The Prose Code: A Journey into the AI-lluminated World of Literary Algorithms

The earliest instance of computer-generated literature is considered to be Jean Baudot’s 1964 A Writing Machine (Heflin 2), and artistic creation has continued apace with technological advances. Far from being an outlier, the connection between art and AI represents a creative form intricately woven into the history of technologically influenced arts, a trend that gained momentum after the digital revolution (Santaella 45). These range from plot and short story generation, poetry, and long-form writing. Allison Parrish’s work is a shining example of computational poetry. She sees herself as a descendent of the Dadaists and employs bots to navigate unexplored regions of semantic space. In one of her projects, she used the Google Books corpus to find ‘voids’, places where two words had never been used together before. Two examples of her findings were ‘angiography adequate’ and ‘abreast annihilates’ (in Miller 220). 

When it comes to prose literature, AI-generated fiction can be found both in the commercial and literary spheres. Philip M. Parker is a businessman with software capable of scouring the web, extracting information on almost any subject, and transforming it into print-on-demand books. In 2008, he asserted that he was customising his algorithms to generate romantic fiction (Miller 192). Less mercenarily, in 2018, Ross Goodwin, former speechwriter for President Obama, published 1 the Road, an AI rendition of a road trip reminiscent of Jack Kerouac’s On the Road. Throughout the journey, the AI generated words in response to its surroundings and the places it traversed, crafting somewhat disjointed yet evocative sentences, such as, ‘It was nine seventeen in the morning and the house was heavy’ (quoted in Miller 230).

Two patterns emerge from this brief overview: either these are tightly-controlled narratives in forms that already lend themselves to predictability such as romance novels, or they are thought-provoking forays into meaninglessness. Both are a far cry from AI cognitive and computer scientist, Marvin Minsky’s, 1982 prediction: that AI would soon possess the capability to generate prose at the level of Shakespeare (Amabile 351). Indeed, in 2018 neural model Deep-speare rose to media attention through its brainteaser ‘AI or not AI: that is the question’. Yet, a Shakespeare scholar was quickly able to discern which was which. This was partially due to the errors in Elizabethan prose but also its limited readability and emotional impact (van Heerden and Bas 177). The issue with older, template-form AI was its restricted scope, limited capacity for creative generalisation, and labour-intensive human engineering (Chun and Elkins 106). New probabilistic models, in their acquisition of language through trial and error and studying vast transcripts of human-generated language, do not have these limitations. Instead, they exhibit factual errors resulting from a lack of real-world knowledge, manufactured facts, illogical thinking, logical leaps or omissions, repetition, and grammatical problems (Chun and Elkins 107). The longer the narrative, the more these mistakes become apparent. 

As AIs currently have no awareness and lack any form of intent, there will always be a human in the loop, even if that involvement is limited to pressing start. Photography is a useful analogy here. In the 19th century, the definition of visual art was revolutionised by a technology that could capture the world at the click of a button. It would be quite preposterous, however, to state that therefore the camera is the artist. The intent and conceptualisation remained in the hands of the photographer, as nowadays the use of AI remains in the hands of the computational artist. Poet Allison Parrish is firm in her stance regarding AI authorship: ‘When I put out a book of poems it’s by Allison Parrish, not Allison Parrish and a poetry bot […,] in the same way that a Jackson Pollock painting is not by Jackson Pollock and a paint can’ (quoted in Miller 218). She continues: ‘furthermore, it will always be a mistake to attribute volition to the computer and not to the people who programmed it because attribution of volition is removing personal responsibility: the algorithm did it, not me’ (quoted in Miller 223). Thus, AI is an artistic tool that, as we have seen in the examples above, can allow artists to explore exciting and novel conceptual spaces. Like any tool, however, its use must be tempered by a consideration of its ethical implications, and in AI these are, unfortunately, myriad.

Stanford University’s Senior Fellow at the Institute for Human-Centered Artificial Intelligence, Michele Elam, pinpointed a key issue in the use of literary algorithms: the narrowing of narrative possibilities. The content and style of AI storytelling are increasingly leading to the culling of narrative forms that do not support the ‘addictive’ storytelling that corporations are pushing at scale (293). Thus, creative AI could restrict, not broaden, the narratives available for public consumption. Her analysis of a ChatGPT adaptation of Maya Angelou’s poem ‘Still I Rise’ yields further insights. She demonstrates how it fails to differentiate between the last two centuries’ Black vernaculars and expressive forms, flattening their historical and cultural specificity. With this undifferentiated language there is a concurrent loss of linguistic richness and meaning. This is especially problematic considering the way a poem about racial empowerment has been folded into a ‘jumble of blues, Black power, racial uplift, and Ole Man River minstrel,’ given the very different political connotations of these forms (287). There is a clear misuse of cultural forms through the operation of a machine emanating from wealthy, usually white, businesspeople and software developers that works to neutralise an expression of Black political power.

The power dynamics encoded in Elam’s specific example of the use of ChatGPT are echoed across the AI industry. The Global North and its gigantic data companies exert significant influence over cultural development globally and run the danger of widening the North-South digital divide (Santanella 51). Elam further points out that, although AI texts present themselves as impartial, truthful accounts of reality, helped by the lack of transparency of how their databases are created and used, they in fact come with a set worldview extracted from their database (292).  Since creative works are products of the culture in which they are produced, data derived from them will inherently display institutional and cultural biases. Director of the Communication, Media and Cultural Studies at the University of Canberra, Jenn Webb, writes: ‘We live immersed in representation: it is how we understand our environments and each other. It is also how we both are, and how we understand ourselves; representation is implicated in the process of me becoming me’ (3, italics in original). Implied in this remark is how representation influences the construction of identities, cultures, and communities as well as in the expression of difference. Rudine Sims Bishop’s famous 1990 article on representation in children’s literature concludes that children receive a potent message about how they are devalued in society when they have a complete absence of or negative representation in the books they read. This remains true outside of childhood, with a 2021 ViacomCBS study finding that 59% of people poorly represented on screen feel ‘unimportant, ignored, or disappointed’ and that this deeply influences their sense of self-worth and confidence (41%), sense of belonging (40%), and prospects in life (34%) (Guerrier). 

The significance of representation in media thus means that biases in the products of creative AI cannot be ignored. Indeed, a central recommendation of the UNESCO’s 2022 MONDIACULT Declaration regarding AI ethics is that of cultural diversity, stating that ‘generative AI models should be trained and corrected to avoid perpetuating biases and stereotypes and ensure that diverse and representative stories or images are produced and disseminated’ (quoted in Bailleul). In generative AI models, biases can be found in two places: in the raw datasets, which include previously created works of art, and in the design stage, which may represent institutional prejudices (Flick and Worrall 76). For example, let’s say that Philip M. Parker has refined his software and is able to produce romance fiction through AI. The programme is then tasked with writing a commercial bestseller. As popularity is the basis for the criteria, it will seek to create fiction for the ‘average’ romance reader. A survey in 2021 found that 92.2% of romance novels were written by white authors, 73% of romance readers were white, 45% had a college degree, 82% were women and the average reader age was 42 (Curcic). Thus, the algorithm is likely to produce a white, middle-class boy-meets-girl narrative, and unlikely to write a boy-meets-boy, a Black-girl-meets-Japanese-boy or a novel with well-rounded working-class protagonists. Lack of representation due to the commercial incentives of AI literature is not the only problem. AI has been known to produce racist, sexist, and other hate speech, such as with Microsoft’s TayBot (Flick and Worrall 81). A key recommendation of the ACM Code of Ethics is relevant here: ‘a system for which future risks cannot be reliably predicted requires frequent reassessment of risk as the system evolves in use, or it should not be deployed’ (Gotterbarn et al.).

AI developers, however, appear resistant to confronting the problem of bias in their algorithms. In December 2020, co-leader of Google’s ethical AI team, Timnit Gebru, was ousted after refusing to remove her and her co-worker’s names from a paper surveying the pitfalls of generative language models (Harris). Gebru went on to establish DAIR in 2021, an independent, community-driven organisation which aims to limit the unchecked power of large technology companies over AI research, development, and usage. According to Gebru, marginalised communities likely to be damaged by algorithms need to have a voice in the creation of AI for it to serve the interests of all members of society (Qumer 1). This case underlines the need for decentralised, diverse AI research and why investors should back these kinds of initiatives.

Even with a decentralisation of AI research, the problem of how to tackle biased content in databases remains. A considerable problem is the obscurity of such models, given the way generative AI learns from its data and creates its own rules and associations (Edmonds 63). It is not at present possible to know why the AI produces the text it does. Finding biases in datasets will be made easier once the ability to observe how an algorithm handles and evaluates data is developed. It will not, however, address the bias in the dataset per se. Attempts at debiasing may involve eliminating prejudiced information from texts, such as gender stereotypes (Flick and Worrall 77). However, this will not alter the reality that, for example, romance novels predominantly feature white, middle-class, cis, straight characters. Furthermore, it might not be sufficient to just add more diverse data to correct the imbalance (Flick and Worrall 77). Although debiasing databases and adding content may improve the outcomes, further research is needed before generative AI can be considered safe to use. 

Computer-assisted writing can help artists attain creative breakthroughs and allows for a range of artistically enhancing projects such as those by Allison Parrish and Ross Goodwin. Yet the current ethical issues in its deployment are pernicious, wide-ranging and complex. It could even be questioned whether developers should be looking to tackle the ethical questions embedded in AI’s creation before releasing it to the public. However, is this possible when research in this area is supported by the corporations that AI serves and ultimately financially benefits? In this rapidly evolving field, governments must frequently update their legislation to tackle the negative effects of AI. At the same time, diverse and global initiatives such as DAIR should be valued as a step towards achieving equity in literary AIs.  

In ChatGPT’s story, Whiskers and Luna are drawn to an enchanted garden on the edge of town. They are separated, Whiskers playing hide-and-seek with the mischievous fairies while Luna converses with a wise old owl. Their paths converge once more in the heart of the garden and there they share their stories, ‘weaving a tapestry of enchanting tales that would be whispered among the town’s cats for generations to come.’ The moral of the story is that fun and games, which is present in most uses of literary AI, must be combined with wisdom. As readers and writers in a field where computation is increasingly used as a creative tool we would do well to heed this message. 

References

Amabile, Teresa M. ‘Creativity, Artificial Intelligence, and a World of Surprises.’ Academy of Management Discoveries, vol. 6, no. 3, Sept. 2020, pp. 351–54

Bailleul, Cyril. ‘UNESCO unites diverse perspectives to inform policies for AI in the creative sectors.’ UNESCO, November 2023. Accessed January 13, 2024 at: www.unesco.org/en/articles/unesco-unites-diverse-perspectives-inform-policies-ai-creative-sectors. 

Chun, Jon, and Elkins, Katherine. ‘What the Rise of AI Means for Narrative Studies: A Response to “Why Computers Will Never Read (or Write) Literature” by Angus Fletcher.’ Narrative, vol. 30, no. 1, 2022, pp. 104–13.

Curcic, Dimitrije. ‘Romance Novel Sales Statistics.’ WordsRated, October 2022. Accessed January 16, 2024 at: wordsrated.com/romance-novel-sales-statistics/

Edmonds, Ernest. ‘AI, Creativity, and Art’. Edited by Craig Vear and Fabrizio Poltronieri. The Language of Creative AI. Springer, 2022, pp. 56-72

Elam, Michele. “Poetry Will Not Optimize; or, What Is Literature to AI?” American Literature, vol. 95, no. 2, 2023, pp. 281–303.

Flick, Catherine and Worrall, Kyle. ‘The Ethics of Creative AI.’ Edited by Craig Vear and Fabrizio Poltronieri. The Language of Creative AI. Springer, 2022, pp. 73-88.

Gotterbarn, Don et al. ACM code of ethics and professional conduct. Association for Computing Machinery, 2018. Accessed January 16, 2024, at: www-acm-org.uoelibrary.idm.oclc.org/code-of-ethics. 

Guerrier, James. ‘The Effects of Poor Representation Run Deep.’ Paramount Insights, November 2021. Accessed January 16, 2024, at: insights.paramount.com/post/the-effects-of-poor-representation-run-deep/. 

Harris, John. ‘There was all sorts of toxic behaviour: Timnit Gebru on her sacking by Google, AI’s dangers and big tech’s biases.’ The Guardian, May 2023. Accessed January 16, 2024, at: www.theguardian.com/lifeandstyle/2023/may/22/there-was-all-sorts-of-toxic-behaviour-timnit-gebru-on-her-sacking-by-google-ais-dangers-and-big-techs-biases. 

Heflin, Judy. AI-Generated Literature and the Vectorized Word. Yonsei University, 2020. 

Mitchell, Melanie. Artificial Intelligence: A Guide for Thinking Humans. Pelican, 2019.

Qumer, Syeda Maseeha. ‘Timnit Gebru: Seeking to Promote Diversity and Ethics in AI.’ The Case for Women, 2023, pp. 1–36.

Santaella, Lucia. ‘Artificial Intelligence and Creativity Under Interrogation.’ Edited by Craig Vear and Fabrizio Poltronieri. The Language of Creative AI. Springer, 2022, pp. 43-55.

Sims Bishop, Rudine. ‘Mirrors, Windows, and Sliding Glass Doors.’ Perspectives: Choosing and Using Books for the Classroom, vol. 6, no. 3, 1990.

van Heerden, Imke, and Bas, Anil. ‘Viewpoint: AI as Author – Bridging the Gap Between Machine Learning and Literary Theory.’ The Journal of Artificial Intelligence Research, vol. 71, 2021, pp. 175–89.

Webb, Jen. Understanding Representation. SAGE Publications, 2009.

“Write me a story about two cats” prompt, ChatGPT 3.5, OpenAI, January 14, 2024: chat.openai.com.