Can Readers Tell If a Book Is AI-Written? What the Studies Show

Can readers tell if a book is AI-written? Studies of 2,587 readers say no better than chance, and our 13-model count shows the tells they do notice.

By Published November 18, 20264 min read

Try KudoNovel for free
  • No subscription needed
  • Unfiltered & uncensored
  • Story Bible that updates itself

Mostly not. In the two biggest tests on fiction, ordinary readers were no better than chance at picking the AI story: 1,682 adults in one study, 905 more in the follow-ups. What readers do notice is a handful of habits: the same names, stock phrases and punctuation. Those habits differ a lot by model, and they are the part you can fix.

What the studies found

Checked 18 November 2026. The main experiment is by Sydney Sears and Deena Skolnick Weisberg of Villanova University, published 5 August 2026 in Judgment and Decision Making (the paper). Each story ran about 1,000 words, a five-minute read, and the AI ones came from ChatGPT 4.0.

TestReadersResult
Study 1: rate stories told "human" or "ChatGPT"1,682AI stories scored higher on quality (1.54 vs 0.97) and absorption (1.42 vs 1.00)
Study 2: pick the human story424167 right, 39.4%, below chance
Study 3: same, new stories481250 right, 52.0%, no different from a coin

Source: Sears & Weisberg, 2026.

Two details matter more than the headline. Readers who were told a story was human rated it higher (quality 1.40 against 1.12), whatever it really was. So readers can't separate the two on the page, yet they mark down the word "ChatGPT".

Expert readers are harder to fool when the AI is only prompted. In Chakrabarty, Ginsburg and Dhillon's study (latest version 14 November 2026), MFA-trained readers rated a prompted imitation of 50 award-winning authors worse on quality (odds ratio 0.13), while general readers preferred it (1.82). The same paper found detectors flagged 97% of prompted outputs, and 3% once the model was fine-tuned on the author's books.

Why the two groups split: Marco, Gonzalo and Fresno (ACL Findings 2025; 1,471 stories, 101 annotators) found surface-focused readers, mostly non-experts, reward easy reading and varied wording, while holistic readers, mostly experts, reward how themes develop. A casual reader and a critic are scoring different things.

What this doesn't show: all of it is short fiction in a controlled setting. Nobody has tested whether a reader spots an AI novel after 300 pages.

The tells readers do notice

Detection is a bad bet; repetition is not. Over 300 pages, a name that appears in every book or a phrase that appears every chapter registers even with a reader who couldn't say why. We counted these in the AI Slop Index: 102 scenes from each of 13 models, October 2026.

HabitLowest modelHighest model
Stock words per 1,000 wordsQwen3.8 Max, 12.3Llama 4 Maverick, 52.6
Em dashes per 1,000 wordsClaude Opus 5.5, 0.04Kimi K3, 9.68
"Not X, but Y" per 1,000 wordsGPT-6.1 Sol, 0.64Qwen3.8 Max, 3.21
Most common character nameMara, in 7 of 13 modelsElara, Elias, Kaelen, Lena in the rest

Source: KudoNovel's AI Slop Index, October 2026.

Mara was the most-used name in seven models, and in GPT-6.1 Sol she turned up in 23% of scenes. A reader who has met Mara in three AI books has stopped being fooled, with no detector involved. The model-by-model word list is in AI writing words to avoid.

Where the other options fall short

  • Detectors don't settle it. In our own run, one free detector flagged 7 of 60 classic-novel passages as AI at 60 words (72 passages).
  • Prompting a chatbot gives the 97%-detectable output from the Chakrabarty study. Fine-tuning fixed it for $81 an author, but that needs the author's whole catalogue, which you don't have as a first-time novelist.
  • Disclosure rules are the real risk. Platforms and stores ask whether AI wrote the text (KDP's AI disclosure), and the Villanova label effect says the answer costs you readers.

A checklist that removes the tells

  1. Search the draft for names you didn't choose. Rename anyone called Mara, Elara, Elias or Kaelen unless you meant it.
  2. Count em dashes in a chapter and cut to your own habit.
  3. Hunt "not X, but Y" and swap each for the plain claim.
  4. Read one paragraph aloud. If three sentences open the same way, rewrite one.
  5. Keep the rough paragraph you started from, and compare.

Start from your words, not a blank prompt

The tells come from asking an AI to invent. In KudoNovel you type your own rough paragraph and press Expand: the AI writes it out from your words, in your story's language, with the names and events you gave it, and your story bible supplies the rest. Steer takes one line to pull a paragraph back ("less formal, more her"), and every change is a version you can Undo, so your original line is never lost. There is no filter of its own on what you write. Open a page and try it on a scene you've already drafted; one free credit comes with the account.