1 Introduction

The use of artificial intelligence (AI) to enhance productivity has rapidly gained traction across industry worldwide, including in publishing and related sectors. AI has been heralded for its ability to reliably undertake repetitive or dangerous tasks, thereby—theoretically at least—freeing humans from the shackles of boredom and/or risk [62]. Its application has expanded across all sectors, from industrial automation [37] and airport security [57] to driverless vehicles and assistance for people living with disability [61]. It is being trained to solve complex mathematical and medical problems, with or without human involvement, with claims of being able to provide solutions beyond the capacity of even the best human brains [45]. In the creative sphere, in which we work and research, generative AI (GnAI) is now capable of producing new and novel creative texts, music and images [13], even if these are not wholly original due to the very nature of current applications, which literally implement predictive modelling to existing content. No act is exempt, no matter how creative a practice or circumscribed a profession, not even journalism [35] or peer-reviewed academic publishing, it seems.

This paper focuses on the role of professional editors and editorial practices within the knowledge economy, particularly with regard to the hitherto underrepresented aspect of editorial labour that deals in, and with, ethics. It engages with significant ethical and industrial implications relating to advancements in AI, specifically the potential application of GnAI in creative industries, emphasising and illustrating the enduring value of editorial engagement in assisting authors to shape quality literature in ways that are ethical and sustainable. To date, there has not been much research into editors’ editorial expertise, possibly because that knowledge is largely tacit, often instinctive and likely individualistic—though this characterisation of professional editing is changing with the outsourcing of training to tertiary institutions and development of specialist publications such as Editing for Sensitivity, Diversity and Inclusion: A guide for professional editors [47]. It is also the result of the lack of research on the practices and processes of editing in the professional space. How ethical issues might manifest differently depending on whether human or GnAI processes are involved depends, first, on identifying best practice in editorial conduct. Human editors are trained—formally or on the job, through the age-old apprenticeship model [46]—to bring a critical perspective that can identify ethical considerations, even if the request to do so is not in the editorial brief. GnAI is not. To receive feedback or advice on an ethical quandary from GnAI, an author (or editor, depending on who is doing the prompting) needs first to be aware that such potential issues might exist.

Arguably, AI may be the latest “best-thing” in 2024, but with the decades-long evolution of AI in industry, society and art, the idea of it is not new. The precursors of AI can be found in the myths and legends of antiquity, as well as in contemporary storytelling. When ancient Chinese, Indian and Greek philosophers began formally to describe human thinking, they assumed these processes to be structured methods of deduction akin to mechanical manipulation of symbols inside our brains. By the seventeenth century, European philosophers including Thomas Hobbes and René Descartes were formally articulating hypotheses that would form the basis for development of AI, while Karl Marx’s Manifesto identified many of the tensions and issues we debate today. By the twentieth century, we saw AI beginning to appear in the popular imagination: notably in science fiction novels and silent movies—conspicuously in the 2-min silent film, Le Voyage dans Le Lune (A Trip to the Moon, 1902) by French director Marie-Georges-Jean Méliès (1861–1938), who drew on the literary works of Jules Verne and H. G. Wells. More recently, science fiction, and with it AI, emerged from a mid-century entanglement with horror genres into an increasingly sophisticated and diverse, immensely popular genre that is now more closely associated with the rise of technology. Today, genre fiction is far from the only category concerned with AI; recent literary examples include Klara and the Sun (2021) by Nobel-prize winner Kazuo Ishiguro and 12 Bytes: How we got here. Where we might go next (2021) by Jeanette Winterson.

Novelty and utility aside, professional editors are now face-to-face with the reality of AI, well before the writing profession and the publishing industry have managed properly to ask, let alone reach consensus, on fundamental questions about how we could or should use it. Or when, and why. Publishing is, again, at the forefront of technological advances “whose cultural influence promises to be no less revolutionary than the introduction of moveable type” [18], (p. 8). This time it is not “desktop publishing” disrupting the status quo but GnAI: presently, its use by authors (and its use of authors, with uncited works in copyright being used to refine its processing), but soon, surely, editors also.

We explore the ethics of engaging GnAI in the process of editing fiction—by in-house and freelance editors, or by writers in revision—and the ethical issues arising from its use. We focus specifically on literary fiction, what Ken Gelder [22] calls “Big L Literature” (p. 11), partly because of how this category exemplifies editorial (and authorial) excellence—Gérard Genette [23] refers to it as “simultaneously (and intentionally) aesthetic and technical” (p. 51)—and partly because editing commercial genre fiction, and for that matter academic papers, operates in a different market, with different expectations and a different author–editor relationship. Given that GnAI is specifically designed to provide the most likely text response, this is a tool for those wanting predictable and formulaic text. But for an industry oriented towards novelty, can a predictive model offer appropriate solutions? An extended conversation about creativity, how it develops within—and outside—contemporary creative industries, how it is supported—and thwarted—by the knowledge economy is outside the scope of this paper. It is, however, a debate we care deeply about, and a discussion we wish to prompt. It is also one that, crucially, may be jeopardised by the rise of GnAI, with its inherent challenge to authenticity, originality and current privileged practices of peer support and professional collaboration.

We ask: how might the rise of GnAI influence editorial practices, support or undermine the role of professional editors in the development of literary fiction? If we, or our authors, work with GnAI, is it still our/their work? How do we credit that collaboration? (Changing protocols regarding how to cite the use of AI in writing, from bibliographical additions to acknowledgements or upfront statement, reflects an increasing recognition of the importance—and increasing influence—of GnAI on academic and educational publishing.) Should its use be encouraged? Are editors (only, also) prompt writers? What are the potential benefits, and risks, for creative workers—writers and editors—in the current and evolving knowledge economy? Using data from an iterative, comparative, multi-stage test case over several months, we compare and contrast suggestions arising from prompts to GnAI with those of human experts, highlighting the ethical principles and practices that underpin professional editing to consider how their tacit and explicit knowledge may or may not translate in the use of GnAI. We then extend this discussion into a broader consideration of the pros and cons, risks and opportunities that GnAI presents for the publishing of literary fiction.

1.1 Background

Writers and editors who use applications like Microsoft Word, Grammarly and PerfectIt have long benefited from the use of AI, whether via spellchecks, predictive text, correction of punctuation, grammatical suggestions or in-built referencing systems. Now, a platform using GnAI has emerged that can create and recreate “original” content, including narrative text, realistic and “artistic” images, music, audio and video content. GnAI is driven by Large Language Models (LLMs), which are trained using “a set of human-created content and corresponding labels” [40] to generate content that is, supposedly, “similar to the human-created content and labeled with the same labels” [24]. Already, it has been announced that the aforementioned applications—in particular Microsoft Word, which is often installed by default and is now practically universal—will in future incorporate AI. In the latest version of Microsoft Word, editing has become more “invisible” than ever, as text changes are no longer obviously tracked. The materiality of the editorial process continues to diminish. Will we soon see a rise in requests for creative practitioners to “show our workings” to prove originality?

As with AI more broadly, the use of GnAI in private and public spheres has accelerated in ways few could have predicted [2, 17], from coding to copywriting, student submissions to machine marking. Governments and businesses, professional and creative communities are grappling with potential and emergent risks of its unfettered use across fields as distinct but also—for our purposes—interrelated as policy, research, publishing and the arts.

Meanwhile, the difference between content generated by AI (albeit with human prompting) and that created by humans is diminishing. In the publishing space, it has been suggested that reviewers are increasingly less likely to be able to distinguish between texts produced by AI and those written by humans [11]. There has been considerable discussion about the use of LLMs in publishing and the media as GnAI has developed a somewhat circular role: being used by publicists to develop media releases and by journalists to analyse their content [2]. In book publishing, will we see GnAI review stories it has written? (As dreaded in Italo Calvino’s If On A Winter’s Night A Traveller).

GnAI is already being used to draft copy for marketing and correspondence, to assist with brainstorming, analyse sales and marketing data, and to script metadata [44]. Writers and editors have expressed grave concerns about the ethics of using GnAI to craft creative works. Among them are the reported clandestine and unauthorised use of copyrighted works to train GnAI [35], the potential for GnAI to be used to spread misinformation [11] and the perceived threat that professional editors—along with authors—may soon be relegated to the ranks of retired professions such as phototypesetters and offset printers. In 2023, Editors Queensland, a branch of the Institute of Professional Editors (IPEd), hosted 3 AI-themed events for members [1], while IPEd’s 11th biennial conference gave considerable attention to discussion about the use of GnAI in editing practice. Indeed, the profession’s considerable disquiet about GnAI can be read in its conference title: “Futureproofing the Editing Profession” [31].

2 Methodology

To frame the shifting gaze from the utility of AI to opportunity and risk (or threat) in publishing, we draw on social theory—specifically, the precedence of social and human capital as critical resources in the production and distribution of knowledge through the publication of books and story. Given the multitude of definitions for social capital, we highlight the social economics tradition in which social networking theory (per Mark Granovetter) and social capital theory (per James S. Coleman) are combined in generating “outcomes from social interaction and cooperation” [5]: (109).

Social capital is thereby inherently beneficial in industry for its generative qualities of “reciprocity, stakeholder relationships, community engagement, trust, development and social cohesion” [5]: (109). Indeed, social capital has long been recognised as an essential ingredient in the success of key players in the publishing supply chain who contribute to the industry’s economic capital, including literary agents and acquiring or commissioning editors [58]. Ayios et al. [5] explored alternative readings that characterised social capital as, among others, neo-capitalism—primarily a propensity towards utilitarianism and defined as “a cost–benefit analysis wherein an action is morally right if it leads to the greatest good for the greatest number (the greatest happiness principle) and minimises harm” (p. 111)—or Kantianism, which identifies individuals “as forms of capital rather than being afforded individual respect as persons” (p. 112).

The editorial process represents a model for social relationships within the publishing context—to wit, “the various kinds of exchanges and unilateral transfers of control that actors engage in to achieve their interests” [14]: (300) using their individual resources, or social capital. In this model, the optimal development of a manuscript relies on the social interdependence of its key players—author, editor and publisher—as well as cooperation through other social relationships that ensure the successful functioning of systems within the publishing supply chain and the creation of value in terms of economic capital, human capital, social capital, intellectual capital and symbolic capital [58].

2.1 Method

Within the digital humanities and publishing studies disciplines, we utilised digital hermeneutics as a focused method for the study. Traditional hermeneutics is concerned with the interpretation of artefacts of words and texts via human-produced “linguistic actions and phenomena” [60], (1341). To date, much of the literature on digital hermeneutics describes the process of interpreting algorithms and big data collection, selection and arrangement, rather than examining their application, as we have done in this study. We found this approach conducive to the practice of editing because of its emphasis on the interpretation of artefacts of words and texts (in this case literary manuscripts being prepared for book publishing). While digital hermeneutics theory focuses on algorithms and big data collection, scholars note the important human input of selection and arrangement and of humans as “interpretative animals” [54], (73), steering and curating the digital data. Latour [38], for example, describes the act of “delegation” as an exchange of interpretation between machine and “interpretational agency”. In seeking to mirror the stages of editorial assessment, including structural or developmental editing, copyediting and proofing stages, we adopted a three-fold mimesis of preconfiguration, configuration and reconfiguration [52, 53]. Preconfiguration can be seen as the original, unedited manuscript (collection); configuration is the editorial perspective and initial assessment of the work (both from the human and AI perspectives (analysis); and reconfiguration is the exploitation and manipulation of the work directly resulting from the editor’s and AI’s response (interpretation). In traditional editing practice, the final stage would entail a process of negotiation between the editor and the author; our experiment, therefore, added the GnAI program to the editor’s role in this final step, with the editor acting as a co-discriminator with the AI [20], (295).

2.2 Materials

Our test case material consisted of a literary fiction short story, written by a co-author of this paper [42], at three different stages of editorial development. We conducted multiple tests with each of the three different versions of the manuscript, in order to compare the experience of working with a professionally trained human editor with the process of (self)-editing using GnAI.

We used Chat Generative Pre-Trained Transformer (ChatGPT), perhaps the best-known LLM to date. Released in late 2022 and designed to create text, ChatGPT uses natural language processing to simulate human conversations. Its list of main topics invites users to “Ask me anything” (https://openai.com/chatgpt). Just as in a conversation between humans, the user can ask ChatGPT a question, receive an answer and then follow up with further questions. In providing answers and explanations, ChatGPT “remembers” its earlier responses and even apologises when it gets things wrong. There are distinct parallels with editing by humans: we have written elsewhere on how our creative practice is a “conversation” in which professionals “prompt” authors to address problems—rather than necessarily making in-text suggestions or reductive corrections [43]. Except that “chatting” with GnAI is not actually a conversation but rather an engagement with text designed to resemble one: “Their primary goal, insofar as they have one, is to produce human-like text” [27].

We used the third iteration of the software, ChatGPT3, which we deemed the most likely version to be used because of its zero cost (the updated ChatGPT4 now includes a paid version). We began each round of testing with the same initial prompt:

Hi ChatGPT, could I please ask for your editorial suggestions on my short story, which I’d like to submit for publication in a literary journal? Here is my draft story, the title is, “The Ninch”.

This simple prompt is parallel to the initial engagement between an author and editor, whereby an author submits their manuscript to a publisher for consideration. The expectation is that the publisher/editor will bring their own “pre-loaded” process and expertise, and will not require (or desire) a highly instructional, iterative brief from the author.

3 Results

3.1 Test 1

In this first round, our test case material consisted of the original draft short story that was submitted to an established literary journal for possible publication. It was, in the end, declined by this journal, but the editor offered sophisticated insights on the narrative construction and point of view of the story. Regarding an early device using quotes about the physiological experience of drowning, the editor wrote: “[T]he quotes about drowning are heavy-handed; they’re given the job of foreshadowing suspense, creating unease in the reader, rather than the narrator doing that job. Structurally they also raise the question of who our narrator is, given we have this omniscient framing element out of nowhere.”

When considering the plausibility of the main character’s emotional response to an event, the editor reflected: “[F]or me stories are driven by choices and I’m not clear what decision our narrator, or anyone else, in the story faces.” Importantly, the editor flagged their feedback in the context of a reader, understanding that if she was missing some point of understanding then potentially readers would too, and that the job ahead was to clarify these structural issues before getting down to the level of a language edit.

We contrast this selective, individualised, high-level feedback with ChatGPT’s offering. After opening our first round of testing with the simple prompt and full copy of the short story text, ChatGPT responded with a summary of the perceived theme of the story, plus a 12-point list of suggestions:

Your story, The Ninch, explores themes of motherhood, connection to nature, and the mysteries of the ocean. It has a poetic quality and a unique narrative perspective. Here are some editorial suggestions to enhance your short story:

  1. 1.

    Clarify the setting: Begin with a bit more description of the setting. Is this a coastal town? What does the park look like? Where does the story take place? This will help readers ground themselves in the environment.

  2. 2.

    Character descriptions: Provide more physical descriptions of the characters, especially the daughter, Else. What does she look like, and what are her characteristics beyond her connection to the ocean? This will make her more relatable to the readers.

  3. 3.

    Pacing and structure: Consider breaking up long paragraphs to create a more reader-friendly pace. Shorter paragraphs can make the narrative flow more smoothly, especially when transitioning between different aspects of the story.

  4. 4.

    Dialogue: While the story doesn't seem to rely heavily on dialogue, adding some exchanges between the characters can help convey their personalities and emotions. Dialogue can also break up exposition and provide insights into their relationship.

  5. 5.

    Tighten the prose: Some sentences are quite long and complex. Simplify them where possible to maintain reader engagement. Clear and concise prose can help convey your ideas more effectively.

  6. 6.

    Character development: Delve deeper into the thoughts and emotions of the characters, particularly the mother. This will help readers connect more with their experiences and struggles.

  7. 7.

    Symbolism: Expand on the symbolism of the ocean and fishing. What do these elements represent to the characters, especially Else? Symbolism can add layers of meaning to your story.

  8. 8.

    Show vs tell: In some places, you tell the reader about the characters’ emotions or experiences. Instead, try to show these emotions through their actions, thoughts, and dialogue. This will make the story more immersive.

  9. 9.

    Foreshadowing: Consider incorporating subtle hints or foreshadowing earlier in the story to build tension and intrigue about Else’s disappearance. This will engage readers and keep them guessing.

  10. 10.

    Resolution: The story leaves readers with a mystery—what happened to Else? Consider providing some form of resolution or closure, even if it’s a subtle one. This can be satisfying for readers.

  11. 11.

    Title: Reconsider the title to make it more evocative and reflective of the story’s themes.

  12. 12.

    Read aloud: Reading the story aloud can help you identify awkward phrasing, pacing issues, and areas where the narrative might benefit from restructuring.

We were intrigued. At first blush ChatGPT did seem to pick up on elements unique to the story (character names and actions, themes, style) and offered questions worth considering. But as we weighed how we, as authors, would apply or respond to this feedback, we began to question how much was in fact specific to the text. Writing students will be familiar with the broad advice to consider showing versus telling, for example, or reading aloud to help smooth out the prose. But which sections (“in some places”) needed more showing or had too much telling? Which phrases were awkward? The suggestions to include more details about setting, character descriptions and development, as well as to explain the ending, seemed entirely at odds with the conventions of the literary short story, which are deliberately elliptical and open-ended. We began to doubt the value and authority of the feedback. Was it relevant or rote?

We then asked ChatGPT for help in revising the text, in line with its suggestions. The response was a significantly revised and shortened text that changed lyrical expressions into cliché and replaced the subtle hint of mystical underwater beings with the main character “holding up a shiny fish”.

This response would be considered a wildly inappropriate overstep if it came from a human editor, and in dereliction of expected and acceptable professional standards. It raised questions about editorial judgement as well as the nature of what readers would consider authentic authorship.

3.2 Test 2

In this second round, our test case material was the revised short story, which had been submitted to a national anthology for possible publication. The story was shortlisted for inclusion but not selected in the final cut. One of the judges, an experienced editor the author knew personally, provided further editorial feedback.

The editor’s feedback was a light touch, as would be expected when an editor works with an experienced writer on a well-developed piece, with the editor recognising and respecting the creative sensibilities of an already published author with a unique literary style.

The suggested edits, via tracked changes in the manuscript, included adding pronouns and small phrases for clarity, and alternative verbs to enhance the artistic vision of creating less-concrete certainty and more mystery for the protagonist, and consequently, for the reader. Importantly, and in contrast to ChatGPT’s suggestions (below), the editor kept the play of words in the story’s ending, which hinted at the key subtext of “otherness”. Sometimes good editing is knowing when the text is working well.

In response to the same opening prompt as in the first round, ChatGPT offered a list of editorial suggestions very similar to that provided in the first experiment with the previous version of the story; that is, general and non-specific advice on matters that could be applied to most fictional stories.

This time ChatGPT also picked up on the use of both present and past tense in the story (which it deemed to be potentially confusing) and urged consistent use of verb tenses. As professional editors, we reviewed the text and confirmed that all verbs were conjugated correctly in the relevant tense and that the transitions between the back story and front story were clear. Indeed, the skilled interweaving of back story and front story is a key technique in any fiction writer’s repertoire. ChatGPT seemed unable to assess the text at this level.

In a further prompt, we asked ChatGPT to correct any error it could spot in tense, spelling and punctuation. It came back with the advice that the 2375-word piece was too long to edit in one go, so requested sections of 300–500 words at a time. Duly provided in chunks, these revised sections then had to be knitted back together into a single document. The edits were not visibly flagged in any way, so we created a compared document in MS Word so that we might review the edits. This version looked most closely like a tracked changes version that a human editor would provide. We noted the changes in spelling from Australian English to US English (which was not requested), and the insertion of unnecessary prepositions, articles and verbs that essentially served as “padding” in what were originally stylistically lean sentences.

3.3 Test 3

In this third round, our test case material was a further revised version of the short story, by now accepted for publication in a national literary journal and later longlisted in an international writing award.

Upon acceptance of the manuscript, the journal editor (or rather, its freelance editor) made some minor edits, working on the fine detail of singular word changes to smooth sentences here and there (for example, to expand contractions such as “she’d” to “she had”). That was it. The piece was then published.

In contrast, ChatGPT’s response to the same initial prompt on this polished and, in the end, published version was again to proffer a similar list of general suggestions on common writing issues that did not pinpoint specific areas in the text where this feedback might be applied.

Crucially, ChatGPT did not pick up that this version of the story had already been published, and did not flag potential plagiarism, which is a key consideration for the human editor. Given how ChatGPT is trained—using materials scraped from the internet on the assumption that if it’s online, it’s fair game (that is, it’s in the public domain)—we wondered, is the concept of plagiarism even part of its programming?

Learning that ChatGPT copes better with a focus on singular tasks, we proceeded with separate rounds of specific requests for further editing, based on individual suggestions provided in its initial feedback. First, we asked for alternative titles for the story, then places in the text where the pace of the story could be sped up or slowed down, followed by places where there was too much imagery and not enough action.

ChatGPT’s suggested titles for the story replaced the location-specific, colloquial reference in the original title with generic phrases related to themes of family and the ocean. Interestingly, its suggestions for places to modify the pace or action did flag certain scenes and alternative approaches, but by this stage we were not convinced ChatGPT could grasp the author’s literary style, aims and positioning of the text. Its interventionist feedback was a stark contrast to the minimal changes the human editor had made to the final version prior to publication. This brings us back to the question of how useful or effective ChatGPT currently is (or isn’t) for authors or editors seeking to use it as a tool for developmental editing.

4 Discussion

Our analysis was informed by decades of professional editing experience (a century between us), as well as considerable expertise in workshopping creative writing with peers in tertiary, private sector and community-of-practice contexts. Indeed, even the experience of writing this paper as a collective reminds us of the value of beta-readers and wide-ranging editorial input, reinforcing our conviction that many members of the writing and publishing industry—particularly those on the periphery, whether as self-published or emerging authors, for example—may be inclined to submit their work to a GnAI program for editing. Why wouldn’t they? Good writers are accustomed to receiving editorial input as prompts rather than directives; to identify potential problems (often framed hypothetically as “the reader might interpret this as”), rather than as a checklist of corrections that must be taken in.

We see in this experiment a contrast between machine editing and the use of expert human editors who provide prompts and specific, practical suggestions at the key stages of revision prior to publication of this literary short story. The larger research project, of which this test case is but a part, seeks to examine the possibilities—and possible consequences—of GnAI for editors of long-form literary fiction. Specifically, whether GnAI can be harnessed to increase productivity or “is capable of being a co-pilot on complex tasks” [15]. Among other key findings, the test case highlights ethical issues that invite further discussion and dissemination:

  • the genesis of ChatGPT’s training with pirated copies of works in copyright

  • the consequence of this in terms of ChatGPT’s inherent bias towards producing similar works

  • the depth (or lack thereof) of “editorial” intervention recommended by ChatGPT with regard to the literary fiction short story

  • underlying assumptions that resulted in actual or perceived bias in ChatGPT’s responses.

This analysis sets the foundation for further research and open-ended discussion of ChatGPT’s utility, popularity and potential application in and for established (editorial) creative practice.

4.1 The role of the publishing editor

The task of editing a manuscript for publication may fall within the purview of a range of roles, professional, non-professional, representing formal and informal responsibilities of the in-house and freelance editor. In this study we focused on the role of the professional editor, who may be the employee of a publishing company or a freelance consultant, and who adheres to a professional code of ethics and conduct such as those required of members of IPEd [34]. In contrast, GnAI can be used—indeed, is being used—by authors to seek editorial advice that does not adhere to or aspire to industry best practice. Importantly, ChatGPT did not seek to verify the originality and authenticity of the author’s work. In this sense, GnAI failed to contribute to the development of capital—intellectually and symbolically—in the publishing process.

In large publishing houses, editorial tasks are usually split between different job titles: publisher, acquisitions editor, commissioning editor, managing editor, editor (senior editor, line editor, copyeditor, desk editor etc.), production editor, permissions editor, project editor, marketing editor, proofreader, editorial assistant, and so on [41]. In small and medium publishing houses, editors may take on management of the entire publication project, from concept to promotion of the published book. These editors may seek to collaborate, or may be responsible for liaising with, a range of external suppliers who contribute to the editorial process and the development of the manuscript, such as other editors, production personnel, graphic designers, illustrators, photographers, typesetters or desktop publishers, and marketing personnel [19] whose roles are similarly reliant on their social interdependence to achieve their mutual interest in the publication of a manuscript. These supply chains continue to exist even given digital disintermediation, persisting in particular in small presses where publishers may also undertake the press’s editorial functions, or editors at large may work as de facto publishers. In this way, professional editors utilise their human capital via individual resources (intellect, knowledge and expertise) to enhance the economic, social and symbolic capital inherent in the traditional publication of literary fiction.

IPEd’s Australian Standards for Editing Practice [33], “sets out the core standards that editors should meet and tells employers what to expect from the editors they hire” (p. vi). Fundamental in the Standards are five core, skills-based domains: professional practice, management and liaison, substance and structure, language and illustrations, and completeness and consistency. Whether these skills and related tasks are considered notional or discretionary, or are defined in the editor’s position description and/or contract, a human editor may reasonably be expected to undertake some or all the tasks listed in the Standards, as determined by the scope of the given project. While the role of the human editor may vary according to the needs of the project, the parameters and the goal of editing are imbued with expectations about the standard of work that is to be undertaken. In comparing this with ChatGPT’s response in our case study, we consider three key points.

  1. 1.

    LLMs such as ChatGPT are, at best, considered adjuncts to human editors—tools to enhance productivity rather than creativity [8].

  2. 2.

    GnAI is a real and present threat to the livelihoods of creators and editors alike [7, 29], even if it does not, as yet, have a clearly defined role, goal or value in the publishing process, and no risk mitigations are in place. Examples of risks inherent in use of GnAI for editing include the lack of critical oversight outside of the specified brief (the prompt), and non-editorial staff using the technology for development editing without training in industry best practice and ethical standards.

  3. 3.

    The engagement of writers with GnAI cannot, at this point in time, be considered a “relationship” in the same way as that of the author–editor relationship. The human editor’s relationships with the author/creator and their employer or client are pivotal to their effectiveness in editing the manuscript, as is trust in their execution of their role [36].

Professional editors contribute significantly to the economic, social and symbolic capital and viability of publishing. In the practical world of professional editing, the role of the editor is most frequently characterised as one of mediation between opposing forces, whether these are economic (e.g. obtaining the best cover design at the cheapest price), informational (e.g. mediating between comprehensiveness and succinctness) or managerial (as the late, esteemed Australian editor Janet Mackenzie [41] wrote, “the editor must be careful not to appear to take over—even if that is, in fact, what you are doing” (p. 51)). However, this materialist perspective belies the fundamental value of the editor in relationship with an author: that of collaborator, teacher and confidante, and which often goes beyond the development of a single manuscript. That little has been published about this aspect of the author–editor relationship and its possibilities is perhaps due to the “romantic myth of the [writer as] solitary genius” [10], but the capital value of the professional editor is surely undisputed [58].

4.2 Editing and ethics

In professional editing, questions about ethics typically arise in respect to originality and merit in a text (intellectual and symbolic capital), and maintenance of professional boundaries (social capital). We explore each in turn.

4.2.1 Copyright infringement

While publishers require authors to verify that a manuscript represents their own original work, editors nonetheless are taught to be alert to signs of plagiarism and other forms of copyright infringement [48]. Academic editors working on texts being prepared for assessment must ensure they do not overstep the bounds of the editorial role, which are clearly articulated in their professional guidelines for the editing of research theses [32]. The editor of a work of fiction is driven by a different, though not wholly unrelated, imperative. Publishers and readers of fiction are continuously seeking new works that inform, entertain and delight, that are “evocative, intriguing, original” [15]. These are the qualities that underpin the criteria for literary merit (social and symbolic capital) that editors of fiction are seeking to assess in the manuscripts they are assigned.

At no point in our test case did ChatGPT seek to verify the authorship of the short story. Even more disturbing, it also did not identify that the story had already been published in a well-known and current literary journal (Meanjin). (The first paragraph of the story is available online, the rest behind a paywall.) This suggests that the LLM’s training had not included any kind of deductive processes that would resemble the moral and ethical deliberations of the human editor.

According to the ASA [4], ‘It is undisputed that the works in the training datasets have been copied without permission from, or payment to, creators’ (emphasis in the original). As at June 2024, it is unknown whether the courts in the United States or elsewhere will sanction GnAI owners for stealing the original content of creators and copyright owners [3], or prevent them from doing so in future. If a professional editor had similarly overstepped the bounds of the social contract and/or the obligations of their employment, there would be severe consequences, the least of which would be not to continue to profit from the underlying infringement.

The capital values developed in the publishing process are predicated on a warrant by the author that the work they are supplying for editing is their own original creation. This could be seen as similar to the suspension of disbelief when a reader engages with a work of fiction, or the fourth wall in a theatre production, and in the process of editing the editor becomes a conduit to such engagement. Now that it is possible for GnAI to become a party in the creative process, questions arise about the nature of authentic creativity when using GnAI and, ultimately, the ownership of the intellectual and other capital that is generated. Further considerations include the fidelity of the author–editor relationship and its contribution to the development of the author as a writer.

4.2.2 Professional boundaries

Manuscript editing occurs on two dimensions: first, with the editor working directly on the manuscript (commonly in electronic form, typically using the industry standard software, Microsoft Word and historically on hard copy—a practice some experts continue to use so that the author is prompted to consider each textual suggestion rather than defaulting to the apparently mindless act of “Accept All Edits”), in “relationship” with the text and, second, simultaneously in relationship with its author and publisher (the client or employer). The editor’s approach—and indeed, their editorial suggestions—is informed by the nature of these social relationships as well as their professional and legal or contractual obligations to the client (which can be the publisher or the author), their moral obligations to the reader and their adherence to a professional code of ethical conduct, and even in their initial assessment of a manuscript’s potential.

Aside from contributing to its economic capital, in developing a manuscript for publication, the editor is simultaneously perpetuating the role of the book as a cultural object and acting as a cultural mediator [39], by seeking to “mak[e] each book the best conceivable version of itself,” and is in intimate relationship with someone else’s creation in a manner that presents a “paradox of simultaneously holding and letting go” [10]. We editors are invested in the story’s success, emotionally, professionally, financially. Needless to say, an LLM has no such investment, and this was evident in its responses in our test case with ChatGPT [16].

As noted, the human editor is usually employed or contracted to undertake specific tasks in the development of a given manuscript, subject to professional standards and codes of ethics and conduct. These tasks may include a sensitivity read, a structural edit or a copyedit, and yet the editor is also likely to have a weather eye for issues beyond the stated remit in the brief (including ethical issues). ChatGPT has no such obligation.

4.2.3 Hermeneucity

When assigned to work on a manuscript destined for submission to a literary journal, as in our case study, the professional (human) editor is expected to have a sound understanding of what that entails, including industry definitions as to what constitutes a “literary” work, and the conventions of natural language—not to mention a mutual understanding of what a particular publishing house might consider “quality”, or a market might rate as “value”. In our experiment, there was no indication that ChatGPT “understood” the parameters of style, idiom and imagination that writers and readers alike expect of literary fiction. ChatGPT turned a speculative subtext hinting at a “southern selkie” into a literal shiny fish [16]. This lack of literary sensitivity and insight significantly undermined the capital value of this creative endeavour.

A further weakness identified in LLMs such as ChatGPT is polysemy [59]—not that it did not identify that a word or phrase can be taken to have more than one meaning but that it did not detect the intended meaning, not from the context of the inquiry (“Hi ChatGPT, could I please ask for your editorial suggestions on my short story, which I’d like to submit for publication in a literary journal?”) nor from the text itself. It is early days for this technology, and future updates may well remedy this problem by prompting the user to state the intended genre or by asking questions about the user’s intended meaning or creative intention, but it’s clear that ChatGPT has no obligation to improve the text or to enhance its social, cultural or capital value. Its “brief”, such as it is, is simply to respond: literally, efficiently and predictably to a prompt.

Our study is not the first to encounter a lack of hermeneucity in LLMs. In their study of hermeneutic value in prompt engineering, Henrickson and Meroño-Peñuela [25] found that ChatGPT produced responses that ranged from trite to pastiche, in effect, “bullshit: text that sounds persuasive, but has little regard for truth or relevance”. In contrast, by convention any work a human editor undertakes on a manuscript must seek to improve it. That is what editing is understood to be: enhancing, clarifying and correcting rather than changing text for the sake of it. The extent or outcome of such improvement may be debatable, but no author or publisher would expect a manuscript to end up in a worse state than when it went under the editorial scalpel, or, indeed, accept the manuscript for publication, or commission the editor for future work. Undue or unjustified heavy-handedness on the part of the human editor would likely be considered a breach of the social contract that binds editor and publisher. A real risk for writers who use GnAI for editing is that it is prompted by the author, who doesn’t have the critical distance of an editor or the inside knowledge of the publishing industry. Thus, ChatGPT offers no better authority on how to edit than any (human) author.

4.2.4 Editorial bias

Human editors all come with their own particular influences and biases. Some biases are observed as positive, such as favouritism for a particular style; others are negative and discriminatory [26]. Apart from being unprofessional, in Australia as in elsewhere, it is illegal to discriminate against someone because of their ethnicity, religion, age, sex or disability. While editors who subscribe to professional ethics must act according to those codes of conduct, bias and discrimination can come in many forms, some not-so-obvious ones include systemic and hidden biases such as unconscious biases [6]. Much of the bias detected in manuscripts may be readily apparent, but implicit bias is not always easy to recognise, particularly if the author and editor share cultural or socioeconomic similarities. Implicit bias can include the ways in which certain groups are excluded, marginalised or perceived, and the ways in which others are given prominence “where experiences, thoughts, opinions and biases of one group are over-represented in the social reality” [49]. Author bias must be identified and acknowledged, their manuscript viewed through a variety of lenses that can offer diverse perspectives and a “critical reading of the underlying themes” [47]. Overt discrimination may be abhorrent, but other biases, sometimes even positive biases, can result from authors seeking to respond to readers’ expectations of a particular genre or style, perceived as “giving the readers what they want” [44].

Human editors can be trained to be aware of their hidden biases, and they can be held to account through their employment and engagement in the profession, as well as through professional supervision, tertiary education and other forms of peer support [44]. The question of whether AI, including applications such as ChatGPT, can be trained in ethics and ethical conduct has been considered in fields as diverse as health care, education, cognition and non-academic settings (see e.g. [28, 30, 50]). A consistent finding is that, while AI can be a valuable resource and its use can bring significant capital and other benefits to society, trust and reliability are contingent upon human judgement, particularly in settings requiring critical decision-making. Further, “there is a risk of overdependence on AI, which may diminish the role of healthcare professionals, thereby reducing critical thinking and human interaction” [30]. Perhaps editors, and in time authors, will be trained to use this tool ethically, even if its very application is contrary to the social relationships, economic and cultural value upon which the knowledge economy and cultural industries are founded. But will GnAI be “trained” to share the same goals as human editors? Goals that extend beyond the technology’s desire “to provide a normal-seeming response to a prompt” [27]: to impart information, expertise and editorial advice?

It might be assumed that AI is, by definition, agnostic to the emotions that can drive bias and discrimination in humans. However, LLMs, including ChatGPT, are not without undue influence or bias. Indeed, as Cave and Dihal [12] found, “machines can be racialised and […] this racialisation includes transfer of the attendant biases found in the human world” (p. 699); for example, resulting in representational harms that portray the world as homogenously white and middle class. Other studies have reported undue influence through word embeddings and algorithms in older and newer versions of LLMs, including gender bias [49], racial bias [21] and political bias [55]. A study by Rutinowski et al. [56], which applied personality tests to gauge “self-perceived personality traits” in ChatGPT, returned unexpected results. Using a range of well-known tests, including OCEAN (Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism), Myers–Briggs Type Indicator (MBTI), Dark Factor and a political compass questionnaire, the researchers found that ChatGPT was slightly biased in favour of libertarian and progressive views, which complemented its “perceived personality” as open and agreeable. ChatGPT did not achieve a high score in the Dark Factor test but scored relatively high in the sadism and ego traits [56]. Such studies (see also [9]) refute the assumption of neutrality or agnosticism in GnAI.

5 What this study adds

In less than a hundred years, digital communication has become ubiquitous across societies and economies worldwide. Ever since the launch of the first programmable digital computers in the 1940s, scientists have sought to make intelligent machines. From an ethical perspective, whether this project succeeds matters less than how it is conducted. What is gained must be weighed against what is lost, and at what cost. Relationships, principles and practices, as well as poor handiwork and creativity, and diminished social and other capital. When developers ran roughshod over the intellectual property of creators, many considered this a failure and a betrayal that must be addressed. Indeed, the writers whose published works have been plagiarised to populate and train ChatGPT may turn from the very idea of using GnAI in the editorial development of their next work. Trust is fundamental in the author–editor relationship. A related but often overlooked factor is the social capital inherent in creative collaboration between author and editor in the literary space—editors are prized for their ability to accompany an author on their creative journey, a skill that goes well beyond ensuring consistent grammar and punctuation. At its heart, editing is a role that demands the ability to engage in an active and continuing conversation that includes judicious prompting where needed, and also knowing when only the lightest touch—and possibly some diplomatic reassurance—is necessary.

Amid the fanfare that accompanied the release of ChatGPT in late 2022 there was a sense of disquiet in some quarters. Would human creators, such as writers and editors, soon be replaced by AI? The Writers Guild of America galvanised members in an unprecedented and effective strike. Others roundly dismissed the notion (e.g. [8, 51]). Our study sought to explore the potential impact of GnAI, in its current guise, when it comes to editing literary fiction. Is it a concern of ethics, or industrial relations? Of course, it is both, but the former has not previously—or comprehensively—been addressed by our profession or industry.

This study has highlighted how the editorial process represents a model for social relationships and the development of social, economic, intellectual and symbolic capital within the publishing context. We have demonstrated how the addition of GnAI has complicated and compromised interdependent social roles, interrupting the essential conversations that take place between author and editor in the development of literary works, thereby diminishing the social and capital value of the publishing endeavour. The study demonstrates that there are pressing questions for further investigation.

We have explored critical aspects of AI’s potential impact and influence on traditional editing roles and responsibilities, extrapolating likely risks from a detailed discussion of a unique test case we conducted at the start of 2024. We examined the potential benefits and risks—with a particular focus on the ethical implications—for authors, editors, publishers, the knowledge economy and the creative landscape, asking how the increasing application of GnAI might jeopardise the roles of professional editors and the essential editorial practices we rely upon in developing literary fiction manuscripts for publication. Working from an experiment comparing three rounds of edits by established literary editors at esteemed journals with ChatGPT’s suggestions, we conclude that GnAI applications appear to have been created within a development paradigm that is, at best, agnostic to notions of social and cultural capital and, at worst, at odds with the ethics and conventions that underpin the delicate balance of commercial, cultural and individual interests that enable book publishing.

Given its demonstrated limitations, we would not expect GnAI to have a clearly defined role in the editing process any time soon. “Editing” by ChatGPT (or other LLM) is not a requirement prior to submission of a manuscript for publication. And yet, in just over 18 months since the release of ChatGPT, writers worldwide are looking to GnAI as a shortcut, or value-add, in the development of their manuscripts. Certainly, it is a less-expensive option than paying a human editor. Professional editors are right to be concerned—not that GnAI might do a better job than a human editor, but that many may not care about that distinction. Ultimately, ethically, the buck is deemed to stop with the human who chooses to engage with GnAI.

6 Conclusion

While ChatGPT’s overall performance in our study did not come close to that of any of the three human editors we compared its work to, it should not be dismissed outright. (We also note that our participants each had their own distinct styles of communication: preferences, inclinations, specialisations and enjoying personal, sometimes established, relationships with the author of the story at the centre of our test.) We found, as the industry has more broadly, ChatGPT could play a useful role in supporting the work of human creators—writers and editors—by taking on the “grunt work”, such as correcting spelling, recommending punctuation, ensuring consistency in tense, and other copyediting tasks, so long as the parameters were clearly and narrowly defined [16]. We also recognise the rich collaborative potential for writers and editors in the early brainstorming stages of developing story ideas, when GnAI can provide suggestions to help clarify their thinking on rough drafts prior to in-depth work with a professional editor. But it is no substitute for the “real thing”, even if the only audience writers are writing for are fans, friends and family. When authors seek feedback and editorial input, they are usually not asking for solutions per se, but rather help to identify problems they may not be able to see—perhaps because they are too close to the subject, or they have read the draft too many times and cannot see the wood for the trees. Sometimes the best input from an editor is simply to say, “Look, here something is not—not yet—quite right.” The very nature of the LLM, based on predicting patterns rather than deducing or inducing insight, is completely counter to best practice in editing. The fact that GnAI does not represent a publisher undermines its value. Perhaps another form of AI will emerge and these forerunners will be forgotten, along with BETA cassettes. Certainly, GnAI currently cannot edit—not as we know it. It is NQR (not quite right) yet, but can be a companion for certain types of creative writing. By uncovering GnAI’s limitations in this space, our experiment has detailed the potential benefits and risks, and highlighted the inestimable—and often unappreciated—value of the human literary editor and their contribution to social, cultural, symbolic and economic capital.