Scan Your Draft for AI Detector Triggers
Paste a paragraph to check for uniform sentence length, overused AI vocabulary, and repetitive transition phrases before submitting your work. Everything runs in your browser.
Everything runs in your browser. Nothing is sent anywhere, and nothing is stored.
In April 2026, Hachette cancelled a horror novel. Not for plagiarism, not for poor sales projections, but because of an AI detector score. I have been writing and editing professionally for two decades, and I think every working writer now needs a file they hope they never open: proof they wrote their own book.
This guide covers what actually happened in the two cases every editor is now thinking about, what the published accuracy evidence says about the tools used, whether editing with Grammarly triggers Turnitin, and the provenance checklist I recommend you start keeping today.
Mia Ballard’s horror novel Shy Girl started as a self-published book, got picked up by Hachette, and was pulled before its US release. According to Slate’s April 2026 account, readers on Reddit and YouTube had raised suspicions for months, then the detection company Pangram analysed the full manuscript and flagged it as 78 percent AI-generated. After a New York Times story presented the evidence, Hachette cancelled the book.
CBC’s reporting adds the detail that matters most for the rest of us. Hachette cancelled both the US and UK releases. Ballard denied using AI to write the novel, but said it was possible an editor she had worked with on the self-published version might have. And the industry split: some accepted the accusation and Hachette’s response, while others considered the punishment prejudicial because detection software is imperfect. John Degen, president of the Writers’ Union of Canada, told CBC: “I really would have preferred to see her publisher stand up for her and stand behind the work themselves because they trusted their own process.”
Notice what decided the outcome. Not a confession, not a court, not the publisher’s own editorial process. A score.
A month later the same script ran in literary fiction. Jamir Nazir’s story “The Serpent in the Grove” won the Commonwealth Short Story Prize for the Caribbean region and was published in Granta. Then, as The Conversation reported in May 2026, screenshots circulated on X of Pangram reports claiming 100 percent of the text was AI-authored. Granta’s publisher, Sigrid Rausing, even consulted Claude, an AI chatbot, which concluded the story was “almost certainly not produced unaided by a human”. Her final position: “perhaps we will never know.”
Two things in that story deserve attention. First, no formal determination was ever made; the doubt itself became the verdict in public. Second, the Commonwealth Foundation declined to start running detectors on submissions, saying that scanning unpublished work “would raise significant concerns surrounding consent and artistic ownership” and that a prize must operate on a principle of trust.
I keep a full evidence page on this site, with every figure linked to the study it came from: AI detector accuracy: what independent tests find. The short version:
The pattern that should worry fiction writers: the Authors Guild concluded that polished, edited prose by experienced human writers shares many characteristics with AI output. Clean writing is the false-positive profile. You can test your own draft for the exact lexical and rhythm patterns detectors look for using our free Signs of AI Writing Checker.
This is the single most common reason students and freelancers get blindsided by an unexpected AI flag. The answer depends on which half of Grammarly you clicked:
If you use Grammarly on academic papers or client work, stick to mechanical grammar fixes and rewrite awkward sentences in your own words rather than accepting one-click generative rewrites.
Because false accusations became so widespread in universities and freelance marketplaces, Grammarly shipped two features you should know how to use:
None of the following requires expensive software. It requires the discipline of not deleting things.
Ask, in writing, exactly which tool was used, on which text, and what score it produced. Any accuser acting in good faith can answer in one sentence. Then respond with provenance, not indignation: version history export, Grammarly Authorship report, dated drafts, correspondence. Point to the published false-positive evidence, ideally the Authors Guild test and the benchmark page that collects it. If the stakes are contractual, as they were for Ballard, that is the point to involve an agent, an authors’ organisation such as the Authors Guild or the Society of Authors, or a lawyer, before agreeing to anything.
A score is not evidence of anything on its own. The record of your work is. Build it now, while nobody is asking.
Standard Grammarly spelling, punctuation, and single-word grammar corrections do not trigger Turnitin's AI detector. However, using Grammarly's generative AI prompts or full-paragraph rewrites lowers sentence perplexity and burstiness, which can cause Turnitin or GPTZero to flag the passage as AI-modified.
Keep Google Docs or Microsoft Word version history enabled, save dated draft files, preserve your research notes, and enable Grammarly Authorship (free for Google Docs and Word), which records keystroke-by-keystroke origin and generates a shareable human-authorship report.
Grammarly Pro's built-in AI detector is a helpful pre-submission check to spot uniform, robotic phrasing before you turn in work, but like all statistical detectors it can produce false positives on heavily edited or formal human writing.
One practical writing tip, a new tool, and a fresh contest deadline, every Thursday. No spam, unsubscribe anytime.
While you wait for that email
Claude Code for Writers, $29
Go past templates and actually build with AI.