As AI’s endless clichés continue to encroach on human art, the true uniqueness of our creativity is becoming ever clearer

Creative writing used to be a human prerogative: do it well, do it badly, but either way endorse the consensus that to write about human experience was worth the candle and the coffee. Here was an essential human act, so much so that poetry formed a critical part of the computer pioneer Alan Turing’s original test: to determine whether an unseen respondent to a series of questions was human or a mechanical imposter. The Turing Test is often simplified to denote a single crossing point between two territories, human and machine. Pass the test, and artificial intelligence can stroll on over to our side of the line. Take a look around. Decide what to do with us. But, first, it has to pass.

In the paper ‘Computing Machinery and Intelligence’ (1950), published in the journal Mind, Turing set out his objective: ‘to consider the question, “Can machines think?”’ In true human fashion, he immediately re-phrases the question, at some length, and eventually arrives at the ‘imitation game’, modelled on a drawing-room entertainment from before the internet, before television. The original game he has in mind involves a guesser in the hotseat who poses questions to a man (X) and a woman (Y), who are out of sight and hearing in a separate room. The guesser has to determine from their written answers which is the man and which the woman. X tries to mislead, and wins if the guesser is wrong; Y wins if the guesser is right. Try it, it’s fun.

In this context, the first question posed in Turing’s proposed test is less surprising than at first it seems: ‘Will X please tell me the length of his or her hair?’ Next, Turing asks, equally politely: ‘Please write me a sonnet on the subject of the Forth Bridge.’ Two questions in, and the contested boundary between human and machine thinking is already looking for answers in literature, in art. Turing’s 1950s version of X – the participant aiming to mislead – replies: ‘Count me out on this one. I never could write poetry.’ To imagine this answer, in the second phase of his game, Turing’s complicated brain is playing the role of a machine playing X, hidden from sight and typing its answers, pretending to be a man (who previously played the game pretending to be a woman). I know, but if the test were easy an air-fryer could pass it.

Turing isn’t suggesting that a machine can’t write poetry. In the convoluted logic of the imitation game, X calculates that in 1950 ordinary people didn’t write poetry, a commonsense assumption that every computer masquerading as human should know. Among other prejudices from the mid-20th century, Turing’s paper makes incautious references to race, religion and the Constitution of the United States. He likens the inability to see computers as sentient as equivalent to the ‘Moslem view that women have no souls’. Turing wades in: he doesn’t compute as we would now.

And neither do the future computers of 2026 that he was trying to envisage. Any of today’s large language models (LLMs), like ChatGPT or Claude, can write an instant sonnet on the subject of the Forth Bridge. I typed in Turing’s test question, and Claude 4 threw up 14 lines of poetry including the abbreviated word ‘mathemat’cal’, for the scansion. The poem made sense, and was formally a sonnet, and appeared in seconds.

Whether or not this counts as thinking, Turing intuits that the frontier he’s marking out will be picketed by the arts. In his paper, he picks a fight with an eminent neuroscientist of the time, Sir Geoffrey Jefferson of the Royal Society, who believed that ‘Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain – that is, not only write it but know that it had written it.’ For Jefferson, addressing the Royal College of Surgeons in his 1949 Lister Oration, mechanising the efforts of the infinite monkeys on typewriters didn’t really count.

These days, in the arts, it’s harder to share Jefferson’s confidence. The advances made by AI prick at artistic vanity – the work of a human artist can’t be all that special if a machine can replicate the results almost instantly. That hurts. A great human artist, we’d like to believe, amplifies and defends the exceptionalist spirit of our species but, in an echo of the anxieties that haunted early photography, a demonised version of AI threatens to steal away our souls. Encroaching on the best of what we can do and make and be, machine art intrudes onto sacred territory. Creative artists are supposed to be special, inimitable.

Turing’s Imitation Game paper was published 14 years after the first Writers’ Workshop convened at the University of Iowa, in 1936. Turing may not have known, with his grounding in maths at King’s College Cambridge, that elements of machine learning had already evolved across the Atlantic in the apparently unrelated field of creative writing. Before Iowa, the Muse; after Iowa, a method for assembling literary content not dissimilar to the functioning of today’s LLMs.

First, work out what effective writing looks like. Then, develop a process that walks aspiring writers towards an imitation of the desired output. The premise extensively tested by Iowa – and every creative writing MFA since – is that a suite of learnable rules can generate text that, as a bare minimum, resembles passable literary product. Rare is the promising screenwriter unfamiliar with Syd Field’s Three-Act Structure or Christopher Vogler’s Hero’s Journey: cheat codes that promise the optimal sequence for acts, scenes, drama and dialogue. In the same way that an LLM is designed to ‘think’, these templates are a form of reverse engineering: first study how the mechanics of Jaws or Witness made those movies sing, then identify transferable components for reassembly to achieve similar artistic success further down the line.

To a computer-programmer, reverse engineering as a machine-learning mechanism is known as back-propagation. In A Brief History of Intelligence (2023), Max S Bennett shows how this methodology has already helped in the development of image recognition, natural language processing, speech recognition, and self-driving cars. Supervising coders work to isolate the required answer in advance, then go back to nudge input responses until the artificial neural network arrives at the pre-set solution.

The mysterious magic ingredient has been debated in English on the printed page since at least 1580

If only writing were so simple. According to figures from Data USA, up to 4,000 students graduate each year with creative writing MFAs in the US. No one expects that number of Great American Novels to show for so much studying, despite the fact that many hopeful writing careers start with the prompt mentality invited by Chat GPT: I want to write a bestseller like the one that blew me away last summer. Or, for the more adventurous: something new but relatable, a novel/memoir hybrid with literary credibility and strong narrative momentum, like a cross between Lee Child and Annie Ernaux. Thank you. I’ll wait. But not very patiently.

Clearly, when the end result is compared with the original intention, the back-propagation method is fallible for creative writing courses and LLMs alike. To revisit Jefferson, as quoted by Turing, the finished work is undermined when inspired by the wrong ‘thoughts and emotions’, whether blind ambition in student writers or blind obedience in computers. Something more is required, and the mysterious magic ingredient has been debated in English on the printed page since at least 1580, when Sir Philip Sidney reached for the essence of exemplary creative writing in An Apology for Poetry. When it worked, he concluded, good writing could both teach and delight. It provided a guide to living well in a more accessible form than theology or history or philosophy. Creative writing was special.

So special, in fact, that no one has yet been able to break down the findings of English literature departments – what makes literature work – into sufficient granular detail to reformulate as instructions actionable by an LLM. Or by a creative writing student. Nor are the efforts being made in this area by other art-forms particularly encouraging. ArtEmis is a large-scale dataset designed to record and subsequently predict emotional responses to works of visual art. The scheme matches emotional annotations from more than 6,500 participants to textual explanations of what they’re seeing, and from this data ArtEmis hopes to enable the back-propagated creation of artworks that provoke equivalent emotional responses.

The understanding seems to be that if a machine can create a visual image that generates a controlled set of feels, then art will have been successfully created. Which sounds plausible, except human emotional responses are notoriously capricious. The ArtEmis procedure already has an analogue precedent in Hollywood, but if focus groups worked reliably for the arts, then cinemas would be full of bangers. It’s worth remembering that the 2023 strike action by the Writers Guild of America won significant protections against the use of generative AI in screenwriting, specifically disallowing the replacement of human writers by AI. This hasn’t noticeably boosted the production of great art movies. Human writers still make so-so films. Without any intervention from AI, we continue to paint indifferent canvasses and write forgettable novels.

Bad art is something human beings love to do, in vast numbers. It’s part of who we are, and when abandoned by inspiration we trust in the same methods we’ve programmed into LLMs. As predicted by Turing, ‘digital computers … can in fact mimic the actions of a human computer very closely’, and for the creation of failed artistic product we’ve taught artificial intelligence all our dodges. Creative writing that falls short, whether originating in a garret or in an Nvidia chip, ‘writes’ by selecting language units that commonly fit together, as recognised from published material available in the public domain. Familiar word combinations are assembled into almost convincing sentences, a tired use of language formerly called out as cliché. LLMs are cliché machines, trained on a resilient human weakness for generating maximum content with minimum effort.

This explains the headline in June 2025 in the British publishing industry’s leading trade magazine The Bookseller: ‘AI “Likely” to Produce Bestseller by 2030’. The headline referenced a conference speech by Philip Stone of Nielsen, a company that compiles UK book-sales data. I expect he’s right about that bestseller, because LLMs will come for genre writing first – police procedurals, spy thrillers, romances – re-treading identifiable formulas with proven popular appeal. Eager to please (‘Hi Rich, how are you today?’), AI also has the advantage, shared by surprisingly few human writers, of being able to churn out derivative product without embarrassment.

Fortunately for everybody else, the endless capacity of an AI to deliver rule-bound and resolution-directed narrative has an unexpected benefit: AI is the tool that will prove not all writing has the same value.

Writing has been reluctant to imagine new ways of reading, despite the vistas opened up by new technologies

To escape the dead man’s handle of cliché, readers live in hope for organic associations, speculative leaps and surprise inferences. Whereas, to an AI, which is fed the answer before the question, ‘surprise’ remains an elusive concept. This objection to machine thinking was raised as long ago as 1842 by Ada Lovelace about one of the earliest computers, Charles Babbage’s Analytical Engine (for historical context, Iowa’s first creative writing get-together, though informal, took place in 1897). ‘The Analytical Engine has no pretensions whatever to originate anything,’ Lovelace observed. ‘It can do whatever we know how to order it to perform.’ Her italics emphasise the contrast with human thinking where originality, among artists at least, is a cherished value.

The visual arts, more than literature, have kept alive the modernist imperative to ‘make it new’. The Turner Prize, for example, awarded to the strongest UK contemporary art exhibition in any given year, is permeated by a sense that if it’s not new it’s not art. For visual artists, formal curiosity comes with the job, exploring new ways of making to invite new ways of seeing. Writing, on the other hand, happily rewards the comfort of familiar forms, which justifies, in the UK, the existence of a separate Goldsmiths Prize for fiction that ‘extends the possibilities of the novel form’. Because most other prize-winning novels aren’t doing that.

Writing has been reluctant to imagine new ways of reading, despite the vistas opened up by new technologies. Transferring books wholesale to Kindle and Audible is little more than digital haulage, and makes literature, in its complacency, vulnerable to proficient AI re-runs of familiar material, lowering the odds on that imminent AI-generated bestseller. Writers, or more accurately their publishers, seem to have mislaid any sense of urgency around the importance of difficulty, or curiosity about the astonishing returns that formally daring work can provide. It’s a rare book proposal submitted to a mainstream publisher that dares promise a book that’s not like all the other books.

Lovelace, thinking about how machines think, instantly identified the importance of originality. Or as a Marianne Moore poem has it:

these things are important not because a
high-sounding interpretation can be put upon them but because they are
useful; …

Originality, in the arts as in science, enables the human project to move forward. Any discovery that is new and true extends the scope of reality. In this context, art that only pretends to be original won’t get us anywhere very interesting.

The Turing Test is basically a test of lying. Can a machine, adopting a recognisably human strategy, pretend to be something it isn’t? Passing Turing’s Test calls for an act of deception, leaving the deceived human interrogator vulnerable to primitive fears about impersonation and imposture. Art is supposed to see the truth beyond this kind of lie, and the original creations worth defending are in a category so extraordinary, because of the intensity and authenticity of Jefferson’s ‘thoughts and emotions felt’, that they exist in a permanent present tense. What Toni Morrison does is unbelievable.

Whereas what an AI does is probabilistic. An LLM’s calculation of the most likely sequence of words is the least likely way to create great writing. Anyone working at a more emotionally engaged level than statistical probability, genuinely creating new work, has a better chance of resonating with readers, however that affinity is expressed. ‘If literature is a street brawl between the courageous and the banal,’ Greg Baxter wrote in his memoir A Preparation for Death (2010), ‘I bring the toughest gang I know: the pure killers, the insane.’ Baxter’s literary gangsters do not kneel before the most likely next word. Baxter values his ‘pure killers, the insane’, while computers as envisioned by Turing receive instructions to be ‘obeyed correctly and in the right order.’

We can defy AI creep by encouraging the human ambition to make art, unassisted, whether successful or otherwise

I don’t doubt that LLMs can be asked to imitate transgression, but obeying that instruction makes them ludicrously phony and the enemies of art, even though in their advanced contemporary forms they appear better equipped to respond to Turing’s stabs at English literature. In 2026, for example, ChatGPT and Claude make short work of the Turing Test challenge of 1950 to explain Shakespeare’s creative choices in Sonnet 18. Why a ‘summer’s day’ and not a ‘spring day’? Easy (just ask them, they know the answer). LLMs now ace most of Turing’s original questions, and if they can’t write a sonnet like Shakespeare, then neither can I. That doesn’t mean I can’t think, and Turing makes the same reasonable allowance for computers. They too are allowed their limitations, and his attitude to machine intelligence follows the logic of Denis Diderot’s parrot: if the illusion of understanding is sufficiently convincing, it qualifies as understanding. The machines are faking it until they make it.

Or in Turing’s words: ‘God has given an immortal soul to every man and woman, but not to any other animal or to machines. Hence no animal or machine can think. I am unable to accept any part of this.’ Turing invokes God for the sake of the 1950s, but he rejected the idea that humankind is ‘necessarily superior’ to the rest of creation, whether man-made or otherwise. He sides with materialist philosophers like Democritus and Thomas Hobbes, seeing the mind – whatever it might be – as located entirely in the physical structure of the brain. An AI is a physical structure, leading Turing to judge that whatever an AI can’t do, it can’t do yet.

In which case, how should writers and artists react to this situation as it stands now? We can attach stickers to the dustjackets of novels saying ‘Human Written’, as recently trialled by the UK publisher Faber and Faber. Visual artists have labels that say ‘Created with Human Intelligence’ or ‘Not by AI’, and maybe hashtags can keep AI at a distance until a generational talent arrives to save human honour in a blaze of truly original style and content. Take that, AI. See how much catching up you have to do.

More proactively, in the meantime, the rest of us can defy AI creep by defending and encouraging the human ambition to make art, unassisted, whether successful or otherwise. Art is an affirmation of human existence, the transmission and reception of messages about encounter and connection. One inner life can touch another and, for best results, nurture a creative process that no LLM can imitate. Marcel Duchamp called art ‘this missing link, not the links which exist’, an insight that arrives in the 21st century as a straight refutation of the imitative LLM creative model, stuck in its feedback loops and repeating existing sequences. Not for ChatGPT the electric shorting between inner lives, which in writing is most readily accessible in memoir. What anyone remembers is theirs alone, an undigitised storehouse of authentic human experience.

When Turing was deep in thought, according to his biographer Andrew Hodges, he used to scratch his side-parted hair and make a squelching noise with his mouth. Inside his head, at around the time he devised the Turing Test, he heard sceptical voices telling him a computer would never be able to be ‘kind, resourceful, beautiful, friendly’. His future machine brains wouldn’t ‘have initiative, have a sense of humour, tell right from wrong, make mistakes, fall in love, enjoy strawberries and cream,’ and so on. Turing was making comparisons with his remembered lived experience. What AIs couldn’t do was memoir.

Taking this idea as my starting point, I recently launched the Universal Turing Machine, a human proposal for a new way of writing and reading. The Universal Turing Machine is an expandable online grid of 8 x 8 squares, like a chessboard, and writers are invited to claim a grid for themselves and fill each of the squares with 1,000 words of memory. The reader can move randomly between memories and voices, playing an equally active role in the space Duchamp identified: ‘art is the gap’. Twice a year, I plan to tile new grids to those already online, steadily increasing the size of this collective experimental memoir, amplifying the diversity of human existence and creating a subjective encyclopaedia of true-to-life experience.

The Universal Turing Machine format is designed to encourage writing as a mode of thinking, which is what the arts – seeing, listening, writing, reading – have always offered. A memory that knows it’s being remembered is up there with the hardest, cleverest kind of thinking we can do, and why, for the purposes of his test, Turing couldn’t keep his hands off literature. AIs can’t yet emulate writing as its own mode of thinking, or reading, or remembering, and it doesn’t help to learn to write by reading everything. Just as memoirs aren’t improved by total recall.

The communication between writer and reader, artist and audience, is the nearest we come to telepathy

To see the miracle of human artistic selection in action, consider the French experimental writer Georges Perec’s novel La Disparition (1969), or A Void. This is the book that contains no instances of the letter ‘e’, the kind of systematic constraint an LLM could replicate in a flash. What the computer brain can’t do is add Perec’s life experience. The letter ‘e’ in French sounds like ‘eux’, meaning ‘them’. Perec’s father died while fighting in the war. His mother was deported from Paris to Auschwitz by the Nazis. The two of them are missing from their son’s life and from his novel, which then becomes the opposite of a disappearance, drawing attention to their distorting absence in a triumphant act of artistic reclamation.

Towards the end of ‘Computing Machinery and Intelligence’, Turing unexpectedly mentions that ‘the statistical evidence, at least for telepathy, is overwhelming’, and ‘If telepathy is admitted it will be necessary to tighten our test up.’ The communication between writer and reader, artist and audience, is the nearest we come to telepathy: to transmitting and receiving information between minds. Turing recognises that his machines will struggle to match this human refinement, and although not everyone discovers telepathy through art, anyone with an individual experience of strawberries and cream can try. The effort itself is worthwhile, and encouraged by projects like mine: the Universal Turing Machine welcomes human contributors, no test required.

Or more accurately, to do the work of recomposing memory in writing – to think in this distinctly human way – is itself an act of resistance. It reframes Turing’s test in favour of the part played in his original imitation game by Y, who aims to tell the truth, and doesn’t seek to mislead. X can’t have memories on your behalf; can’t fake it, won’t make it, and a knowledge of self remains now as always an assertion of cognitive sovereignty. In writing the self, Y becomes convincingly human. Y wins. The boundary between human and machine thinking remains intact, refortified by a self that won’t and can’t be outsourced.