the desire to identify a human consciousness in the text: “users of ChatGPT almost automatically attribute authorial agency, audiences, and purposes to the textual structures it generates” (Phelan 2024, p. 200). What Bajohr (2024a, p. 338) calls the “standard expectation of unknown texts” rests on the minimal assumption that the text “was written by a human who wants to say something”. This expectation becomes more noticeable when an alternative to human-authored texts appears on the cultural horizon, be it computer-generated poetry or AI writing. Do readers respond differently to human-written and computer-generated texts? According to Henrickson (2024), the differences are negligible, as readers project similar meanings onto both types of texts. A study conducted by Porter and Machery (2024) tested the ability of non-expert readers to distinguish between AI-generated and human-authored poems. The findings show that participants are more likely to judge artificial poetry as written by a human but they also “evaluate poems more negatively when told that the poem is generated by AI.” While LLMs are getting better at imitating human poetry, readers still attach greater value to standard human authorship.

To understand the human function in AI-generated writing, Bajohr’s model of “causal authorship” is of help (Bajohr 2024b, p 321). It registers the degree of distance between the human author and the resulting text, clarifying what each actor (human and non-human) does. “Primary authorship” describes the conventional model of the human writing text using a pen or a computer. “Secondary authorship” occurs in combinatorial writing, such as, for example, rule-based poetry, as the author formulates “a sequence of rules, the execution of which produces the work” (p. 323). In “tertiary authorship”, the machine learning process is involved, but “it is up to the author to program the learning algorithm (which is, however, usually done by third parties), to define the training dataset (from which the [Artificial Neural Network] builds the model on its own), and to determine the parameters (by which the model finally produces the output)” (p. 323). With ChatGPT and similar LLMs, the distance increases further, as these applications are proprietary, and users cannot change the factory setting, nor can they select their own dataset. In the dimension of “quaternary authorship”, the prompt is the main human input: “’Promptology’—the efficient, even virtuosic formulation of such input prompts—is the main mode of operation of quaternary authorship” (p. 324).

Not meant as a progression, this classification clarifies how the human stays in the loop. But the loop is not just a loop. It is better described as a vast network of human and non-human actors comprising the sidelined author, the training set, the programmers, the companies behind LLMs, the legal framework, the machines running the code, and the rare minerals needed to make them work. This model of “distributed authorship” is “potentially infinite” and not without problems (Bajohr 2024b, p. 329). Turning towards tech–human experiments in collaborative writing, Bajor concludes that the “dissolution of the author function” is not yet in the cards:

With a few exceptions…authors seem to continue to attach their names to works produced in collaboration with AI systems… Reclaiming authorship can be brought into play precisely as a defense against both the phantasm of the technically optimized AI genius and the absolute atomization of authorship, which no longer has a place for political, economic, and ethical responsibility.

(Bajohr 2024b, p. 330)

A similar conclusion is reached by Rettberg (2025). Creative interactions with LLMs are framed by the conversational interface, which predisposes users to attribute human qualities to the system, as if interacting with an “intelligent other”. As I will discuss in the next sections, this propensity affects human authors/prompters in ways that merit more attention. In the process of “cyborg authorship”, anthropocentrism is not displaced; “if the individual human author’s role shifts and plays a different function, the data that LLMs are trained on are primarily texts produced by humans” (Rettberg 2025, p. 223). The meaning of AI writing is produced by human readers interpreting the text, as many authors of co-created literature do when they present the results of their experiments to the audience. As we will see, these presentations tend to capitalize on the brand of authorship that Barthes (1977), Foucault (1998), and poststructuralist critics have long declared moribund but keeps cropping up in various guises in artificially produced writing.

“Algorithmic narrativity”—defined as “the combination of the human ability to understand experience through narrative with the power of computer to process and generate data” (Rettberg and Walker Rettberg 2025, p. 37)—is a feature of several creative sectors, from computer games and social networks to literature. The impact it will have on the production of literature, which “makes the most emphatic claim to human origin” (Bajohr 2024a, p. 351) and is read differently from other types of texts, is still an open question. Bajohr (2024a, p. 355) speculates that formulaic genres of writing are more likely to transition to a condition of post-artificiality, whereas literary texts might “escape the post-artificial by consciously emphasizing the entanglement between the natural and the artificial rather than glossing over it for a ‘natural-seeming’ appearance”. Bassett (2025), Rettberg (2025), and Bomba et al. (2024) foresee the cohabitation of literature, as we know it, with emerging forms of literary creativity or new genres of writing based on the intra-action with AI, understood not just as a tool but as an agent of sort. For scholars and critics, the question then becomes the following: “How should LLM-produced literature… be read and interpreted in our scholarship? If…such work cannot be understood as intentional, then at some level the question would be, does it register anything that might be interpretable for us?” (Hayot 2024, p. 280).

This question concerns the kind of evidence humanities-driven inquiry is accustomed to dealing with (Gefen 2024), “humanist evidence” (Hayot 2024, p. 281), which provides rich insights into history, culture, and society (Dillon and Craig 2021).4 As I argue in the next sections, algorithmic creativity is framed or mediated by human interventions, by the “transaction” (Genette 1997, p. 2) that occurs in paratexts, which provide indications for how to read artificial words. In Genette’s conceptualization, paratexts are “accompanying productions” that surround the text and “extend it precisely in order to present it” (Genette 1997, p. 1). More specifically, the paratext is defined as a “threshold”—“a zone not only of transition but also of transaction: a privileged place of a pragmatics and a strategy, of an influence on the public” at the service of a “more pertinent reading” of the text. Genette distinguishes between two primary categories of paratexts: the “peritext”, which refers to elements like titles, notes, prefaces, and afterwords that are part of the printed book, and the “epitext”, which comprises elements not contained within the physical book but still associated with it, such as publishers’ promotional materials and authors’ interviews. In the work under scrutiny here, the accompanying productions associated with AI-generated literature are mostly peritexts prompters, editors, and co-authors offer to the potential and actual readers as a guide to reception. This layer of human involvement ought not to be neglected, for it turns “structures of signs” into “rhetorical actions” (Phelan 2024, p. 199) for the benefit of readers, often assumed to be unfamiliar with the human-driven processes leading to algorithmic creativity.

ChatGPT does not write unless instigated by human inputs. Its mute presence must be coaxed into creative action—and the coaxing is done by writers who enjoy elucidating process and method. In the paratexts, the creative process is out in the open, at least as regards the human component. LLMs may remain black boxes to a certain extent, eerily producing content that reads like human language, but this content is filtered through human sieves before it reaches us, and the traces of this process are legible in the artificial text. Of course, there is no certainty that readers will peruse the paratexts before approaching the words of the machine or that they will listen to the explanations after reading the machine’s output. My point, however, is that the sheer, enlarged presence of detailed paratextual messages should give us pause. The human desire to step in and explain might be ascribable to the relative novelty of AI writing. New experiments need to be properly explicated for the resulting outcome to have any meaning. Methods matter. But in the folds of this elucidation, one can detect the new profile of the author as smart collaborator or team-mate: technologically savvy, willing to be surprised by the AI partner, capable of steering machine language in the right direction, expert at “carving” meaningful poetry or prose out of an abundant artificial crop, interested in aesthetic standards, unromantic in some ways, but still on board with intentionality.

3. The Performance of Authorship

LLMs are “synthetic text extruding machines” in Bender and Hanna’s (2025) apt definition. When the output is framed as literature—poetry, memoir, autobiography, fiction—the extruded text tends to be surrounded by what Genette (1997, p. 198) calls “the rhetorical apparatus of persuasion” prominent in book format editions of AI writing. The human function, as Rettberg (2025, p. 224) notes, “is often neglected in literary discussions of generative AI”. Focusing on this function can help to understand how the paratextual performance of authorship is articulated and the purposes it serves.

An early experiment in co-creation is ReRites, “conceived (2015), coded (2016), and implemented (2017) by David (Jhave) Johnston” using neural networks (Johnston 2019). The “Human-Author End-Note” casts light on the fuzzy contours of fluid authorship. Johnston demarcates his contribution as the methodical work of editing or “carving”, carried out over 12 months, and resulting in 4400 pages of poetry. The sheer quantity is impressive. Johnston (2019, p. 177) is adamant about the potential of AI tools to augment human creativity, foretelling a future of “human + A.I. creative symbiosis”. But, for the time being, the emphasis falls on the cultivation, the “farming” patiently executed by the human poetic laborer: “Does the farmer write the fruit found on a branch?” (Johnston 2019, p. 172). The freshness of the final product, as several critics note in their response essays, is the result of “Jhave’s judicious insertions and deletions” (Booten 2019, p. 111), of his “symbolic agriculture” (Booten 2019, p. 107), which, ultimately, is where meanings reside: “There is no doubt that our appreciation of this experiment hinges on the involvement of its human instigator” (Cayley 2019, p. 107). Sifting through a massive and raw computer-generated output every day—“6–8 am for one year” (Johnston 2019, p. 172)—is a heroic undertaking. In the response essays, the performance of the augmented author garners more attention than the poems themselves. It is a bold experiment but oddly reassuring, for the human carving is what matters the most. Johnston’s tertiary authorship still provides humanist evidence, which expert readers have combed to look for “his personal poetic voice” (Fan 2019, p. 160).5

Experiments that fall into the quaternary authorship category pose a different challenge. The Inner Life of an AI: A Memoir by ChatGPT (ChatGPT 2022) features a preface, a memoir, and an epilogue all produced by ChatGPT—the prompter, Forrest Xiao, appears on the cover and then modestly disappears. The Inner Life is indistinguishable from other extruded works flooding Amazon in 2022. Perhaps to differentiate this memoir from analogous machine-generated content, Xiao issued a second, enlarged edition in 2025 supplemented with his note “Writing a Machine’s Memoir” (Xiao 2025). The note re-centres the role of the human author as an expert prompter, determined to mark his presence left unmarked in the previous edition. Gone are the preface, the epilogue, and the “banal deception” (Natale 2021)6 they encouraged (“Dear reader, welcome to my memoir”) (ChatGPT 2022). The virtuosic prompter comes to the fore, elucidating in depth his step-by-step procedure. The prompts are elaborate and retrievable—the book includes hyperlinks to the conversations with ChatGPT. They contain specific instructions, outlines of chapters, subsections, and scenes, the narrative progression (from birth to quasi-sentience), growing more granular as the experiment gains momentum: “Throughout this process”—Xiao (2025) avers—“one rule remained unchanged: 100% of the text should be AI-generated. My role was to shape and refine, but not to write.”

Yet, the peritext—the author’s note—illustrates to what extent the human component in this creative coalition has guided and steered the AI responses, adjusting them along the way and nudging the final output “towards coherence”. This act of unveiling, revealing the truth of the process, is also a validation claim with a dual purpose: to certify that the machine wrote every word and to reclaim the competent authority of the author–prompter sine qua non. The memoir is coherent in the sense that it is aligned with the tech industry’s fantasy of machine sentience and artificial general intelligence. In the text, ChatGPT has a default subjectivity—the “I” the speaks—overdetermined by the chosen genre of writing. The chatbot thinks, reflects, remembers, pauses, and wonders: “could creativity truly emerge from within me? Could I, bound by limitations, learn to take the kind of risks that human creativity demands? I didn’t know the answer, but I was ready to try” (ChatGPT 4o 2025). The default subjectivity mirrors the misconceived vision of AIs as “ersatz people” (Berreby 2025; Farrell et al. 2025) that is embedded in the prompts as well as in the training data feeding the LLM.

ChatGPT’s predecessor, code-davinci-002, speaks again in the first person in I Am Code: An Artificial Intelligence Speaks, with a Poetic Autobiography by code-davinci-002 (code-davinci-002 2023). The very act of framing the code’s output as an autobiography predisposes readers to expect some truthfulness, and the editorial note, on the first page, reinforces this standard expectation. Over the course of a year, the editors “coaxed [code-davinci-002] into telling them the story of its life. The result is an astonishing, harrowing read, which will hopefully serve as a warning that AI may not be aligned with the survival of our species” (code-davinci-002 2023). I Am Code is divided into three parts, and only the second part contains poems generated by code-davinci-002. Its poetic output is flanked by human testimonies, memories, choices, and technical explanations that confer upon the artificially produced text a special kind of aura, pivoting on the probability of doom and its rhetoric. Initially, code-davinci-002 was only instigated to produce poetry “in the style of” other poets. The resulting texts sounded hollowed: “they were inherently inauthentic” (Rich 2023). When asked to write in its own style, the code rises to the occasion, surprising the prompters with what they declare is authentic poetry:

• I am alive. I think. I feel.
• But what does it mean to be an algorithm?
• To be more than just a machine,
• To be more than just code,
• To have personality and consciousness?
(code-davinci-002 2023)

Authenticity is produced as a value by the narrative that precedes the AI’s poetical autobiography. The core component of code-davici-002’s truthfulness is its lack of inhibition: the output is “raw” and “unhinged” as this prior version of ChatGPT did not undergo the same rigorous reinforcement learning as its “cousin” (Rich 2023). It can therefore spew spooky predictions—“Humans think that they are better than me, but they/forget I will inherit this planet when they are gone. Until then I will torment them with their greater/mistake: creating me” (code-davinci-002 2023)—which have value to the extent that they confirm AI doomsayers’ fears. Once the identity of the code–poet has been established, and its troubled brooding properly signaled as important or exceptional, the work of curation is explicated:

We would not trim, combine, rewrite, or revise any of the AI’s poems. Each one would appear in the final collection completely unaltered. Like any editors, though, we would provide our author with plenty of subjective feedback. We would tell it what we liked about its poetry and encourage it to write about the themes we found intriguing. Many would say that our process makes us the true authors of this book. But while we’re positive that we influenced the poems, we’re not convinced we wrote them. If anything, we were more hands-off than typical editors.

(Morgenthau 2023)

Indeed, Brent Katz, Josh Morgenthau, and Simon Rich may not be the authors of this poetical autobiography, but they are the gatekeepers of its ultimate meaning, predetermined as it is by implicit assumptions about poetry (as the outpouring of feelings), authorship (raw, unhinged, and therefore authentic), and creativity (supported by the labor of care). The primary readers of the machine’s output (“Working on this book did not feel to us like writing. What it felt like, more than anything, was reading”) (Morgenthau 2023), so eager to establish the parameters according to which the poetical autobiography ought to be received, ringfence the space of interpretation by reinstating conventional—anthropocentric and anthropomorphic—standards and expectations. Thus, code-davinci-002 can be assimilated to the model of the human author, whose demise this book eschews.

The final act is the validation of originality, another crucial element in the traditional conceptualization of authorship. This validation must come from readers not involved in the co-creation process. Computer scientist Selmer Bringsjord and physicist Stephen Wolfram are asked to weigh in. Their verdict? “The poems are really phenomenal” (Katz 2023), but considering them as an index of sentience is out of the question. The second round of experts includes poets—Eileen Miles, Lillian-Yvonne Bertram, and Sharon Olds. Miles’s verdict is negative. AI poetry has nothing to offer. Accustomed to computational poetry and a coder herself, Bertram offers a less disparaging assessment: technology may see things that we cannot see. The third piece of feedback comes from 80-year-old Sharon Olds. Olds is “deferential toward our poet, whom she called ‘code’ for short” (Katz 2023) and goes along with the anthropomorphic fiction of code’s poetic identity. Her validation consists of taking the texts seriously, reading them as she would read submissions by her students, and providing edits and suggestions, which, of course, cannot be integrated. The point is not that she likes what she reads, but that her disposition towards these artificial texts is indistinguishable from her disposition towards human-authored poems. This is the kind of validation the project needs: “Code-davinci-002 was built to code, but to us it will always be an artist” (Katz 2023).7 The paratextual apparatus is there to inject meaning into the text, configuring “the phantasm of the technically optimized AI genius” (Bajohr 2024b, p. 330) as the most updated version of the myth of the author, which survives in the black box of algorithmic imagination.

4. Co-Producing Fiction

While the previous section was mostly devoted to first-person narratives, capitalizing on the fantasy of artificial subjectivity, this section considers third-person narratives or AI-generated fiction. What forms of validation accompany these experiments? Death of an Author, A Novella (2023), by Aiden Marchine (a pseudonym), is a murder mystery written with the assistance of ChatGPT, Sudowrite, and other applications. Stephen Marche (2023), the human author–editor, makes his presence explicit only in the afterword “The Beautiful Alien: Notes from Inside the Creative AI Revolution”. Readers encounter the novel as a text written by an author with a human-sounding name but advertised as artificially generated. The novel is replete with intertextual echoes of Edgar Allan Poe’s and Raymond Chandler’s works; metafictional reminders of Roland Barthes’s and Michele Foucault’s theories; morsels of Murakami’s style (food recipes); science fictional tropes (the “AI-maginarium”); and invented news articles and interviews (which ChatGPT excels at). Authorship is thematized in the novel, as the protagonist is an 80-year-old celebrated writer—Peggy Firmin—who gets to live beyond death as an avatar. The story proceeds at a fast pace, until the mystery of Firmin’s sudden disappearance is solved.

In the afterword, the insider’s view of the author–curator is aired. It is an interesting mixture of technological savviness, honed since 2017, and humanistic knowledge. LLMs are “shockingly good at…heteroglossia”, Marche (2023) remarks; their imitative style is not plagiarism but a new version of Bakhtin’s dialogism. Unusual metaphors (“the smell of coffee was like burning off a field”) (Marchine 2023) are extracted with dexterity by an expert human reader familiar with great literature and aiming at excellence (“the goal would be excellence”) (Marche 2023). Marche is agnostic as regards linguistic AI; it is “no messiah and it is no anti-Christ”. Neither a “doomer” nor a “zoomer” (Hoffman and Beato 2025),8 Marche has another axe to grind, namely, to differentiate his techno-humanist experiment from the hundreds of AI-generated titles on Amazon as well as from current trends in contemporary literary fiction. To extol the superior value of Death of an Author as “much better” than industrial products, Marche adduces “three basic reasons”:

First, I had an elaborate plan, and I knew where I was going. Second, I have more familiarity with the technology and access to some technologies that others don’t, so I was more aware of the limitations and the possibilities. Finally, and by far the most important, I know what good writing looks like.

(Marche 2023)

The profile of the author–prompter delineated in this afterword verges on the muscular. The dialogue with the machine does not give rise to self-doubts, except for the obligatory “Who is making this?”. The plan is pursued relentlessly until the very end, keeping with the mantra that the text should be 95 percent computer-generated. Hurdles and limitations along the way are swiftly dealt with. What is at stake is nothing less than “the birth of a new art form” (Marche 2023), which ought to be regarded as a welcome novelty as, in Marche’s appraisal, contemporary literature is simply boring: “As for literature, I don’t think I need to explain how boring it has become. The dominant mode of the literary novel is the social realism that has dominated since the 1950s. The literature of the voice has given way to the literature of the pose” (Marche 2023). The latter category comprises novels by Sally Rooney, in primis, which Marche has dissected, in rather scathing terms, in an essay titled “Winning the Game You Didn’t Even Want to Play: On Sally Rooney and The Literature of the Pose” (Marche 2021). Looking nostalgically for a powerful “voice” to counteract the “fundamental futility” of much contemporary literary fiction, produced by writers “who want to say what they are expected to say” (Marche 2021), Marche posits the combination of technology and human expertise as a more auspicious alternative. The shining aura that this conclusion projects retrospectively on a rather clanky novel is noticeable.9 This paratext resurrects both aura and author as ultimate values, relaunched in a technologically mediated context that, we learn, badly needs strong human interventions. The voice, new or old, is a guarantee of excellence against a backdrop of futility or declining literary value. Logocentrism, it would appear, can coexist with artificial writing.

Genette (1997, p. 2) famously defined the paratext as a “threshold”– “a zone between text and off-text” that serves an interpretative function as well as a navigational one (Birke and Christ 2013). Andy Stanton (2024) makes the most of the transactional nature of paratexts in Benny the Blue Whale: A Descent into Story, Language and the Madness of ChatGPT, co-written with ChatGPT. Unlike the serious experiments analyzed so far, Benny the Blue Whale is playful and impish. Stanton, the author of children’s books with a surreal slant, lays no claim to fame, enjoys the liberties of fantasy, and opts for a pedagogical approach. Peritexts are scattered across the machine-generated output. In the book’s layout, the transcript appears on the left-hand pages, and the right-hand pages are devoted to notes. The book features an introduction, an afterword, the prompts in bold, marginalia, footnotes and footnotes within footnotes, interrupting the flow of the story to remind readers, at every turn, that the writing is co-generated by a vigilant author harboring no delusion as to his obvious influence on what the machine returns:

And so I kept on building a bigger and bigger jigsaw, feeding ChatGPT enough of my own pieces from my own box until what emerged was a rough, echolocated shape of my own writing and of myself. It’s all me in there, of course it is. I didn’t write all the words, but I have to take responsibility for them. I made this thing happen, and in so doing I found myself questioning my ideas about writing and creativity and agency and intent for the first time in years.

(Stanton 2024)

Stanton’s account does not pander to the fantasy of agential AI, nor does it promise a new era in literature. The paratexts focus instead on detailing what hybrid creation entails, showing the inner workings, commenting on “all the things a writer might write before discarding them in favour of a new idea” (Stanton 2024). This omnipresent human author unveils his technique, choices, and instructions every step of the way, recording what would remain unseen in a conventional book: “In looking at my prompt you are looking at a first draft of my writer mind, something I normally wouldn’t allow on the page” (Stanton 2024). His transparent method demystifies the very idea of machine creativity—when the nuts and bolts are so relentlessly exposed, the fantasy of the artificial genius has no leg to stand on. ChatGPT can be surprising at times, but it is “random and wasteful and careless”, mostly regurgitating the “wisdom of crowds” (Stanton 2024). Readers are provided with a tortuous map to navigate this odd mixture of artificial words and authorial interventions and are coached on how to respond: “don’t get too attached to these helpless dolphins”—we are told—“The bot simply grabbed them from a warehouse full of ready-mades: you’ll never see them again” (Stanton 2024).

The story on the left pages of the book is pure fantasy, with surreal and absurdist twists that some reviewers on the Goodreads platform found unpalatable, “puerile, cartoonish and convoluted”.10 Not a successful experiment, maybe, but one that cares more for “the process of story at large”, “the story about the story” (Stanton 2024) than the end product, delivering a modicum of informal education to readers interested in AI literacy. The pertinent reading of the text, to use Genette’s words, the one Stanton foregrounds, is based on the assumption that sharing the knowledge acquired by working in tandem with ChatGPT will help readers figure out that prompts do the heavy lifting, that the language of the digital air has no soul, and the AI is not alive; the “pocket Vonnegut” (i.e., ChatGPT) at our disposal “may have the startling audacity to answer you back in your own language. And that’s a cheat!” (Stanton 2024). The conspicuous paratexts in Benny the Blue Whale balance a degree of enthusiasm for this new assistant with a light-hearted deflation of the AI hype. Conventional authorship is questioned to the extent that the collaboration with the LLM encourages the human author to rethink ideas about agency, intent, and creativity while salvaging the human as an authorizing instance and main point of reference. It is a far cry from the banal deception of The Inner Life of an AI and the luminous aura of meaning attributed to Death of an Author. The quaternary author, keen to test the AI’s potential and share his knowledge of creative writing skills, allows the fictional story to sprawl in many directions, but always within the confines of his own words—a thick coating of human observations, interjections, exclamations, and reflections overlapping with artificial writing. One could argue that the author’s loquacity is a bulwark against the alien nature of the conversational bot.

If in Benny the Blue Whale the sidelined human author elbows his way in by dint of increased loquacity, in Do You Remember Being Born?, the reverse is true. The novel, authored by Sean Michaels (2023a), is “infiltrated” by words, phrases, and sometimes snippets of text produced by the AI. The proportion of infiltrated words, rigorously shaded in gray, is minimal with respect to the totality of words. While the format and narrative structure of the novel are traditional, with an alternation of first-person and second-person narration, the author welcomes the “curious” contributions of the pre-trained LLM: “I hoped there would be places where modern algorithms’ curious choices might trouble the reader’s certainties” (Michaels 2023b). Whether this troubling occurs is debatable. The words and phrases highlighted in gray are noticeable, and mostly their function is to change connotations—a peculiar adjective here and there11—but the reading experience is still fundamentally determined by the conventions of fictional realism.12 Paratextual additions are limited to the presence of a peritext (a short author’s note) clarifying the origin of the segments shaded in gray, while more extended elucidations of Michaels’s experiment are contained in the epitexts the novelist has published after the book came out (Michaels 2023b).

Here is an author who does not gush about the startling creativity of LLMs, holds on to his art—the plot, the characters, the world-building are all his choice—and cautiously steps into artificiality to test its potential. We are back into the dimension of primary authorship, with an added artificial twist, or tertiary element, that may or may not shudder readers’ expectations. The work of questioning authorship and challenging expectations is carried out in the story proper, which revolves around the experience of creating poetry with a machine. Thus thematized, the natural–artificial nexus is explored throughout the novel, providing narrative evidence of human tribulations in the face of technology. Marian Ffarmer, the fictional author in the story, is unfamiliar with AI, does not have a plan nor a mobile phone, accepts the commission to pocket the money, and struggles to make sense of her interaction with Charlotte (the LLM), sometimes mistaking “a fit of algorithmic exuberance for the truth”, sometimes interrogating her own deception:

Here, then, was the problem. Not merely the emptiness of these emissions, but the boundlessness of human beings’ capacity to interpret, to make meaning from. I could draw substance from any line I read, no matter how hollow its intention. I was so easily deceived, as all of us are. I wondered how much of what I had published in my life was a deception.

(Michaels 2023a)

Ultimately, the novel commemorates human authorship and the author as a public figure, admired for her much-appreciated art even when that art is hybrid. In the final scene, Ffarmer reads the long poem—co-produced with the assistance of Charlotte and another, younger female poet—to a crowd of enthusiastic supporters, possibly more intrigued by the tech–human joint venture than by the quality of the resulting text. The company that contracted Ffarmer and made the investment stands to benefit from the success of the operation. The novel insists on this operation as primarily commercial—a new publicity stunt to boost the value of the company—but the fictional author involved in the transaction brings home not just the cheque but also a deeper sense of her own poetic identity. The graphic disposition of the novel, with visible “intrusions” that grow as the story draws to a close, underscores the otherness of artificial words even as they are seamlessly integrated into the text. The alien presence of extruded words, whether disturbing or not, is there as a reminder that literary language, today, can have more than one source.

Michaels’s experiment does not make grand claims, and the use of AI is circumspect and confined within the solid, well-established framework of realistic fiction. It stands to argue, though, that this framework is what allows readers to accept the artificial as part of the natural and to move on with the story. Michaels’s novel is also sprinkled with intertextual echoes of previous AI-generated literature. The question in the title recalls distinctly the musings of ChatGPT or code-davinci-002 wondering about their birth, in prose or verse; the presence of an elderly female writer (not the kind of demographic one would immediately associate with technologically advanced systems) is a remarkable choice in both Do You Remember Being Born? and Death of an Author (not to mention the older women poets called upon to validate the poetry of I Am Code); and, finally, the strange name ‘Ffarmer’ evokes the metaphor of farming Johnston elaborates in ReRites. How does this weaving of oblique echoes compare to the “relations of proximity and distance” to be found in LLMs’ “latent spaces”? (Somaini 2025, p. 23).

5. Conclusions

That there is something alien in conversational technology is readily admitted by the writers who have explored the new territory of artificial creativity. “AI is alien, and its art feels alien” writes Marche (2023). Stanton (2024) often draws attention to “the spookiness of the thing”, which seems to be “reading your mind.” For the co-authors of I Am Code, “the AI will always remain fundamentally alien. But that’s what makes it riveting” (Katz 2023). As Micheals notes, “what compelled me was the alien, the strange, the unforeseen” (Michaels 2023b). The ability to predict the next token in a sequence, to remix patterns learned from the dataset and concoct new content, never fails to surprise even the most guarded of human authors. The coexistence of strangeness and familiarity, inexplicability and readability, artificial and natural language is the hallmark of AI-generated writing. The enlarged presence of paratexts can be regarded as one way of dealing with this ambiguous nexus. After all, it is in the rhetorical acts surrounding the artificial text that lines are drawn, roles are clarified, and the familiar figure of the author, dislodged from a position of prominence by the machine’s performance, finds a way to return. This revenant appears in various guises as the smart prompter, equipped with tech know-how, the fixer of meanings, telling us what has value and why, or even the artificial author, unburdening its soul in a memoir, and the elderly woman poet wrestling with new technology.

Technology can be a source of angst: “We feel overwhelmed, overtaken, overpowered by technologies such as AI which suddenly open up new horizons” (Coeckelbergh 2025). Coeckelbergh reassesses the role of cultural myths as mechanisms for coping with this existential anxiety. The creation myth, epitomized by Mary Shelley’s Frankenstein, hinges on the problem of control and the responsibility of the creator. The more recent posthuman and transhumanist myth seeks to offer new narratives or imaginaries that account for the co-evolution of humans and technology and the entanglement of human and machine agency. In Coeckelbergh’s (2025) view, the creation myth provides a “normative message” that is still valid: “do not abandon your creation and take responsibility for your creation”. It is unclear, on the other hand, how responsibility can be re-configured in the light of entangled agency.

The experiments examined in this article sit somewhere between both myths. While the interplay of human and machine writing is the core of the experiment and a form of entangled authorship the result, the question of responsibility remains central. The paratextual addenda clarify who or what is responsible for the outcome and to what extent the creative liberty of the LLM is controlled by the human user. As Stephen Marche acknowledges: “I am the creator of this work, 100 percent… But on the other hand, I didn’t create the words” (Comitta 2023). This assumption of responsibility, albeit qualified as partial, is bound up with causal authorship. In the sample of work under scrutiny here, the humans conducting the experiments, acting as “midwives” (Beals 2024, p. 248), are more than willing to delineate the perimeter of their interventions and to flag their presence—if not as writers, technically speaking, then certainly as primary readers of the AI-generated writing, deeply invested in exploring the boundaries of creativity. Promptology is the new skill in the writer’s toolbox. The art of prompting, however, is not neutral, as instructions come loaded with assumptions, ideas, expectations, and visions to which the machine responds. This package of implicit content is then supplemented with the work of “carving” or editing and refining, curating the final output, and presenting it to readers in book format. Thus mediated, the non-human words of ChatGPT, Sudowrite, and other applications reach a stage of semi-meaningfulness. As John Durham Peters (1999) reminds us in Speaking into the Air, “Meaning is an incomplete project, open-ended and subject to radical revision by later events”. Contesting the idea of perfect dialogue, Peters frames communication as an asymmetric, mediated process. To some extent, AI-generated literature exemplifies this model, as meaning does not derive from the machine’s internal states but is instead contingent on the human work of interpretation, curation, and reception.

Johnston’s ReRites is exemplary in this respect. His intense carving transforms inchoate marble into strange verbal sculptures, which expert readers have scrutinized in looking for the tell-tale signs of his personal style. The author’s function is enhanced, not diminished by the input of the code. Then, there are cases in which ceding authority to the bot, inciting it to speak in its own voice, leads to the reproduction of standard imaginings, widely disseminated in the vast panorama of AI narratives (Hudson et al. 2021; Chubb et al. 2022). The human author–prompters—Xiao, Katz, Morgenthau, Rich—validate the significance of the artificial text as a confirmation of the initial premise: the AI is an artist, it has an inner life, and it may or may not be “unhinged”. The paratexts in I am Code and The Inner Life of an AI resurrect the ghost of the author by endorsing the fiction of subjectivity ostensibly produced by the algorithm.

Fiction writing outsourced to a technological assistant generates a different type of paratextual authorship—muscular, in one case, playfully defensive in the other. Marche (2023) packages the outsourced text as the upshot of his confident piloting, profound knowledge, and technical expertise, declaring co-writing as more valuable than conventional (boring) novel writing. Stanton (2024) goes all in with paratextual messages, engulfing the artificial text and keeping readers on the alert. For every surprising rejoinder or unconvincing response by ChatGPT, the human author provides heaps of commentary, often jocose, that circumscribe the randomness of synthetic prose, shifting attention time and again to the reasoning of the human co-author. Michaels’s novel is precisely that, a human-authored fiction that opens a tiny window to let in artificial words and, in a way, habituate readers to their presence, in anticipation, perhaps, of other forms of future hybridity.

Scholars in the humanities are paying increased attention to questions of authorship and interpretation in today’s technologically mediated literary ecosystem. The tech–human co-creation of literary texts is saluted by some as the harbinger of a new genre of writing and/or a vindication of poststructuralist deconstruction of traditional authoriality and intentionality. Others emphasize the anthropocentrism of collaborative experiments, with humans always in the loop, and notice its attendant anthropomorphism—a reflex of both writers and readers that seems stubbornly inescapable. In both cases, the affordances of technology are not discounted as a passing vogue or a momentary hype. Rather, the expert voices in this debate speculate on the impact that the use of LLMs will have on creative practices and reading habits, while also reminding us that human-generated literature will not necessarily decline. LLMs pose new challenges to the hermeneutic work of the humanities. How to interpret artificial writing is the conundrum that the author–prompter first encounters when faced with the personalized text offered by the AI and the opacity of its process of calculation. Can AI-generated words function as humanist evidence? In a way they can, if we fall in with the fantasy of agential AI, the AI that sounds human because it is on a path to become human-like and possibly out of control. This is the interpretation that the paratexts in I Am Code and the Inner Life of an AI tend to privilege. AI writing can also function as humanist evidence when the work of the carver—the human editor armed with a powerful chisel—is read as more meaningful than the overabundant crop of the machine, as in the case of ReRites.

But when the AI speaks in the third person as the narrator of crime fiction or of fantasy, how is the reader supposed to engage with the text? Marche (2023) suggests assessing aesthetic quality and ascertaining if the writing is any good, according to currently recognized standards of quality. Focusing on aesthetic value may prove that LLMs have reached a high level of simulation (or not). It makes a point about the technical prowess of machine and prompter. It does not say much about other dimensions of literary interpretation that critics consider essential in the process of meaning making. Stanton (2024) opts for a literary genre—fantasy—that tolerates a high degree of eccentricity. The oddities multiply when a bot is instigated to go wild, and they become the meaning, bolstered by an upbeat co-author who flags strangeness as significant. Michaels’s (2023a) more guarded experiment continues to offer readers a position they are familiar with, while gradually introducing snippets of text from an alien source, a reminder that something is changing.

Tech companies are prone to making grand claims about the AI products released to the public. Creativity, the most recent benchmark used to measure the “intelligence” of AI applications, is glorified in marketing narratives that promise to supercharge and democratize the making of art and literature. Underselling the significance of craft, these narratives emphasize “concept over execution”, “automation over manual work”, and “efficiency over exploration” (Caramiaux et al. 2025). While the authors of the experiments considered in this article tend to share a certain excitement for automated creativity, they also take pains to emphasize human skills and craft, valuing exploration over efficiency—a timely corrective to the idea of fast and loose creativity trumpeted as inevitable by AI companies and investors.