AI tell · grade ungraded
How to spot AI writing in fiction, without a checklist
Fiction defeats most AI-tell checklists because a novel has nothing for them to grab onto: no headings, no bold bullet labels, rarely even the em-dash density that flags a blog post. The tells that hold up in fiction come from a different level of evidence entirely, editors reading hundreds of submissions in a row, not one manuscript at a time. Clarkesworld's editor, Neil Clarke, closed submissions in 2023 after describing a flood of stories that were technically competent and interchangeable, several sharing the exact title 'The Last Hope.' Separately, writers on a publishing forum flagged query letters that reached for the same pitch template: a novel described as blending the storytelling of one well-known book with the emotional honesty of another, in near-identical phrasing across different submissions. Neither tell can be seen by scanning a single manuscript. Both require comparing it to a pile.
What it is
Two related patterns, both corpus-level rather than sentence-level. The first is slush-pile sameness: submissions that are each competently written but interchangeable with dozens of others, the same beats, the same stakes, sometimes the same title. The second is the formulaic comp-title pitch, a query letter that describes a novel as a blend of two well-known books in near-identical phrasing across many different submissions. Neither shows up as a flagged word or a flagged sentence. Both show up only when someone reads enough submissions to notice the repetition.
How often it shows up
This page runs on our social-sentiment research (research/reports/social-sentiment-tells.json), not our own model logs; we did not build a corpus of AI-written short stories to test corpus-level sameness directly. Slush-pile sameness carries the strongest sentiment score in that entire research set, a 5 out of 5, corroborated across four separate sources: a Reddit thread, a Horror Writers Association Facebook post, Neil Clarke's own account of closing Clarkesworld submissions, and an NPR report on the same closure. The comp-title pitch pattern is weaker evidence: one source type, a single publishing-forum thread, with a sentiment score of 4 out of 5, and the thread itself notes that writers can converge on similar comps honestly.
Why readers flag it
An editor does not reject any single one of these stories for what is on the page. They reject it for what a hundred other stories that week also put on the page. A story that would read as perfectly fine on its own becomes a flag only in context, next to nine others with the same structure and, in Clarke's account, sometimes the literal same title. The comp-title pitch works the same way: any one comparison to two well-known books is a normal, even useful, pitch technique. A stack of query letters using the identical sentence template is what actually reads as generated.
The fix
underslop edits a single document. It cannot see the slush pile a story is competing in, and it says so plainly rather than pretending otherwise. What a writer can do alone is handle the two things a corpus comparison would catch: give the premise a detail specific enough that it could not describe the next story in the pile, and write a pitch line in your own words instead of the two-books-blended template, even when the comparison itself is accurate.
Examples
This novel blends the rich, dual-timeline storytelling of THE NIGHT CIRCUS with the raw emotional honesty of A LITTLE LIFE, exploring themes of memory, grief, and found family.
The novel moves between two timelines in the way *The Night Circus* does, and it carries a similar emotional weight—though here the register is closer to the blunt, unsparing honesty of *A Little Life*. It's a book about memory, grief, and the people you piece together a life around when your own family isn't one.
The last ember of the old world flickered in her hands as she stood at the edge of everything she had ever known, ready to become the hero the prophecy had always promised.
She held the last ember of the old world in her hands and stood at the edge of everything she had known, ready to become the hero the prophecy had always promised.
Great question about your query letter! Here's a draft: this novel isn't just a fantasy epic, it's a meditation on grief, love, and the price of power. Let me know if you'd like me to revise anything.
Here's a draft for your query letter: This novel isn't a straight fantasy epic—it's a meditation on grief, love, and the price of power.
Questions
How can you tell if a novel is AI-written?
From a single manuscript, it is genuinely hard, fiction has no headings or bullet lists to flag. The clearest signal editors have described is comparative: reading many submissions in a row and noticing that dozens share the same premise, the same beats, sometimes the same title.
How do editors detect AI-written story submissions?
By volume, not by sentence. Clarkesworld's editor described competent but interchangeable submissions arriving fast enough to overwhelm the slush pile, several with the literal title 'The Last Hope.' No line-level checklist produces that observation; only comparing one submission to the hundred around it does.
What are the signs of AI writing in fiction specifically?
Two, and both require a pile to compare against rather than one page: interchangeable stories that are individually competent but collectively generic, and query letters that lean on the same comp-title sentence template, phrased almost identically across different writers.
Can a tool like underslop check a novel manuscript for AI tells?
It can check sentence-level and paragraph-level patterns in whatever you paste in, the same tells covered across this site. It cannot see the slush pile your manuscript is sitting in, which is where the fiction-specific tells actually live. That is a real limit, not a caveat to skip past.
Is a formulaic comp-title pitch proof a query letter was AI-written?
Not from one letter. It becomes suspicious only in aggregate, when an agent sees the same sentence structure across many submissions in a short window. The evidence for this one comes from a single publishing forum thread, and even that thread's own commenters pointed out that real writers naturally converge on similar comps too.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.