A proposal for a different understanding of artificial intelligence in culture

The phrase ‘artificial intelligence’ is undoubtedly one of the most frequently used terms in contemporary technological, social and art discourse. But don’t worry, this short essay has a rather different purpose than dozens of other texts coming at us from various corners of the internet every day, repeatedly telling us how amazing and terrifying AI is. We would like to briefly introduce to you our book The Culture of Neural Networks: Synthetic Literature and Art (Not Only) in the Czech and Slovak Context, which we are currently finishing and which will be published by the Prague-based publishing house Karolinum next year in English. Our book is and isn’t about artificial intelligence, it’s about what is usually referred to as artificial intelligence. Our book is not actually about the technology itself. Rather, it’s about what this technology is causing in the world of art and literature – and, in fact, in the inner worlds of human imagination.

The idea of a thinking, creating machine has fascinated humans from time immemorial. Human imagination seems obsessed with the desire to animate the inanimate, out of the need to express natural human creativity, or perhaps match divine beings. Like Pygmalion, praying to Venus to bring his own creation – a statue of a perfect woman – to life. This is still in the realm of myth. However, there has been a long chain of real attempts to construct an artificial being, from the efforts of the robotics pioneers the Banū Mūsā brothers from the 9th century or Ismail al-Jazari in the early 13th century, who described how to construct programmable human automatons for music, to Leonardo da Vinci’s 1495 metallic knight, Pierre Jacque-Droz’s 19th-century mechanical ‘children’ that could write words and draw pictures, and finally, in the 1950s, the concept of artificial intelligence, based on the belief that human thought is replicable by computers.

And it’s quite interesting to observe how these seemingly purely technical efforts are accompanied by poetry. It’s as if a machine writing poetry would be the best evidence that the humanisation of machines has finally been achieved. It’s not a coincidence that one of the fundamental genres of electronic literature, computer-generated poetry, was born precisely in the 1950s, in a scientific environment, on large and slow mainframe computers. After its birth, the genre of generative poetry gradually developed, improving its results as computer technology advanced, and it established itself in the field of literary experiments. However, it did so without much attention from the media, including the literary media.

This, however, has changed quite recently. Bombastic headlines now appear in the pages of world newspapers, proclaiming that robots are writing poetry that is sometimes indistinguishable from human poetry, thanks to the so-called artificial intelligence. However, you will not find many instances of this term in our book. We have not embraced it due to its unreflective metaphorical and manipulative nature. Simply put, no true artificial intelligence in the strictest sense has been created yet. Instead, we have new generations of software based on machine learning and large language models. They can perform wonders and are improving rapidly, they just aren’t intelligent yet, being barely able to handle one type of task, often at the cost of generating a large amount of unusable output. That’s why we prefer to talk about artificial neural networks as the carriers of this software revolution and, at the same time, as the central point of the cultural formation we simply call the culture of neural networks. This culture arises not only from the activities of these fascinating technologies but also (or rather, especially) from the linguistic games that accompany them – and often inadvertently or even intentionally confuse, just like the above-mentioned newspaper headlines. Which is why our book explores not only the projects of generating literary texts and other kinds of art, but also the communication strategies employed in presenting these texts, and the ways of interpreting them.

Among other things, our writing about literature generated by artificial neural networks is structured by its developmental perspective, which is essentially chronological. It’s quite astonishing when we realise how short the time frame this neural network boom has occurred in. Recurrent Neural Networks (RNNs) became good enough to generate literary texts in the mid-2010s. However, the breakthrough came in 2018 when GPT-1 was released. After that, things progressed very quickly: the release of GPT-2 in 2019, GPT-3 in 2020, GPT 3.5 (a politically correct version of the previous model) in 2022, and finally, on 30 November 2022, Chat-GPT was launched, making generative AI accessible to practically everyone (in March 2023, GPT-4 was released, but it currently has limited access or requires payment and serves, broadly speaking, as a Bing search engine function). In less than a decade, these new and increasingly sophisticated language models were published and made available to a growing audience of enthusiasts.

However, this straightforward narrative can be somewhat misleading, especially if it is perceived as evidence of the continual improvement of AI’s ability to generate literary texts. As demonstrated by the authors of the recently published collection of generated poems ‘I am Code’ and older, seemingly outdated language models (in their case, GPT-3 in the davinci-002 version) can produce literary results that are better than the latest version of Chat-GPT, which is ‘forbidden’ by its developers from talking about itself and expressing negative opinions about human life. Nevertheless, we ultimately arrived at three model phases of the process of the establishment of neural networks in culture. Despite the short time frame in which this process occurred, it cannot be seen as unified but needs to be internally structured, especially considering the differing motivations of human actors who entered it in various phases:

1) Verification Phase: In this phase, the main role was played not by artists but by developers and programmers who, through projects with a literary impact, verified the functionality of algorithms and the sufficiency of language models at their disposal. As an example from the Czech context, we can mention the project of an automatic poet realised by the mathematical linguist and machine learning expert Jiří Materna in 2015, using an RNN network trained on texts from the web written by amateur poets. The generated results were practically indistinguishable from their originals. Similarly, with the same type of network, Cambridge software researcher Jack Hopkins worked on generating Shakespearean sonnets. The determination of scientists to create artificial literary text was sometimes, perhaps, excessive during this phase. This was certainly the case with the Japanese team from Future University Hakodate, which attempted to win a literary competition with a generated novel in 2016. However, it was later revealed that approximately 80% of the text was created by humans, with the computer only ‘remixing’ it. On the other hand, as an example of good practice, we can mention the Czech project TheAItre, which resulted in the generation of a theatrical play. In its creation, scientists from Charles University collaborated with artists from the Švandovo Theatre.

2) Artistic-Subversive Phase: In this phase, artists took on a leading role in conceptualising generative projects. It was no longer about verifying the capabilities of the technology and demonstrating its abilities; instead, it was about fulfilling a conceptual intention that often subverted literary norms or the technologies themselves and the media practices these technologies establish. A brilliant example of an artistically subversive approach to neural networks is the project by Slovak multimedia artist Samuel Szabó, titled Umelá neinteligencia (Artificial Non-Intelligence). In this project, a character-RNN network was trained by several different parodying models, for instance, to generate traditional Slovak nationalist poetry. However, the training intentionally fell short, so the resulting poems mimic nationalist poetry by the choice of topics but at the same time parody and ridicule it due to its semantic nonsense. A critical intervention into the very practice of generating texts using AI is the project ReRites by Canadian media artist David Jhave Johnston. This project involved generating 12 substantial poetry books in 12 months and aimed to demonstrate new possibilities for boosting human creativity while also highlighting the risk of artistic overproduction. Less subversive but artistically interesting results were achieved by the project of the fictional Slovak feminist poet Liza Gennart. The book resulting from this project, a collection of poems generated using GPT-2 but fine-tuned on mostly contemporary Slovak poetry, gained for its human authors Zuzana Husárová and Lubomír Panák the national prize for poetry Zlatá Vlna in 2021.

3) Vernacular Phase: This phase is characterized by a radically democratized approach to artificial neural networks, allowing anyone from the internet user community to engage in the generation of literary texts or other works of art. In this phase, generative literature becomes part of pop culture, or rather, amateur literary creativity, losing its scientific, technological and artistic exclusivity. Soon after the release of Chat-GPT, numerous poem generators connected via API to this attractive chatbot began to emerge rapidly, giving everyone the opportunity to play with creating versified text, generating poems in predetermined genre forms, on user-selectable topics. Thanks to easily accessible self-publishing services offered by platforms like Amazon, dozens of poetry collections by authors you’ve never heard of, and may never hear of in the future, have made their way into the book market. However, these books share something in common – an absolute lack of awareness of the generative literature context into which they unknowingly entered and an unwavering desire on the part of the authors to acquire symbolic capital still associated with poetry and the book as an artefact.

So, what has artificial intelligence, or as we prefer to call it, the culture of neural networks, achieved in the world of literature and art so far? We believe its consequences have mainly manifested in two areas. First, in the above-mentioned vernacularisation of literature and art. This process has been ongoing for a long time and certainly didn’t start with artificial neural networks. It is an organic part of the history of media that began with the printing press which democratised education previously confined to monastic scriptoria. However, neural networks have significantly accelerated this process of involving amateurs in the realm of literature. They have achieved something previously unimaginable and seemingly crazy – the removal of the last barrier on the path to participation in literature, which is the ability to write (literary) text. Having this ability is no longer truly necessary, as demonstrated by numerous generated books available on Amazon. On the other hand, there has arisen a need for new competencies, including the ability to create a meaningful and effective prompt that encourages the neural network to produce an acceptable literary output. Everyone will need this competence, even those who, for example, only occasionally use neural networks in their creative activities, as an ancillary source of inspiration.

However, the second significant area where the culture of neural networks has made its unmistakable mark is in the realm of literature’s self-reflection as a means of communication. It has posed new fundamental questions about who the author actually is and has pointed out the necessity of perceiving even technical actors in creative processes as active co-creators of the final work (after all, even a typewriter used to be such a co-creator, it just wasn’t as apparent). At the same time, neural networks have stripped authorship of a certain pseudo-mystical aura it once had – there is now no doubt that both the poem that appears to touch on the deepest aspects of spirituality and the one that strives to convey the impression of raw reality can be convincingly crafted without spirituality or life experience.

Moreover, the culture of neural networks necessitated a deeper reflection about readers, or the act of reading itself. When reading machine-generated literature, it’s difficult to avoid questions about how the text actually came into being, what it reveals about its data sources, and so on, even though they may not be obvious. Thus, when we read generated text as if it were human, we consciously engage in a game, granting the text the right to its mimicry and occupying ourselves with a literary self-reflective meta-reading.