Shy Girl, AI Slop Books and the Reader-Trust Crisis

For years the argument about AI and fiction was hypothetical. In 2026 it started showing up in real publishing decisions. On March 19, The New York Times reported that it had given Hachette Book Group evidence that large parts of Mia Ballard’s horror novel Shy Girl were AI-generated. According to coverage of that report, Hachette ran what it called a thorough and lengthy review, then canceled the planned US release through its Orbit imprint and pulled the book from UK shelves. Fortune later noted it was also removed from Hachette’s website and Amazon.

Ballard’s position matters and deserves to be stated plainly: she denies using AI to write the book. She first self-published Shy Girl in February 2025, and she has said an acquaintance she hired to edit and format that earlier self-published version used AI tools. In a statement quoted by the Times, she said her name was ruined "for something I didn’t even personally do," and declined to give details because she is pursuing legal action. Hachette has said it requires authors to disclose AI use, but it has not published the specifics of its findings. How the text came to read the way it does has not been publicly settled.

What is clear is how the suspicion started. In mid-2025 a Reddit user who said they worked as a book editor pointed to stylistic tics, such as nouns almost always carrying an adjective, actions described through similes and habitual lists of three. The discussion moved to BookTok, Instagram and YouTube before a major newspaper took it up. Readers, not software, raised the alarm first.

Shy Girl drew attention because a big publisher was involved. The larger shift is happening in self-publishing. On September 14, 2026, Fortune covered a study led by Tuhin Chakrabarty of Stony Brook University and Paramveer Dhillon of the University of Michigan, posted to arXiv in July as Generative AI floods and dilutes the market for books. It has not been peer reviewed, and its numbers should be read with that in mind.

The researchers ran the full text of 14,419 randomly sampled self-published genre-fiction ebooks sold on Amazon between 2023 and 2026 through the Pangram detector. About 20 percent, 2,880 books, had more than a quarter of their prose flagged as AI-written, and nearly 1,000 were flagged at over 90 percent. Matched to sales data, those books made up a smaller share of sales than of the catalog, but their share of sales and of top-ranked positions grew over time. Fortune reported that across three years the number of selling titles per quarter rose 19.2-fold while quarterly revenue rose 8.9-fold, so the average book earned less.

The authors describe this as AI reshaping a market "through scale rather than quality." That is the part that should worry readers as well as writers: the problem is less that machine-written books are winning on merit than that there are so many of them that good work gets harder to find.

The study’s design is careful, but its foundation is a classifier, and AI detectors remain contested. Pangram’s developers report a very low false-positive rate; that figure comes from the company, not an independent audit. The Decoder, summarizing the paper, put the limit simply: a detector gives an estimate, not a source. The study’s own comparisons are associational and do not prove any individual book was machine-written.

Earlier research shows why caution is warranted. A 2023 Stanford study of seven popular detectors found they flagged 61 percent of essays written by non-native English speakers as AI-generated, while scoring essays by US-born students almost perfectly. Tools have improved since, but a score is not a verdict, and no reader should treat one as proof against a named author. The Shy Girl story shows how much is at stake for the person on the receiving end.

Without a reliable machine test, readers are left with judgment. A few signals are worth noticing, as reasons to look closer rather than evidence. Volume is one: Amazon’s own KDP help pages now cap new titles at two per book format each week, down from three a day in 2023, and a pen name releasing books faster than anyone could plausibly draft them is worth a second look. Sameness is another: prose where every sentence has the same rhythm, every emotion is named rather than shown, and the same tics repeat chapter after chapter.

Then there is the thin texture problem. Machine-heavy fiction often summarizes scenes instead of staging them, and characters forget what happened two chapters ago. Reading a sample before buying, checking whether an author has a history of revisions and replies to readers, and noticing whether reviews mention specific scenes or only general praise will tell you more than a score.

One limit to keep in mind: Amazon requires self-publishers to disclose AI-generated content to the company, but according to The Decoder that information is not shown to customers. Our explainer on KDP’s AI disclosure rules covers what writers are asked to declare.

Much of this debate collapses two different things. Amazon’s KDP guidelines draw the line usefully: content created by an AI tool counts as AI-generated even if you edited it heavily afterward, while text you wrote yourself and then refined, error-checked or brainstormed with AI counts as AI-assisted, which KDP does not ask you to disclose. In the assisted case the writer drives the story and every line passes through human judgment. The distinction matters for quality, for honesty with readers and for rights; our piece on who owns an AI-assisted story goes through the copyright side. This is an overview, not legal advice.

Human editing is what turns a draft into a book, whatever tools were involved. Disclosure is what lets readers decide for themselves. The Shy Girl case also shows the risk of outsourced help: if someone else touches a manuscript, the author should know what tools they used, because the author’s name is the one on the cover.

ShortFling is a fiction platform, not an AI chat or companion app, so this question sits close to home. Every writer who publishes signs a creator agreement confirming the work is their own original work and that all characters are adults in a purely fictional story. Pressing publish submits a story for review, and it goes public only after it is approved.

The AI editor in /write is built as an assistant to the writer, free within a daily quota, for drafting help and revision rather than replacing the author. We do not claim to run an AI detector on submissions, and given the evidence above we would not present a score as proof if we did. If you want to see what that balance produces, browse the library or the adult NSFW stories, and if you write, our guide to slow-burn romance with an AI editor shows the assisted approach in practice.