Three paragraphs, from three different hotel reviews. Can you tell which, if any, were AI\u2011generated?<\/p>\n
\u201cThe hotel is in a great location for everything. Lots of places to eat and drink. The hotel itself is always abuzz. The tavern located on the ground floor is definitely a must. Food, service, prices and atmosphere were great.\u201d<\/p>\n
\u201cA good hotel, though the room had the proportions of a well-appointed lift. Slept well, shower was excellent, staff were friendly. Breakfast was busy but competent. Would return, though probably not with a very large suitcase.\u201d<\/p>\n
\u201cExcellent base for a London trip. The room was quiet, the bed comfortable, and everything worked exactly as it should. Staff were helpful without hovering. A smooth, unfussy stay from start to finish.\u201d<\/p>\n
How do you reckon you did? Most people, says Claire Hardaker, a professor of forensic linguistics at the University of Lancaster, get this kind of judgment right only about 60% of the time. Her online test, Bot or Not<\/a>, asks users to identify the fakes in a series of 15 reviews. The middling success rate might come as a surprise to those convinced they can spot AI writing at 50 paces. When doubts<\/a> were raised in May about the authenticity of a prizewinning short story by Jamir Nazir, social media users were lightning-quick in their condemnation. \u201cIf you know, you know,\u201d commented one. Nazir later told<\/a> The Atlantic that he didn\u2019t use AI.<\/p>\n Hardaker says her respondents tend to rely on a few quick rules of thumb to identify AI language, including the presence of cliches and the use of dashes. The \u201crule of three\u201d, where words or phrases are arranged in a satisfying trio, is also thought to be a giveaway. \u201cPeople have learned very simplistic rubrics and now just madly apply them everywhere.\u201d<\/p>\n There\u2019s a problem, though: these \u201ctells\u201d are also characteristic of human writing, which, after all, the large language models (LLMs) that produce them were trained on. \u201cYou could go back to Charles Dickens and say he had AI, because he used the em dash too.\u201d And orators have known about the rule of three ever since Julius Caesar said Veni, vidi, vici<\/a>. In our hotel review examples, only the first one was authentic. Did you clock it?<\/p>\n Perhaps because it is so hard to know for sure, suspicion has become the order of the day. In the literary world, accusations of AI use now bedevil writers, with varying levels of justification. A debut horror novel, Shy Girl, was withdrawn by publishers Hachette after rumours circulated<\/a> online that the author had relied on AI, which she denies; Steven Rosenbaum\u2019s book The Future of Truth, a serious study of \u201chow AI reshapes reality\u201d, was found to contain numerous hallucinated quotations, which the author acknowledged<\/a> in an apology.<\/p>\n Media organisations, including the Guardian, field increasing numbers of complaints about supposedly AI-generated text. These include intuitions about particular turns of phrase, but also comments about typos and grammatical errors. In one case, the word \u201cafter\u201d was inadvertently duplicated in a sentence. \u201cI can\u2019t imagine a human editor\/proofreader missing something like this,\u201d wrote one reader, displaying a touching faith in our copy-editing abilities.<\/p>\n The problem is that not only does AI train on human writing, but humans are stylistically influenced by AI, the interplay creating a kind of linguistic hall of mirrors. Short of an author admitting it, it\u2019s hard to say for certain whether an individual piece of writing is AI or not. That uncertainty is a recipe for paranoia.<\/p>\n And if you\u2019re tempted to reach for a commercial screening tool to sort human from machine, that comes with uncertainty too, says Hardaker. \u201cGiven that some of us naturally write in a way that would be seen as AI-like\u201d \u2013 she mentions neurodivergent people, for example \u2013 \u201cthat will be detected as AI. And you can modify AI output to make it seem more human-like. You put that kind of content into an AI detector, you\u2019re going to get wacky results.\u201d As someone who has served as an expert witness in court, she\u2019s \u201cextremely sceptical\u201d about their efficacy.<\/p>\n The newly popular detector Pangram, which boasts false positive rates of around 1 in 10,000<\/a>, has been shown in independent tests to be highly effective<\/a> at detecting AI writing even when it\u2019s been run through a \u201chumanizer\u201d app to disguise its origin. But questions remain. I was able to fool it on the first attempt (see the screenshot below) by channelling a bombastic register that might well be characteristic of AI, but could equally be the work of someone with a naturally bombastic style \u2013 or, more to the point, a writer who has been steeped in the output of the LLMs that power ChatGPT, Claude and Gemini. That, increasingly, is all of us.<\/p>\n Photograph: David Shariatmadari<\/p>\n Vast amounts of AI writing are now being published every day \u2013 from advertising copy to academic abstracts and fiction. At the same time, it looms ever larger over our lives via auto-generated email suggestions, \u201cAI overview\u201d search results, and the responses to our chatbot queries. At this level of exposure, it\u2019s no longer a question of whether AI is changing language, both the way we speak and the way we write; the question is how. And should we resist, or embrace it?<\/p>\n We\u2019ve known for some time that LLMs generate text that can be slightly different from human writing, on average. Often this only becomes clear when you look at large amounts of material. One eagle-eyed researcher linked the sudden popularity of the word \u201cdelve\u201d to LLMs back in 2024<\/a> after searching a database of scientific papers. Other \u201cfocal words<\/a>\u201d that AIs have tended to overuse include \u201cshowcase\u201d, \u201cboast\u201d, \u201cunderscore\u201d, \u201cgarner\u201d, \u201calign\u201d, \u201csurpass\u201d and \u201cintricate\u201d. But, again, any individual piece of writing could entirely innocently make use of this vocabulary.<\/p>\n In a further twist, some researchers think the \u201cdelve\u201d phenomenon might not be down to the models themselves, but the humans tasked with evaluating and steering them in a process known as \u201creinforcement learning with human feedback\u201d. For workers who are \u201cunderpaid, stressed, and under time pressure\u201d, it seems<\/a> \u201ccertain words are treated as a proxy for quality\u201d and the model is inadvertently trained to use them more often. In other words, \u201cdelve\u201d might owe its meteoric rise to the fact it doesn\u2019t seem like the kind of word an AI would use. (A separate suggestion<\/a> that it appeared more often because it was characteristic of English used in Nigeria, where many RLHF workers lived, isn\u2019t borne out by the data.)<\/p>\n There are other patterns<\/a> we can distinguish: LLMs love nouns, but they seem to use pronouns less than humans. This might reflect the fact they don\u2019t do as much talking about themselves or other people as we social creatures do. They like attributive adjectives (\u201cthe uncomfortable chair\u201d), but not predicative ones (\u201cthe chair was uncomfortable\u201d), perhaps because they prefer to deliver information in small, dense packages, whereas we pad things out. Different models have clear idiosyncrasies<\/a> \u2013 you might even call them \u201cdialects\u201d: Gemini enjoys saying \u201chere\u2019s a breakdown\u201d, while Deepseek often responds with a cheerful \u201cCertainly!\u201d. When asked to edit formal English from around the world, AI tends to flatten and homogenise towards an Anglo-American standard, in a process researchers have termed \u201ccultural ghosting<\/a>\u201d. Thus the perfectly acceptable request in Indian professional English to \u201cKindly do the needful & revert back at the earliest\u201d gets \u201ccorrected\u201d to \u201cPlease complete the task & respond promptly.\u201d<\/p>\n The evidence that aspects of LLM-speak have escaped into the \u201creal\u201d world, changing the way humans use language when AIs aren\u2019t around, is now rolling in. One study analysed<\/a> thousands of unscripted conversations and found that words like \u201cdelve\u201d and \u201cboast\u201d spiked after ChatGPT was released. Another showed the frequency of \u201cdelve\u201d in academic abstracts actually dropping<\/a> after it was singled out on social media, in a sign that AI\u2019s influence might play out in complex ways.<\/p>\n Illustration: Pete Reynolds\/The Guardian<\/p>\n Does any of this matter? Language<\/a> changes all the time \u2013 words come in and out of fashion, and new technology has always been one of the forces behind this. AI does seem to be generating particularly high levels of anxiety, though. Why? \u201cI think where it scares people is that idea of encroaching into sentience, of becoming the new human,\u201d Hardaker says. Since 2023, she\u2019s expanded the Bot or Not project into speech and music, and has noticed just how viscerally people react when a song they\u2019ve enjoyed turns out to have been composed and performed by a machine.<\/p>\n Gary Shteyngart, a novelist who teaches creative writing at Columbia University, noticed a similar strength of feeling among his students at the prospect of AI literature. \u201cWhen one of my graduate students said \u2018as an experiment, I\u2019m going to be writing a part of this piece with AI\u2019, the other students became so angry, they wrote letters to me saying how awful this was.\u201d<\/p>\n \u201cThere\u2019s a kind of implicit bargain between writer and reader where you know the work that you\u2019re getting is generated by a human being, and I think it felt like an assault on that,\u201d he says. \u201cReading literary fiction is this incredible Vulcan mind meld with another human being, entering someone else\u2019s consciousness. With AI I\u2019m entering the simulacrum of another person\u2019s consciousness, one degree removed, or many degrees removed. How sad is that by comparison?\u201d<\/p>\n For Hardaker, \u201cI guess it impinges on what we think of as what makes us special, what makes us valuable and unique\u201d. At the same time, the music-generation model she uses \u201chas generated some absolute bangers. I listen to them, unironically, in my car, and I enjoy them quite a lot.\u201d<\/p>\n Could the same happen with literature? Will a machine-authored novel one day take its place among the 100 greatest of all time<\/a>? Peter Stockwell, professor of literary linguistics at the University of Nottingham, thinks AI may be able to do the basics, but it can\u2019t scale the heights. \u201cIf you want something that\u2019s very familiar and very mediocre and entirely functional, it\u2019s amazingly good at that.\u201d<\/p>\n One way to think of language, he says, is as a series of layers, with words at the bottom followed by phrases, clauses, compound sentences, all the way up to narrative structure. \u201cAI is really good at the lower levels. It\u2019s learned lots of our syntactic structures and so everything looks well formed and grammatical. But, the higher up you go, the less good it is.\u201d The arc of a story is particularly hard for AI to get convincingly right.<\/p>\n \u201cIf you\u2019ve got an AI to write a narrative, it can do a pretty good job of having a sequence of events and something happen at the end. But it wouldn\u2019t be a very tellable narrative,\u201d he continues. \u201cNothing startling or interesting would happen. And if there is anything startling, it will generally look like a mistake, rather than a brilliant twist.\u201d<\/p>\n The secret sauce of great writing remains secret \u2013 even to the academics who study it. \u201cLinguists don\u2019t understand, really, how language works at its higher levels,\u201d at the level of discourse, storytelling, enchantment. \u201cWe can\u2019t build a machine to do something when we don\u2019t know how it works.\u201d We do have some idea of what it might boil down to \u2013 and that\u2019s our fundamentally social natures and, tied in with that, the fact that we are \u201cwetware\u201d \u2013 human flesh, with its spikes of adrenaline, rushes of dopamine, craving for social contact, all of which find expression in language\u2019s structure and the way we use it.<\/p>\n double quotation markWould I work with an LLM? Of course! Why not?Jeannette Winterson<\/p>\n There are two broad models in linguistics, explains Stockwell, one that sees the brain as a computer, parsing grammatical structures and computing meaning from them, and the other that sees it as fundamentally embodied, something reflected in language by the fact that, in many languages, we understand by \u201cseeing\u201d or tend to think of \u201cup\u201d, where our head is, as good (we get \u201chigh\u201d and feel \u201clow\u201d). \u201cOne of the key things is that the current AIs don\u2019t have a body, they don\u2019t exist in the world, so they don\u2019t know what it feels like to be in the world as a human.\u201d<\/p>\n For Shteyngart, feeling is essential: \u201cToday is the first warm day in New York. And if I was to start writing a novel, I think that [it] would be warmer. I think I would filter what I know through the warmth of the day. I think if I ate a really wonderful lunch and sat down to write, there would be more sensuousness in my writing.\u201d<\/p>\n \u201cThe love of the body and its encounters with the physical world are what drives some of the best of literature. So I almost feel sorry for these LLMs, as I\u2019m talking about them, because they\u2019re pushed into some horrible machine in the Bay Area, and they just don\u2019t know how wonderful life is.\u201d<\/p>\n One much feared effect of the mass use of LLMs is that they act as a flattening force \u2013 smoothing away the variety and idiosyncrasy of human language into a kind of beige goo. That\u2019s a legitimate concern, as far as it goes, though it\u2019s not a new one. People have long angsted about the homogenising effects of American film and television on accent and vocabulary, and there are subgenres of language \u2013 political euphemism, customer-service prattle, therapy-speak \u2013 that have spread further from their home territory than many might like. The crucial thing, though, is that their influence tends to generate a backlash \u2013 and there\u2019s no reason to think things will be different this time.<\/p>\n In fact, our capacity for innovation might ultimately be the thing that truly distinguishes human writing \u2013 particularly the literary kind \u2013 from AI. \u201cThe whole point of an LLM is that it\u2019s trained on existing language. So it\u2019s always retro,\u201d says Stockwell. \u201cI could get an AI and say \u2018write me a short story in the style of Virginia Woolf\u2019 and it\u2019ll do a decent job. But what you can\u2019t say is \u2018write me a story in the unique style of the next great, serious literary innovator\u2019. It couldn\u2019t possibly do that.\u201d<\/p>\n That\u2019s because, once again, it lacks the social environment, and the body, that give rise to characteristically human motivations. \u201cWhy does somebody do something new in an art form like literary writing? It can be out of annoyance or irritation with what\u2019s gone before. Or it\u2019s because somebody sees things in a different way than the run of the mill, or sometimes just because people are antsy and wanting to do something different, or a little bit crazy or isolated.\u201d<\/p>\n The are plenty of examples from history, says Stockwell: \u201cAfter the bureaucracy and uniformity of the first world war, you\u2019ve got this sudden, huge, antithetical artistic movement in the rise of surrealism and Dada; similarly, after the austerity of the second world war you get the psychedelic movement, art and literature changes again, quite radically. So there always seems to be that sort of kicking against the norms. It\u2019s hard to think how you would program an AI to do that, because AI works on an existing large body of material. It\u2019s the embodiment of the conservative-with-a-small-c status quo.\u201d<\/p>\n Originality is so important for novelist Jennifer Egan that she\u2019s quarantined herself from the technology entirely. \u201cI feel a danger of infection, to use a kind of loaded metaphor,\u201d she tells me. \u201cI know they stole some [of my] stuff, and there\u2019s nothing I can do about that, but I\u2019m not giving them one more word voluntarily.\u201d Anthropic used pirated copies<\/a> of books, including Egan\u2019s novels, to train its chatbot Claude; most LLMs use language from individual queries as additional training data. She sounds exasperated: \u201cI don\u2019t want to partake of this kind of language spam that they\u2019re offering.\u201d<\/p>\n