This essay by artist and researcher Pip Thornton documents and debates some of the outputs of ‘Writing The Wrongs of AI’ (WWAI), a project which explored creative ways to demonstrate the power that human words, poetics and writing might have in resisting the influence of artificial intelligence in the literary sphere and beyond. The essay documents some of the project’s successes while also reflecting on how the saturation of our current political and cultural economies by, with, and through AI technologies and logics frustrates and complicates the evasive, subversive, and rebellious qualities of language, thus potentially disarming poetry and creative writing as tools of refusal. One of the outputs of the WWAI project – a short satirical play performed at the 2024 Edinburgh International Book Festival – will be reprised (with new second act) as part of the Being Human Festival at the Scottish Poetry Library on 6^th & 7^th November 2025. More details and ticketing info to follow…

Introduction

In his 2012 book, ‘The Uprising: On poetry and finance,’ Italian philosopher Franco ‘Bifo’ Berardi makes a series of assertions about the relationship between language, technology, and the digital economy. Recognising the power of poetry as an “excess of language” that “exceeds economic exchange,” he evokes the revolutionary potential of a poetic language that insists on both financial and linguistic insolvency. According to Berardi, through resistance to the forces of capitalism,[1] language that maintains the potency of irony and the power of enunciation can avoid reduction to “semiotic debt” and can instead “create a new world.”

With the advent of Generative AI I am less hopeful of the power of poetry and its ability to evade monetisation and semiotic capture. For the last decade, I have sought to reclaim the agential power of language (in terms of human expression, emotion, and subversion) as a tool to interrogate the systems of digital capitalism that treat our words as commodified data [2]. I would like to think that my work has had some success in this respect, but what follows are examples from Writing The Wrongs of AI (WWAI), an ongoing research project which while designed to empower writers and the wider publishing industry in the face of technologies such as ChatGPT, has also highlighted how the saturation of our current political and cultural economies by, with, and through AI technologies and logics potentially frustrates and complicates the optimism of Berardi’s poetic resistance.

Writing The Wrongs of AI

In September 2023, I saw a Tweet from Scottish crime novelist Val McDermid calling fellow writers to arms over the theft of copyrighted work by AI companies for use in training Large Language Models (LLMs) such as ChatGPT, and it made me think about how writers could push back.[3] I began to conspire with Jenny Niven, incoming director of the Edinburgh International Book Festival, who was anxious about how AI might affect the Book Festival community. With her support, I secured funding to develop the Writing the Wrongs of AI (WWAI) project.[4] WWAI brought together a group of writers, publishers, poets, translators, and copyright experts to engage in a series of workshops, exchanging knowledge, experiences, and grievances and exploring how a range of language-based creative refusal tactics and interventions might help reclaim some agency in the midst of AI’s crushing inevitability.

There are, of course, precedents for evading power through words. From poems used to conceal or decrypt code in wartime[5] to dialects and vocabularies invented to elude authority,[6] poetic language lends itself well to subterfuge and secrecy, not least due to its ambiguity and liquidity[7]. In a digital age, we can attempt to evade the forces of digital capitalism and political oppression by making texts unscrapable, obfuscated, or inaudible/unreadable to machines,[8] or we can harness human creativity to imagine a language that refuses to be modelled or privatised or a speech so expensive that it breaks banks.[9] Playing with some of these ideas, WWAI allowed us to collectively harness the power of poetics and creativity to question, resist, and refuse AI’s incursion into the literary sphere.

As the WWAI project progressed, we uncovered glimpses of Berardi’s linguistic insolvency, facing off in its existential battle against the rampant commercial incentives of AI-generated text. Prompted in part by Alan Turing’s[10] use of Shakespeare’s Sonnet 18 to illustrate the shortcomings of machinic logic, we played around and tested ways in which poetic writing can refuse AI and where it cannot, revealing three potential sites of tactical refusal: silence, illegibility, and parody.

On silence

I’ve been interested in the idea of poetic resistance to systems of digital capitalism since a spoken-word poem evaded economic capture in my research on the monetisation of Google search terms,[11] but it was only a few years before audio data from video-hosting sites became collectable and, thus, monetisable. Such advances in technology rendered accidental acts of Socratic resistance (an insistence on speech over writing) obsolete. While meaningful resistance and refusal of these forms of AI might seem all but futile, the alternative is silence. Silence of voice, page, and screen – a complete withdrawal of our linguistic labour. Forcing the machines to talk to one another in the ultimate feedback loop might, indeed, satisfy Berardi’s revolutionary call for linguistic insolvency, but it is hardly practical.

There may, however, be ways in which silence can still refuse AI in a more positive way.  WWAI participant, Scottish Gaelic poet Peter McKay,[12] highlighted the power of ‘nothingness’ in poetry, suggesting we might underestimate the agency of the human reader or listener in appreciating the pauses and beats of poetic writing:

There are huge gaps between the lines and the words where the reader has to do a lot of the work… [AI has] got the structures and the patterns, but it doesn’t have the nothingness… the nothingness of poetry can be used to mess with AI.

While it may satisfying to know that these types of pauses and silences evade machine reading, they seem to matter more to the human and less to the machine for which letters and gaps are tokenised in the same way and differences between them are irrelevant.[13] Of course, when you scale up from poetry to the global music industry, ‘nothingness’ might become a more powerful and impactful tactic of refusal, as the recent ‘silent album’ released in protest at the UK government’s plans to facilitate the use of copyrighted material by AI companies shows.[14]

Another variation of silence became apparent in discussions about AI and accents, with participants with strong accents describing the frustration of not being heard or understood by voice recognition and AI tools. In the workshops, we watched the brilliant Burnistoun comedy sketch depicting the growing annoyance of two Scotsmen in a voice-activated elevator trying to get floor ‘eleven,’ a word which, when spoken in a strong Scottish accent, can be incomprehensible to AI.[15] Jo Gilbert, a Doric poet, recounted their own frustration in this respect, as they had lived in the Aberdeen postcode area of AB11 and frequently found themselves ‘without a voice’ when having to enunciate their address to automated telephone systems. We thought about how to utilise this voicelessness as a potential tactic of AI refusal, exploring ways of speaking so that humans, but not machines, could understand but recognised that ultimately, refusal in this way reveals a degree of privilege, as many (often vulnerable) speakers of minority languages with heavily accented ‘English’ voices cannot afford to be unheard or misunderstood and have no choice but to negotiate (with) these systems.[16] Once again, Berardi’s vision of ‘enunciation’ as poetic freedom is compromised.

The concept of silence and pause is also relevant here in respect to the temporal pressures and incentives of AI writing. In one of the WWAI creative interventions (see below), the boss of a fictional GenAI company justifies selling his writing tool with the provocation that “we can’t be spending each and every night rhyming couplets.”[17] But the promise of poetry generated in an instant seems an anathema to the often torturous but ultimately rewarding human writing process. As WWAI participant Sean Wai Keung put it, “do you know how long it took me to get those words in that order?” However, refusing AI by insisting on opting out or by insisting on labour- and time-intensive writing risks being dismissed as luddism, privileged, or exclusionary.[18] Ultimately, we must recognise that when systems of digital capitalism (of which GenAI is the latest) get to decide the value of human labour and creativity and to gatekeep essential systems and services, pausing, silence, and time spent thinking are seriously undermined as tools of resistance and refusal.

On Illegibility

Already marginalised in standardised publishing, writers using minority dialects and diasporic communities using creole, broken, or colloquial English now also face a landscape wherein livelihoods and creative agency are increasingly threatened by AI. But can non-normative language be a tool of power, evading the capture of quantifiable AI models and retaining its integrity?

The WWAI project’s main objective was to explore ways to restore a sense of creative agency in the face of AI. In some ways, however, for many writers, that ship has sailed – their work is now inextricably part of the huge language models that fuel tools such as ChatGPT. For others, though, there is still a chance to refuse. While the GenAI market is already dominated by the desire for rapid expansion in western and other dominant languages, those who write in smaller languages and dialects have had more time to push back. A great example of this is the Singapore writers who refused to let their work be used to train a government funded local LLM.[19]

There is, of course, a significant tension between the value of harvesting minority languages for the apparent purpose of representation and inclusion in commercial AI tools and their cultural value as vessels of (often complicated and contested) geo-specific identity[20]. Some participants were conflicted that feeding the machine[21] with minority languages and local dialects to diversify and democratise AI tools might also dilute or taint their linguistic stock. As poet Peter McKay put it,

there’s an equal risk of actually then passing over something that is at the moment not oversubscribed or overused by AI into the machine, and the machine takes over.

Indeed, it might not be competing languages that are important here, but rather universally agnostic linguistic devices such as metaphor that challenge the contextual integrity of AI-generated text. A fascinating insight came from WWAI participant Nasim Rebecca Asl, who considered that metaphor might be a much stronger tactic against AI than the as/like formulae of similes, “which would let the machine know that there’s an equation.”

We also thought about what can’t be stolen for machine learning models, leading us to think of how font, typesetting, or other non-standard print such as marginalia might resist machine reading.[22] WWAI participant, writer Camilla Grudova recalled her excitement at the realisation that

illuminated manuscripts can’t be read by the AI machines and they can’t steal that. It’s an exciting moment, and it made me look back to medieval literature and back to people like William Blake or Alasdair Gray who have their own fonts and use their illustrations and actually this could cause something really creative to happen while trying to elude the machine.

In a similar vein, we discussed the accessibility issues around the subtitling and captioning of audio by AI systems, whose glitches can sometimes be amusing but are also frustrating and non-inclusive. Spoken word, thus, ‘refuses’ AI, not by design, but by failure.[23] In all these examples of silence and illegibility, it becomes clear that while so-called ‘glitches’ in digital systems might offer a degree of (perhaps illusionary) agency and control, they also have the power to “determin[e] who is legible in dominant modes of computing” [24].

On Parody

The WWAI workshops provided a rich environment for participants to vent against AI that can’t ‘do’ a range of ‘human’ skills such as humour, irony, double entendre, or as WWAI participant Graeme Swanson pointed out, a comedic pratfall. In this respect, we discussed parody as a potential tool of resistance, but then art began to imitate life, and such distinctions went out the window.

A main output of WWAI was the LEX-9000 ‘Human Verses Machine,’ a spoof ‘automatic writing machine’ parodying tools such as ChatGPT and manifested in the shape of a 7ft robot. Jetted over from Silicon Valley, the LEX-9000 promised to alleviate the ‘drudgery’ of creative writing. LEX-9000 assists users who might be short of inspiration but was designed to become increasingly controlling, with users’ words involuntarily ‘optimised’ for economic gain. The intervention’s climax was the performance of a custom-written short play depicting the parody product launch of the LEX-9000 at the Edinburgh Book Festival. We created the character of Danny Celeste, eccentric CEO of fictional company Token AI, as well as a protest group – Page Against The Machine (PATM), which attempts to disrupt the launch.[25]

We had thought that the idea of a big American GenAI company coming to the Edinburgh Festivals to sell its wares was the stuff of fiction and parody. But we were wrong. As we were writing the play, I became aware that the Edinburgh Festival Fringe had partnered with Anthropic, a US AI company whose chatbot Claude strives to rival Open AI’s ChatGPT. [26] Anthropic had sent over representatives to hold workshops with Fringe performers, ostensibly claiming to show how AI can be good for the creative industries but, in essence, offering little more than marketing opportunities to tout their product. Indeed, their AI chatbot Claude is trained on exactly the same dataset of stolen work as ChatGPT, including, somewhat ironically, the creative content of many Fringe performers, past and present.[27] I attended the workshops and watched the Anthropic tech bros insisting that Claude will handle the menial jobs you don’t want to do, freeing you up for creative wonder, while simultaneously showing how their product can write jokes, plays, and computer games. The situation provoked such a deep state of cognitive dissonance in me that I still can’t quite work out whether irony (as Berardi’s poetic tactic) died or flourished at the 2024 Edinburgh Festivals. What I can say is that the potency of poetics and parody are severely restricted when cultural venues such as the Edinburgh Fringe buy into technologies that threaten the creative agency and livelihood of their own performers.

“Impotent Gestures”?[28]

While some of the artistic and poetic critiques described in this article can at first seem to follow Berardi’s faith in the power of poetry and creative writing, experience has shown that ‘artificially intelligent’ systems are now far too embedded in political, cultural, and social discourse to facilitate and sustain useful tools of resistance, at least in terms of outright refusal. Indeed, as more and more lived spaces, as well as online spaces, become sites of textual exploitation through ‘smart’ devices such as doorbell cameras and voice-activated assistants, and as words become the accepted currency for companies such as OpenAI, Amazon, Google, and Turnitin, Berardi’s paradigms of ‘enunciation’ and ‘voice’ lose their impact as acts of solidarity, refusal, and resistance. When words become the masters’ tools, everything we say or write makes their houses bigger.

In spite (literally) of the exploitative conditions in which the current digital economy forces us to exist, the WWAI project did reveal some potentially significant examples of linguistic agency in the age of AI. When faced with ChatGPT’s ‘uncanny’ imitation of her distinctive poetic style,[29] Camilla Grudova recalled the feeling that

It was like looking at a mannequin of myself… it revulses your spirit.

But discussions during the project made her consider more positively, in that

I need to be way more erratic in my creative processes, not just ’cause I don’t want to kind of become a parody of myself, but now there’s this thing chasing me that I want to be like, unpredictable for.

Likewise, another participant harnesses the “awfulness” of AI-generated poetry “to figure out what not to write.” The key here is agency – the insistence on using the machine as a tool and the refusal to let it use us and our words as passive commodities or tools of power. In this way, the insistence on linguistic ‘impotence,’[30] as well as insolvency, might indeed be the poetic key to AI refusal.

There is, of course, also a degree of optimism in insisting on the ‘liveness’ of unstreamed or recorded theatre, stand-up, or spoken word; spaces which can break algorithmic circuits[31], thus “reactivating the emotional body, and therefore reactivating  social solidarity,[32] starting from the reactivation of the desiring force of enunciation”[33] and which – for now – continue to evade data grabbing platforms and to frustrate the logics of digital capitalism.[34] So while we wait for the financial insolvency of companies like OpenAI and Anthropic, we do at least have some agency in refusing AI by insisting on linguistic insolvency – finding refuge in the liquidity of poetics and the humanity of poets.

[1] F.B.Berardi, The uprising: On Poetry and Finance (Cambridge, MA, MIT Press, 2012), pp. 7–35. Berardi uses the term ‘semio-capitalism’ to describe the neoliberalisation of political discourse but cites Frederick Kaplan’s concept of ‘linguistic capitalism’, which initially inspired my research in this area. See F.Kaplan, ‘Linguistic Capitalism and Algorithmic Mediation’, Representations, 127(1) (2014), pp. 57–63.

[2] P.Thornton, ‘A Critique of Linguistic Capitalism: Provocation/Intervention’, GeoHumanities, 4(2) (2018), pp. 417–437; P.Thornton, ‘Data resistance through public art: reclaiming narratives in/of the city’, in Morgan Currie, Jeremy Knox, Callum McGregor eds., Data Justice and the Right to the City (Edinburgh, Edinburgh University Press, 2023), pp. 326–345.

[3] McDermid had found her own work in the ‘Books3’ dataset of copyrighted (but pirated) material used to train models such as ChatGPT, which was revealed as part of an Atlantic investigation. See A.Reisner, ‘These 183,000 Books Are Fuelling the Biggest Fight in Publishing and Tech’, The Atlantic, 2023, https://www.theatlantic.com/technology/archive/2023/09/books3-database-generative-ai-training-copyright-infringement/675363/

[4] Writing the Wrongs of AI was facilitated through a Bridging Responsible AI Divides (BRAID) Fellowship, with funds from the Arts & Humanities Research Council (AHRC). BRAID is dedicated to integrating Arts and Humanities research more fully into the Responsible AI ecosystem, as well as bridging the divides between academic, industry, policy, and regulatory work on responsible AI. See: www.wwai.info; https://braiduk.org/projects-fellowships.

[5] Perhaps most famously, cryptographer Leo Marks gave a love poem (https://allpoetry.com/The-Life-That-I-Have), originally written to his late girlfriend, to British agent Violette Szabo in World War Two. ‘The Life that I Have’ was reprised as a means of subverting modern digital systems by artist and researcher Paul O’Neill as part of the Zoom Obscura project (https://zoomobscura.wordpress.com/paul-oneill/). See also: C.Elsden, D.Chatting, M.Duggan, A.C.Dwyer and P.Thornton, ‘Zoom obscura: Counterfunctional design for video-conferencing’. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, 2022, pp. 1–17.

[6] I’m thinking here of Ogham, the ancient Irish script designed to confound the Romans, and Polari, the slang dialect used in queer subculture to evade the police.

[7] P.Thornton, ‘Subprime Language & the Crash’, in Mark Graham, Rob Kitchin, Shannon Mattern and Joe Shaw, eds., How to Run a City like Amazon and Other Fables (Manchester, Meatspace Press, 2019).

[8] While it relies on voluntary compliance, robots.txt is one way to de-index web content. The internet has also facilitated more creative ways to escape linguistic capture, such as deliberate misspelling to avoid copyright. For instance, Chinese activists use emojis that represent the phonetic sounds of censored phrases such as #MeToo, anti-OCR fonts like ZXX (https://zxx-font-website.webflow.io/) and using written signs to convey political messages on applications such as TikTok. See: A.Hingle and M.Knodel, ‘Robots.txt Is Having a Moment: Here’s Why We Should Care’, Tech Policy Press, 2025, https://www.techpolicy.press/robotstxt-is-having-a-moment-heres-why-we-should-care/; V.Guttenberg, ‘Rice Bunnies vs. the River Crab: China’s Feminists, #MeToo, and Networked Authoritarianism’, Flux: International Relations Review, 11(1) (2021), pp. 31–39; M.Behrens (@MatthewBehren11), ‘People on TikTok are reporting ICE raids in small cities in various states. To get past algorithmic censorship, they’re writing down information on paper & holding’, Twitter, 2025, https://x.com/MatthewBehren11/status/1882625962489372850/photo/1.

[9] I refer here to Alain Damasio’s remarkable 2012 short story ‘Les Hauts Parleurs’ (‘The Loud Speakers’) and its protagonist’s lifelong crusade against the privatisation of language, as well as some poetic experiments from my own research. See: J.Foley, ‘Word of Mouth: Pip Thornton talks Poem.py, Engineering Fictions and the Monetised School of Poetry with Michael Rosen – Tues, 23rd Jan @ 4pm BBC 4’, Engineering Fictions, 2024, https://engineeringfictions.org/2024/01/17/word-of-mouth-pip-thornton-talks-poem-py-engineering-fictions-the-monetised-school-of-poetry-with-michael-rosen-tues-23rd-jan-4pm-bbc-4/. 

[10] A.M.Turing, ‘Computing machinery and intelligence’, Mind, 59 (236) (1950), pp. 433–460.

[11] See: https://pipthornton.com/2016/06/12/poem-py-a-critique-of-linguistic-capitalism; I would print out receipts for poems according to what their words were worth to Google in the context of an advert. The poem, ‘Bog-Eye Man’ by Jemima Foxtrot escaped monetisation as it only existed in audio form on YouTube and Vimeo so I couldn’t ‘scrape’ it.

[12] Peter McKay (Pàdraig MacAoidh) is the current incumbent of the post of Makar, the national poet of Scotland.

[13] See also my review in: A.C.Dwyer, N.O’Grady, P.Thornton, T.Straube, E.Gilbert, and L.Amoore, ‘Cloud Ethics: Algorithms and the Attributes of Ourselves and Others: Louise Amoore. Durham, NC: Duke University Press, (2020), 232 pp. 25.95’, The AAG Review of Books, 9(3) (2021), pp. 36-49.

[14] Listen to the Silent Album at https://open.spotify.com/album/1yUbRSEsj5ng38IpHpcLJt. For more details, see: https://www.isthiswhatwewant.com/

[15] View the comedy sketch at: https://www.bbc.co.uk/programmes/p00hbfjw

[16] Failures in AI translation tools used by refugees and authorities can jeopardise asylum claims. See: J.Bhuiyan, ‘Lost in AI translation: growing reliance on language apps jeopardizes some asylum applications’, The Guardian, 2023, https://www.theguardian.com/us-news/2023/sep/07/asylum-seekers-ai-translation-apps

[17] ‘Page Against The Machine’ was written and directed by WWAI participant, playwright Clare Duffy of Civic Digits Theatre company.

[18] Heated debate surrounded NaNoWriMo (National Novel Writing Month) when it refused to condemn the use of AI by writers on the grounds of classism and ableism. See: M.Herbst, ‘NaNoWriMo Organizers Said It Was Classist and Ableist to Condemn AI. All Hell Broke Loose’, Wired, 2024, https://www.wired.com/story/nanowrimo-organizers-classist-and-ableist-to-condemn-ai/

[19] N.Yong, ‘Writers and publishers in Singapore reject a government plan to train AI on their work’, Rest of World, 2024, https://restofworld.org/2024/singapore-writers-reject-ai-training/

[20] Chapman, ‘Revocalising human geography’, pp. 24–42; Medby, ‘Articulation of geopolitics’, pp. 18–25.

[21] See also: Muldoon, Graham, and Cant, Feeding the Machine.

[22] F.Kaltheuner, ed., Fake AI (Manchester, Meatspace Press, 2021) offers an example of using GANs to play with fonts.

[23] Indeed, some tech-based language solutions are not even programmed to recognise some languages. For example, a student recently alerted me to the fact that, on Duolingo, you can learn Klingon, but not Tibetan.

[24] D.Mahmoudi and A.Sabatino, ‘Witches as glitches: A response to Leszczynski and Elwood’, Dialogues in Human Geography, 12(3) (2022), pp. 389–392.

[25] A recording of the Page Against The Machine play, performed on 19 August 2024 at the Edinburgh International Book Festival can be viewed at https://vimeo.com/1078909157.

[26] In 2025 it was announced that Anthropic are reprising their workshops but this time as the distinctly Orwellian ‘Official Education Partner’ of the 2025 Fringe. See: https://www.edfringe.com/about-us/news-and-blog/anthropic-to-bring-ai-workshops-designed-for-artists-to-the-fringe/

[27] A search tool (https://www.proofnews.org/youtube-ai-search) built by Proof News confirms that content from comedians, including Jack Whitehall, Romesh Ranganathan, Stephen Fry, Ricky Gervais, and Graham Norton, was scraped via YouTube subtitles to train Anthropic’s chatbot. For more, see: A.Gilbertson and A.Reisner, ‘Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI’, Proof News, 2024, https://www.proofnews.org/apple-nvidia-anthropic-used-thousands-of-swiped-youtube-videos-to-train-ai/

[28] A frank assessment of the WWAI creative interventions from a workshop participant.

[29] Camilla asked ChatGPT to write a story in her style and instantly recognised the cliches and themes in her short story collection The Doll’s Alphabet (2017), which, in effect, proved that her work had been assimilated into the training dataset. We later checked via The Atlantic’s Books3 search portal to confirm this was the case. See: Reisner, ‘183,000 Books’.

[30] In the sense of refusing phallogocentric language. H.Cixous, ‘The Laugh of Medusa’, Signs, 1(4) (1976), pp. 875–893.

[31] Dekeyser and Lynch, ‘Control and Resistance’, pp. 53–74.

[32] For more on mutual aid as a tactic of AI resistance, see: D.McQuillan, Resisting AI: An anti-fascist approach to artificial intelligence (Bristol, Policy Press, 2022).

[33] Berardi, The uprising, p. 20.

[34] Another example of a ‘circuit-breaker’ in this respect might be the use of profanities in Google search queries, which Ethical Marketer Jane Shaw recently discovered automatically disables the often-unreliable AI-generated ‘overview’ which now appears at the top of search results. J.Shaw, ‘No. Bloody. Way. You know those annoying AI answers when you search using Google? You can stop them by adding a swear word to your search. I flippin’ kid you not.’, LinkedIn, 2025, https://www.linkedin.com/posts/shawthingjane_no-bloody-way-you-know-those-annoying-activity-7318841962155913216-aN-8?utm_source=share&utm_medium=member_desktop&rcm=ACoAAEm-EpgBaoPBLirW515Z5nqPPS4H92zEDsk.