Foteini Dimirouli explores the implications of AI tools for creative and poetic writing.
Foteini Dimirouli
Research Fellow and Access Fellow at the University of Oxford
Image in public domain by
Christin Hume (via Unsplash).
It is no secret that AI entails possibilities of large-scale manipulation. This is why, theorists routinely emphasise, there is an urgent need for regulation and transparency when it comes to AI-powered applications that are backed by billions of dollars and powerful corporations. As things stand, we might be told that we are interacting with a chatbot, rather than a person, on a customer service platform, but other forms of automation fly under the radar; for example, it is not always transparent whether a marketing campaign is AI-generated and tailored to personal data, or whether a CV is being processed algorithmically, and is therefore subject to a variety of selection biases, before being assessed by a human. If in these instances it is evident why user awareness is important, is this also the case in instances that are not necessarily connected to industry and profit, such as the generation of literary texts and, more specifically, poetry?
AI is now competently branching out into the “creative” domain, a realm where recognition is a currency even more powerful than financial gain. Debates so far have largely centred around whether humans can tell AI-generated writing apart from human writing and whether the quality of AI-generated text can match that of human-written text. Given that new research has shown that humans’ ability to discern the source of a writing sample on the basis of quality alone is limited, it seems more appropriate to train our attention on whether awareness of who it is that writes matters; in other words, whether the human status of the writing subject is of value in itself, despite questions of standalone literary merit.
Large Language Models and Poetry
Generative large language models, known as LLMs—GPT-4, ChatGPT, Google Bard, as well as powerful successors that are no doubt around the corner—accept image and text inputs and produce text outputs. Multimodal LLMs function as a sophisticated version of autofill, with phrases or sentences “completed” in sequences that are the outcome of ingesting enormous amounts of data. LLMs can produce anything ranging from humorous captions and taglines for advertising campaigns to opinion pieces and, pertinently to the subject of this essay, poetry. Outputs are not unique, since the submit button can generate new versions in a matter of seconds. Other than just sifting through those to find which is best suited to one’s needs, human agents can also improve their prompts, as well as train the LLM to produce more appropriate content through feedback, a process called fine-tuning.
In the area of literary writing, the frenzy of automatic text production extends to a highly sophisticated imitation game, as LLMs can produce verse in the likeness of the most recognisable voices of the Western canon. Shakespearean-like verse has been generated with remarkable success, and so have poetic responses to reputed poets’ verse or auto-completed versions of unfinished works.[1] Original verse is also a click away. The input of a simple prompt, such as “write a rhyming, 12-line poem, about a rose and a fox” obediently results in an output-poem matching these specifications, with scope to then improve and modify with the addition of parameters, such as “in the style of Edward Lear”.[2] Even as AI models make passing the Turing test seem trivial,[3] their competence in a genre thought of as the hallmark of human consciousness and introspection can be particularly unsettling. Resonant of the uniquely human processing of experience, poetry writing is a mode of expression that many assumed would never be composed at a satisfactory level by a non-human entity.
Is AI poetry any good?
Is LLMs’ capacity to write verse all that remarkable? There is some evidence that humans are prone to being, if not duped, at least flattered, by machine capacity. An equivalence for this sentiment can be found in everyday occurrences, for example when somebody with whom we do not share the same first language addresses us in our mother tongue more fluently than expected. Awe-struck, we are likely to react with praise in a scenario that exceeds our expectations, no matter how imperfect the interaction itself might be. In this sense, when humans encounter verse written by computers, the cards may well be stacked in the machine’s favour; as the social scientist Sherry Turkle puts it, there is an element of “human complicity in a digital fantasy” at play.[4] In other words, it might be machines’ unprecedented and largely unexpected capacity, rather than the quality of the output, that impresses us in AI creations.
At the same time, we need to interrogate how capable non-expert audiences are in spotting machine-written verse. In a blind experimental research project, humans were found unable to distinguish between human-selected high-quality AI outputs (the term used was ‘human-in-the-loop’, indicating human interference in selections) and the literary prodigy Maya Angelou’s writing.[5] Further, the methodical, close reading of poetry is hardly common practice. Outside erudite or academic circles poetry is largely associated with the esoteric and inaccessible, its intricacies obscured to the untrained eye. Even if we were to accept that ChatGPT cannot approximate the mastery of the likes of T.S. Eliot, as Carmine Starnino astutely argued in his essay “Poetry & digital personhood”,[6] we would still need to reckon with the fact that poetry’s popularity today does not reside in the celebrated classics and the echelons of high literature. Increasingly popular amongst younger generations, Instapoetry has been met with wide acceptance.[7] The trend has been redefining poetic writing by adjusting to the specifications of a social media platform that encourages brevity and relatable messaging, as well as the convergence between text and image.
In this context, it is unclear whether what we have come to identify as poetic novelty, the invention of a new style, or innovative nods to longstanding traditions, will host the same general characteristics in the future. The argument that AI poetry is deficient because it lacks nuance, complexity, and, importantly, novelty, might therefore be losing momentum. By this reasoning, and supposing that AI poetry-generating capacity extends, at the very least, to the production of easily digestible verse that reads as human and can capture audiences’ interest, it is unclear whether distinctly human labour is a necessary presupposition for success and popularity in the poetic realm. AI’s foray into poetry generation makes it increasingly difficult to argue that LLMs cannot be intelligent to the extent of being creative, or at least be creative on the premises of inclusive aesthetic criteria and varied understandings of taste.
Human-made VS Human-like
Is there any difference between artworks that are human-made and those that are human-like? Placing emphasis on the value of the artwork, independent of the creator, has been inherited by a long history of critical theory that has challenged the primacy of the writing subject. By the time AI entered the scene, the idea of the author as a cohesive subject had long been called into question. Post-structuralists pronounced the death of the author in the 1960s, attempting in this way to free artistic creation from the tyranny of the creative personality and to open up avenues of active and involved reading. The key message was the following: if we set the life and intention of the author to one side, we are better equipped to see how creation is always to some extent derivative and shared; a product of linguistic inheritance and a collective tradition without which the work would not exist, as well as of a reader bringing in their own experience and interpretations.[8] But AI is now leading us to rethink this movement toward depersonalisation. If the machine ingests preexistent data to produce anew, is mimetic, derivative, and creative quite despite of individuality, personal history, or intention, then AI creations might not differ from ours in any substantial way.
Obviously, when criticism moved away from the writing subject as a locus of grounded subjecthood and intention, it did so on the presumption that the writing subject was human. The realist poet and the surrealist poet, the eponymous poet and the anonymous poet, had their humanity in common, and it was on these grounds that the author’s relationship to the work was deliberated and challenged. But now renewed emphasis on subjecthood is in order. Since AI creativity has come to the forefront of attention, the notion that a machine might be becoming just like us is alarming. LLMs have been referred to as “stochastic parrots”, capable only of mimicking and offering up random combinations without understanding or critical overview. A Google engineer made headlines after claiming that the AI chatbot he was working on had become sentient, meaning that it could feel and emote in a human way, before he got suspended on the grounds of these claims.[9] Jason Allen, who won first place in an art competition for the AI-generated artwork “Théâtre D’opéra Spatial”, emphasized that his own input was instrumental to his success. In a post, he stated: “I have been exploring a special prompt that I will be publishing at a later date, I have created 100s of images using it, and after many weeks of fine-tuning and curating my gens, I chose my top 3 and had them printed on canvas after unshackling with Gigapixel AI”.[10] Competition awards are a golden standard of artistic worth, and Allen’s defense of his own contribution to the creative process—notice the language of personal distinction when he refers to “a special prompt”, “many weeks of fine-tuning”, “my gens”—betrays that human input is still held sacred.
At the same time, the very fact that Allen’s input consisted of belaboured prompts and sorting through vast output is a predictor of what the future holds: the illustrator or artist who is not in some way also a prompt engineer, using AI tools to their benefit, might soon be out of work. We might already be there if one considers the extent to which Midjourney and Dall-e have replaced human creators in the industry of illustration. What is really at stake in these debates is not so much how human the machine really is—concerns about AI being “sentient” can be discounted on the grounds of various philosophical arguments about consciousness—, but what it means to be human and, at the end of the day, whether being human holds significance for our creative future. Poetry, thought of as a hallmark of introspection, is often the platform on which such tensions are thrown into relief. To reach conclusions about this would mean to radically rethink how we approach creative subjecthood, identity, and authorship.
To ponder whether human creativity matters broaches broader questions populating this new era. The series Black Mirror, in an episode about a deceased spouse, branches into the idea of human replicas.[11] Overcome by grief, the surviving partner pays for a service that enables her to live chat with her dead husband through a platform trained on the vast database of texts the couple had exchanged during their time together. The result is uncanny, as the chat service faithfully replicates the husband’s tone of voice and sense of humour. She then opts for an upgrade and orders a humanoid robot that is a replica of her husband in both appearance and behaviour. The replica-robot ends up residing in the attic, in a peculiar living arrangement spurred by a strong attachment to the past and the relationship. Such fictional speculations bring into view questions about whether we would be content with an exact robotic replica of a fellow human, indistinguishable in every way from the real thing. Similarly, to us as readers, they interrogate whether we are content with a reading experience that does not emanate from a creative spirit that has a beginning and an end in the shared temporal dimension of life. Even when poets are anonymous, or unknown, there is a sense of a shared understanding of joy and suffering that does not merely reside in words, but also operates upon assumptions about the writer’s humanity. Even if language is an imperfect medium through which to truly understand the pain of the other, we still know that there is something universal about, for example, being heartbroken, which has been largely communicated through an empathetic bond between cultural consumer and creator, not just in poetry, but also in music or film.
Intention as a Political Act
Another consideration concerning the humanity of the writing subject is that of intention, not in the sense of free will—although it is true that the machine always obeys orders and cannot make independent creative decisions, cannot say “not today”—but also in the sense of political intent. In 1984, when the first book of poetry written by a computer programme was published and attributed to the mysterious figure “Racter” (under the title The Policeman’s Beard Is Half-Constructed), reviews deemed it “surrealist” in its effect, and this language survives in the assessment of various AI-generated endeavours. Granted, poetic absurdist experimentations—including the genre of “nonsense” poetry, most famously represented by Edward Lear—are characterized by unexpected combinations that do not obey grammatical rules or logical sequences. Playfulness and incoherence, and even metaphor and metonymy, can be as profound as they can be random, and easily generated through AI training on preexistent patterns. The entire surrealist movement was premised on the idea of automatic writing, which derives from the subconscious and the dream state, unbound by linearity and accuracy, able to express part of the psyche that structured prose and verse are not able to access. Yet the value of automatic writing did not merely reside in its execution. It was the product of the intention to disrupt norms, the outcome of observing Freudian ideas about the fragmented self and the dream state, and a response to the political realities of the First World War and the collective disillusionment these brought about. If we exclude the intention to create radical art, the ability LLMs have to yield human-seeming absurdist experimentations altogether misses the point. This is another instance that highlights why the quality of the final output should not be the sole consideration in thinking about the value of AI-written poetry.
Conclusion
When we test the notion of AI-generated poetry against majorly influential movements in literary criticism and the history of ideas—the death of the author for the sake of the birth of the reader, or the automatic writing of the surrealists—it becomes clear that additional categories are needed to grapple with poetic value in its intrinsic relation to the manifestation of a distinctly human subconscious. For this to be the case, transparency is indispensable in the creative realm, too. Knowledge of who it is that writes is key to our ability to determine whether ‘human-in-the-loop’ is a standard to be seriously upheld as part of a perennially shifting and volatile set of criteria that determine what we think of as creative worth.
Footnotes
[1] Some examples of this trend see news releases on AI poetry “ChatGPT amazes Twitter users with Shakespearean-style poem on climate change”, by Shubham Singh in Business Today, 19 March 2023 <https://www.businesstoday.in/latest/trends/story/chatgpt-amazes-twitter-users-with-shakespearean-style-poem-on-climate-change-374009-2023-03-19> and “Robot artist to perform AI generated poetry in response to Dante”, by Alison Flood in The Guardian, 26 November 2021 <https://www.theguardian.com/books/2021/nov/26/robot-artist-to-perform-ai-generated-poetry-in-response-to-dante>.
[2] Published pieces offer advice to poets on how to utilise ChatGPT to their creative benefit, see for example “ Structuring Creativity: Poetry Generation with ChatGPT”, by Paweł Sierszeń in Medium, 4 December 2022 <https://medium.com/@Pawel.Sierszen/structuring-creativity-poetry-generation-with-chatgpt-e7ffb4568196>
[3] The Turing test being the homonymous creation of Alan Turing in the 50s, used to test the ability of a machine to exhibit intelligent behaviour equivalent to, or indistinguishable from, that of a human.
[4] Sherry Turkle referred to this human complicity as “The Eliza effect”. The term occurred to her upon observing her students interact with the chatbot Eliza and sharing deeply personal information, despite knowing that they were dealing with an intelligent machine. Turkle, Alone Together: Why we Expect More from Technology and Less from Each Other (New York: Basic Books, 2011), pp. 31-2.
[5] Nils Köbis and Luca D. Mossink, “Artificial intelligence versus Maya Angelou: Experimental evidence that people cannot differentiate AI-generated from human-written poetry”, Computers in Human Behaviour 114 (2021): 106553.
[6] Starnino argues that “poetry as an art of brilliant accuracies, of realities redescribed in ways that bind sound to perception”. Starnino, The New Criterion, April 2022 <https://newcriterion.com/issues/2022/4/poetry-digital-personhood>
[7] Niels Penke, Instapoetry: Digital Image Texts (Palgrave, 2023).
[8] As indicative of this strand of critical thought, see Roland Barthes, “Death of the Author” (1967) (Norton Anthology of Theory and Criticism), Michel Foucault, “What is an Author” (1969) (Norton Anthology), Julia Kristeva on ‘intertextuality’ in “Word, Dialogue, and Novel” (1966) and “The Bounded Text” (1966-67), but also older texts such as T.S. Eliot “Tradition and the Individual Talent” (1919), and E.M. Forster “Anonymity: An Enquiry” (1925).
[9] See Guardian news “Google fires software engineer who claims AI chatbot is sentient”, 23 July 2022 https://www.theguardian.com/technology/2022/jul/23/google-fires-software-engineer-who-claims-ai-chatbot-is-sentient
[10] Matthew Gault, “An AI-Generated Artwork Won First Place at a State Fair Fine Arts Competition, and Artists Are Pissed”, Vice, 31 August 2022, https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissed.
[11] “Be Right Back”, directed by Owen Harris, Black Mirror Series 2, Episode 1, 2013.
Foteini Dimirouli
Dr Foteini Dimirouli is Research Fellow in English and Access Fellow at Keble College, University of Oxford. Her forthcoming book Authorising the Other: C.P. Cavafy in the English and American Literary Scenes (Oxford University Press), explores how the poet C.P. Cavafy entered the worldwide literary scene during the 20th century. Her new research project investigates the rise of digital poetics and the refashioning of literary authority in the digital sphere. It has been funded by the Athens-based Research Centre for the Humanities and the Onassis Foundation. As part of this project, she organised the international symposium Digital Poetry/Media Poetics at the Princeton Athens Center in September 2022.