winning

On May 18, several X users They made their suspicions public about the last Commonwealth Prize winner of short stories: the winning story in the Caribbean category, published in ‘Granta’ (a very prestigious British magazine that for decades has been the thermometer of the Anglophone literary canon), reeked of ChatGPT. The author’s photo, in fact, did not seem real either. And when the magazine responded to the scandal, it did so in a way that ended up confusing everything: They asked Claude if the text was from AIand Claude said no. How it is detected. Recognizing the prose of a language model is not as simple as it seems, but it is not as difficult once the eye has been trained for some time. Language models do not write looking for the right word– They generate the statistically most probable token, taking into account the context, and it is a process that can be identified. For example, the famous “it’s not X, it’s Y” structure, used as the Rosetta Stone of text AI identification. But there is more: accumulation of metaphors without clear referents, verbs like “go deeper into”… annotators hired to fit the models using RLHF (reinforcement learning with human feedback) reward that kind of bureaucratic clarity, which also makes everything more obvious. What does the story have? In ‘Granta’s story things are said like “the humming noon” or the “sweet air with the smell of cane and oblivion.” Some authors, such as Benjamin Breen in this excellent analysis have detailed, many of these turns speak of a special attraction for ambient sounds and vague emotional states (nostalgia, sadness, oblivion), which seem to want to touch upon a materiality that the model does not have and, of course, does not understand. The accumulation of sensory stimuli is a creative writing textbook instruction that models apply mechanically and without discrimination. It’s easy to see once you’ve learned to identify it. Why detectors don’t work (yet). The problem is that recognizing that writing intuitively is one thing, and demonstrating it objectively is another. The first generation of automatic detectors (GPTZero, Originality.ai, Turnitin AI) They have a long history of errors. OpenAI, for example, retired its AI Text Classifier in July 2023 after acknowledging that it only correctly identified 26% of AI-generated text and marked almost 9% of human text as artificial. The only exception to this trend documented so far is Pangram. His technique, called mirror datatrains the classifier with pairs of stylistically identical texts but with different origins. The result, according to the first benchmark independent In September 2025, false positives are close to zero and false negatives are between 2% and 4% in medium and long passages; competitors are scoring around 10-40%. However, again, it is not so easy to trust a tool that sells a “humanizer” along with its text reports about the presence of AI, sometimes with outrageously high percentages. The educational part. In the world of books we are facing a specific scandal, that of ‘Granta’, but in American universities we are seeing a permanent escalation of hostilities. In this extensive reportfor example, we are presented with ten students and teachers trapped in a spiral with no way out: the teachers pass the assignments through AI detectors, the detectors generate false positives on students who have not touched any chatbot, and those students resort to humanizers (or directly to writing worse) to avoid accusations. Joseph Thibault, founder of Cursive, has tracked 43 humanizers with a combined audience of 33.9 million views. Grammarly has developedFor example, Authorship, a tool that records the writing session so that students can prove that they wrote the work: according to the company itself, five million reports of this type were generated in the last year. A teacher states in the article: “The better you write, the more the AI ​​believes you are AI. I put my own articles in the detectors just to understand how they work, and it marks me at 98% every time, without having used AI at any time.” Below are the essays. The ‘Granta’ scandal occurs while the publishing market registers another symptom of the same problem. According to data from the Wall Street Journal that is also collected by Res Obscura, the best-selling non-fiction book of April in the United States (‘London Falling’, by Patrick Radden Keefe) placed 13,468 copies in its first week; the first novel was close to 105,000. The president of Harper Group attributes it to podcasts: according to a recent survey, 62% of men and 54% of women listened to one last month, compared to 46% and 39% in 2023. The reason is the same that led to YouTube before: they promise to satisfy in forty minutes what it takes a book to satisfy three weeks. AI is the next step: it does not compete with books or podcasts, but rather replaces them with a summary generated in ten seconds that answers questions without anyone having to write, edit or read anything. The problem, as Breen points out, is that this model aims for immediate responses and eliminates precisely what makes the nonfiction book valuable: the need for attention, permanence over time and all the reflective nuances that this implies. When the AI ​​copies a specific author. The writer Vauhini Vara went further. As he told on Voxcommissioned researcher Tuhin Chakrabarty to train a model on his three published books and several journalistic articles to generate passages from, theoretically, his next novel. He then mixed them with his own fragments and sent them to his closest friends. No one knew how to distinguish them. What’s more: another conclusion they drew from the experiment is that readers tend to prefer AI text over imitations written by humans when they don’t know the origin. When the source is revealed to them, the text ceases to interest them. Vara draws a conclusion from this: what matters to readers is not whether the text sounds human, but knowing that there is someone real on the … Read more