AI Literature: Will ChatGPT Be the Author of Your Next Favourite Novel?

In Karl Schroeder’s science fiction novel Lady of Mazes from 2005, an artificial intelligence (AI) floods humanity with millions upon millions of exceptionally well-written stories. The number is so vast that they simply could not be read in a single lifetime – an inexhaustible contingent of stories, all precisely tailored to humanity’s particular preferences. Since the publication of ChatGPT in the November of 2022, this scenario – in which an artificial intelligence offers a continual supply of hyper-individualised pieces of literature – is no longer inconceivable. Nor does it seem in any way an unrealistic expectation of a near future: the proliferation of large language models has already exponentially accelerated the production of written content authored by machines, including literary texts, especially in the non-fiction, sci-fi, crime, or children’s and young adult genres.

One sector of the market particularly affected by AI publications is that of self-publishing platforms, such as Kindle Direct Publishing. Both publishers and writers are confronted with a market in which the stranglehold of the attention economy rapidly continues to tighten. Even without literary production augmented by AI, the industry has long complained of a decline in readership. Moreover, publishing houses are no longer solely in competition with each other, but increasingly challenged by novel forms of literary mediation such as BookTok or digital publishing platforms, which have changed the way literature is marketed and create low-threshold offers for independent authors.

“Language models seem to question our idea of writing as a distinctively human activity.”

In the age of Amazon and Goodreads, in which readers are encouraged to comment on and rate their consumption habits, literary production tends to be heavily affected by current demands. Algorithmic data analysis has always been the petrol which drives platform capitalism, and AI technologies will continually improve and refine this evaluation of data. Though data analysis via adaptive algorithms is now standard practice in the literary industry, the use of generative AI – capable of producing natural language in order to change the how books are written – is relatively new. But the use of AI tools in the writing process is moving progressively into the realms of a digital mainstream, as can be seen in the integration of AI in word processing programs. AI-supported writing programs promise a low-threshold participation in literary text production and the optimisation of manuscripts.

The way in which these tools are used in the creative writing process of human authors is currently the subject of heated debate. It makes international headlines whenever prominent authors express their willingness to make use of AI in their respective creative processes, as can be seen in the case of Japanese author Rie Kudan, who used ChatGPT for her novel Tokyo-to Dojo-to, for which she won the Japanese Akutagawa Prize for Young Writers in 2023. Language models seem to question our idea of writing as a distinctively human activity: where lies humanities distinctiveness if ChatGPT and LLaMA are able to reach the same literary heights?

Predictable Storytelling

Demystifying large language models (LLMs) and revealing their inherent deficiencies requires an understanding of their technical background and functionalities. LLMs are text-producing systems which operate on the basis of vast data-corpora by means of elaborate probability calculations. ChatGPT possesses neither general knowledge nor embodied knowledge. Rather, it is a system which can recognise patterns in language and apply these proficiently to synthesise new texts. Language models can thus also be used for literary writing and help in the creation of stories and novels. However, the often-undisclosed data sets used to train the LLMs tend to rely on the illegal use of copyrighted material. An increasing number of lawsuits have been filed in this very regard; the New York Times, for example, filed a lawsuit against Open AI in 2023.

“For the most part, LLMs currently generate schematic stories that reproduce clichés and stale phrases, always tainted with unique biases.”

In recent years, these systems have made significant improvements in terms of storytelling, in part due to the enlargement of context windows. While GPT-3, the predecessor to ChatGPT, could barely produce cohesive stories, ChatGPT can now create a coherent short story or scene from a single prompt. The quality of the results remains debatable, and the more autonomously the systems are allowed to work – i.e., the shorter the prompt – the more generic the literary texts. For the most part, LLMs currently generate schematic stories that reproduce clichés and stale phrases, always tainted with unique biases. This is partly attributable to the data set in use, but also to the difficulty in clearly defining what makes a compelling story. Categorising artistic and literary production around a fixed median is itself problematic, just as mediocrity is no guarantee of successfully telling a story.

Descriptive possibilities are offered by concepts from the fields of poetics and narratology, which concern the organisation of plot, characters, space and time in narratives. Conventionalised plot structures, such as Campbell’s hero’s journey, appear time and time again as common narrative patterns in text created by LLMs. Formulaic storytelling does not necessarily equal bad storytelling – the hero’s journey has proven to be a broadly effective narrative template in the global West, a blockbuster guarantee since the international success of the Star Wars franchise.

Resorting to conventional plot models can be useful in writing processes, as narrative schemes provide the fundaments to many genres. In his diagnosis of the status of books in the age of Amazon, the literary critic Mark McGurl points out that the need to read is often articulated in a desire for repetition; children actively demand to have their favourite story read to them, even if they already know it by heart, and masses of fans eagerly await the next installation of Marvel’s multiverse – sure to uncannily mimic its predecessors in every conceivable way. Templates are omnipresent in our cultural production. Generative AI excels at repeating them. Does this imply that AI writes the most captivating stories?

In experiments on literary texts that we are currently conducting at the Hildesheim Institute of Literature with the help of LLMs, we predominantly obtain results that meet all the necessary criteria for supposedly captivating stories, but due to their script-like realisation, feel devoid of what they should transport.

Collaboration Over Competition

Instead of hypothesising on the collapse of the culture industry due to AI or glorifying the usage of LLMs as a guarantee for bestsellers, we should look for answers in the spaces in-between, rather than thinking in binary terms. The main potential in regard to literary production lies above all in collaboration with AI. In this context, LLMs function rather as tools or aids. Or, depending on the perspective, even as co-authors.

“Instead of hypothesising on the collapse of the culture industry due to AI or glorifying the usage of LLMs as a guarantee for bestsellers, we should look for answers in the spaces in-between.”

If we view the writing process as an abstract model, three phases can be distinguished: conception, text generation and revision. AI is currently most suitable for creative processes in the conception phase. Ideas can be discussed in the chat – suggestions are posed and the system reacts to them. Lists, titles, settings and character biographies can easily be generated, the course of a story planned and fine-tuned. All of these elements can have a positive effect on literary processes. The systems currently perform weakest in the second phase of text generation. But here, too, any evaluation must account for different approaches to genre as well as individual taste.

In Lady of Mazes, the sudden abundance of exceptionally good, AI-authored literature has no effect on the creative ambitions of Earth’s population. In Karl Schroeder’s novel, people continue to write anyway, because creativity – according to Schroeder – is what defines people and makes them human in the first place. Given the discursive climate towards generative AI, in which the role of humans as a central element in literary and artistic processes is repeatedly emphasised, it seems clear that Schroeder’s fictional scenario is mirrored in our contemporary world. The greatest potential does not lie in entirely outsourcing literary work, but rather in viewing it as a supplement to the creative writing process – a hybrid, rather than a whole.

Learn more in this book chapter

Jenifer BeckerKönnen Chatbots Romane schreiben? Der Einfluss von KI auf kreatives Schreiben und ErzählenFrom the Open Access volume KI:Text (Gerhard Schreiber & Lukas Ohly, eds.)

You might also be interested in the episode “Kann KI auch Literatur?” of our podcast “Wagnis Wissen” (available in German on Spotify, Apple Podcasts, Youtube and wherever you listen to podcasts):

[Title image by Ilya Lukichev/iStock/Getty Images Plus]