The Rise of AI Writing Accusations: Who's Being Targeted and Why?

The past year has seen a surge in AI writing scandals, particularly those involving literary prizes, raising significant questions about the criteria and fairness of accusations. The Caine Prize for African Writing, awarded to Simbarashe Kundizeza for his short story “The Blue Fighter,” became a focal point when The Atlantic questioned its authenticity, citing the AI detection tool Pangram which flagged nearly the entire story. Kundizeza vehemently denied using AI, stating, “For the avoidance of doubt, I wrote ‘The Blue Fighter.’ I did not use generative AI to write it. It is entirely my own work.” This incident highlights the growing tension in the literary world, fueled by AI writing tools and detection software.

The literary industry has been rife with similar controversies. In March, Hachette canceled the US release of Mia Ballard’s novel “Shy Girl” following a New York Times report suggesting AI authorship. In May, Jamir Nazir faced accusations of AI use for his Commonwealth Short Story Prize-winning story “The Serpent in the Grove.” Researchers at Stony Brook University also suggested that H.M. Wolfe’s bestseller “Daggermouth” was significantly AI-generated. Further accusations led to debut author Jerry Falade losing a $2 million book deal, and Haitian Canadian author Thélyson Orélien was removed from the Goncourt prize longlist after his work reportedly failed an AI detection test.

As generative AI becomes more accessible, distinguishing between human and machine-generated text is increasingly challenging for readers, judges, and publishers. This has created an atmosphere of anxiety and suspicion, leading to the proliferation of AI detection tools. Companies like Pangram claim near-perfect accuracy in differentiating human and synthetic writing, aiming to mitigate the perceived harms of AI models. Max Spero, CEO of Pangram, a machine learning engineer by trade, positions his company as a protector of the literary ecosystem, aiming to reduce the friction caused by the ease of AI content generation.

Pangram gained significant attention after the Commonwealth Foundation announced Nazir’s story as a winner. While some readers found the story's ornate style peculiar and indicative of AI, others pointed to common stylistic devices that are also prevalent in human writing. The notion that certain stylistic choices, like em-dashes or specific phrasing, are exclusive to AI is often debunked by their common use among human writers. Patrick Nathan, an author, suggests that AI writing often mirrors “bad writing” because it’s trained on vast amounts of human-generated text, much of which is not of high quality. Accusations against Nazir were bolstered by Pangram’s 100% AI flag, with figures like Ethan Mollick, a professor at the Wharton School, sharing results on social media.

A notable aspect of these AI detection efforts is the background of some of the most vocal accusers, many of whom come from the tech sector rather than the literary world. Nabeel S. Qureshi, a former engineer at Palantir, has publicly discussed markers of AI writing. Furthermore, questions arise about the selection of works subjected to scrutiny. The accusation against Orélien originated from Samuel Fitoussi, a conservative French newspaper columnist. Many of the prominent scandals have involved non-white authors, including Nazir, Ballard, Wolfe, Falade, Kundizeza, and Orélien. Doyle D. Calhoun, a professor at the University of Cambridge, suggests this trend may be linked to a broader anti-woke sentiment in conservative online discourse, noting a potential resentment towards the changing demographics in the literary establishment and the tendency for such accusations to disproportionately affect non-white writers.

Pangram asserts that its tool only flags content with high confidence, claiming a false positive rate of 1 in 10,000, supported by some studies, though a University of Maryland study indicated a 2% false positive rate. Older studies, predating Pangram’s creation, found higher false positive rates for non-native English speakers. Pangram’s internal study reported a lower rate for such texts. The challenge for writers is proving their work is original when AI detection tools, even with low false positive rates, can still lead to widespread doubt. Without confessions or access to an author’s complete digital history, definitively proving AI use is difficult. The Caine Prize judges opted for a “human-centered approach,” requesting draft history from Kundizeza, raising questions about the extent to which writers must reveal their process for exoneration.

The literary community grapples with the implications of using AI detectors, viewing it as a potential technological overreach. John Warner, author of “More Than Words: How to Think About Writing in the Age of AI,” argues that relying on automation for discernment can lead to a de-skilling of critical judgment. While the Commonwealth Foundation chose not to use AI detection software for its prize, relying instead on a multi-faceted review of authors’ creative processes, the literary magazine Granta canceled its partnership after feeding a nominated story into Claude, an AI model not designed for detection, which suggested it was unlikely to be human-authored. Granta’s publisher, Sigrid Rausing, expressed the ambivalence of the AI age, noting the irony that AI itself might be the most effective tool for identifying AI-generated content, even as definitive proof remains elusive.