Last July, the Hachette Book Group announced in a now-deleted post that it was “thrilled” to acquire BookTok sensation Mia Ballard’s latest horror title, Shy Girl. After its self-publication in February 2026, discussions and allegations on social media communities began to spread that the book contained artificial intelligence-assisted writing. A month later, the publisher pulled the plug on Ballard’s contract, turning it into one of the biggest conspiracies in the literary world in recent times.
Among the many allegations, and a New York Times article, was a Brooklyn-based, now University of Maryland- and Chicago-backed AI detection tool whose “near-zero false positive rate” methodology investigators swore by. What Pangram Labs unleashed with the Shy Girl controversy opened the Pandora's box of career-making or career-breaking social media trials.
The debate is no longer confined to accusations and detection. In France, the newly instituted Prix Pascal de Littérature IA, announced in October 2026, takes the opposite approach, inviting submissions of works co-written with AI tools, provided authors disclose the models and methods used. The prize offers a striking counterpoint to the growing scrutiny of AI-assisted writing in publishing.
Here, we analyse in detail how the world is coming to grips with distinguishing between human- and machine-generated writing.
How did the use of AI-generated writing pick up pace?
Generative AI chatbots like OpenAI’s ChatGPT gained prominence after a wave of large-language models (LLMs) sparked public curiosity when OpenAI unveiled it in late 2022. With capabilities for researching, drafting, fact-checking, and rewriting long, detailed essays, users soon began using it for human creative tasks. What began with “check this text for grammar and spelling errors” and “does a blue shirt go with khaki trousers” swiftly transitioned into “generate me a 3,000-word essay for my assignment”.
As capabilities grew, newer models became better equipped to loop in human feedback and, hence, became more conversational by the day. Moving from research labs to everyday life at a rapidly adapting speed, generative AI apps started breaking application download records soon after, with users ranging from university administration and students to governments, hospitals, journalists, researchers and businesses.
What is academia doing to tackle the problem?
Universities, among other educational institutions, had little to no time to react to the AI boom. As the primary playground of AI use—and also its most threatening one—they didn't set standards quickly. Many US professors recorded AI-assisted text plagiarism within weeks of new application laws, which led to a spree of bans across public schools in districts like New York, Los Angeles, and Baltimore. Policy then took a U-turn before eventually settling on a middle path that embraced the inevitable. By this time, independent ‘AI-assisted’ AI-text detection software like Turnitin, QuillBot, and GPTZero had been gaining popularity among evaluators.
In parallel, India watched the developments unfolding. Institutions—mostly private—published AI codes in an independent capacity, with some capping AI text alongside existing plagiarism tolerance to promote ‘responsible usage’, while others having strict no-AI-adherence policies in the bid to protect the academic integrity of students.
While the Union government tried to formalise statutory guidelines alongside existing ones, execution has continued to flatline. None of the codes laid out specific guidelines for standardised assessment, leaving a big void in AI literacy and subjective tolerance among teachers and examiners. A case in point is the death of a second-year IIT Bombay student who was allegedly caught using ChatGPT during an exam, which sparked widespread debates and protests regarding malpractice allegation-led agony and the institutional asymmetry that lacks enough teeth to curb it.
How do AI detection tools work?
Five years ago, a layman coming across an article online or an examiner marking grades on a mid-term exam sheet would scrutinise the copy mainly for its creative form—assuming a human wrote it. But a lot has changed. And some argue they must know what they’re reading isn't written by a machine. Naturally, this gives rise to a whole industry of detecting and flagging ‘slop’.
Most AI detectors are binary classifiers that label texts as human-written or AI-generated based on probability scores. This includes n-gram frequency analysis or counting repetitive or predictable patterns of specific word or phrase usage.
Older models like GPTZero followed statistical methodology by looking for a balance in the length of a sentence or rhythm variation deviating away from natural styles and towards ‘statistical signatures’ left by known LLMs ChatGPT, Claude and Gemini. Most newer models use LLMs themselves as “meta detectors” to interpret text.
At its core lies a classification problem. Hidden by the noise of marketing gimmicks is the ever-rising chicken-and-egg problem of what can be truly attributed as AI-generated text if it’s being flagged out by AI models which use similar text to train their models, which, in turn, have been under suspicion all along.
And why don’t they really work?
Damaging as it may be, widespread access to detector tools with disclaimers taking a back seat has resulted in smear campaigns on social media with AI text scores being flashed as mugshots attached to writers. Worse, they can carry, or even reinforce, existing human prejudices.
A breakthrough 2023 paper by Stanford University scholars led by Weixin Liang showed that AI detectors were significantly more likely to flag text as AI-generated from non-native English speakers than from native speakers. While detectors attributed essays by US-born eighth-graders as “near perfect”, over 60 percent of essays written by non-native students were flagged by detectors. Apart from Ballard’s Shy Girl, H M Wolfe’s Daggermoth and Jerry Falade’s Call Me, I’ll Hide The Body fell victim to similar allegations. All of them are non-native English writers and people of colour.
It’s not as if the market is unaware. Grand View Research valued the AI-detector market at $580 million in 2025 and projected it to grow sixfold by 2033. But tools showing prejudice among users can be bad for business. Yet, somehow, we don’t learn much.
Turnitin, another well-known AI-text detection tool used in journalism and academic circles, promotes itself with a false-positive rate below 1 percent. However, the software says it is only reliable at detecting AI use when documents under the lens contain at least 20 percent AI content. But that does not negate false positives. In fact, an Australian federal government standing committee noted that a false-positive rate as low as 0.7 percent in detecting AI content can result in over 60,000 texts being flagged over a semester in the country. Such a figure is definitive and consequential when software is embedded as judge, jury, and executioner in policy.
Are newer models any better at resolving bugs?
Most newer models like Pangram understand that texts can fall on the borderline, with Pangram Labs co-founder and CEO Max Spero admitting to making an intentional choice of “erring to the side” that is mostly responsible for a higher false-positive rate in these models. But there’s another addition to the chicken-and-egg problem: text humaniser.
The latest paper by Jonathan Karr Jr and others, titled ‘Why AI Detection Fails for Academic Integrity’, confirmed that Pangram’s 3.2 model (second to the latest with minor bugs) could not differentiate AI editing from full LLM drafts and treated both as misconduct in a series of tests. It concluded that “Commercial detectors on published abstracts punish acceptable AI-assisted editing, compound risk for English-language learners who use AI for translation and fail to catch humanised synthetic text.”
Two months ago, Spero pushed the idea of his tool becoming something like a ‘nutritional label’ for written prose. This time, the text in question was Haitian-Canadian Thélyson Orélien’s latest book C'était Ça ou Mourir. Amid recent contentions over Pangram’s state-of-the-art and ‘rigorous training model’, allegations of creative dishonesty ring hollow for many, including writers and publishers who feel directly under attack. Spero, whose X bio says “slop janitor”, remains active on social media forums like Reddit and LinkedIn, engaging in discussion.
While the world contemplates entrusting such tools to lighten the burden of upholding the creative or academic integrity—or not at all—emerging tech needs to show promising results with robust methodology.