Not reliably. AI detectors catch long, unedited passages from some models, miss a good share from others, and get worse on short excerpts, where they also start flagging human prose. In our run on 14 November 2026, a free detector caught all six scenes Llama 3.1 wrote but only two of six from Mistral Small, and at 60 words it called 7 of 60 passages from classic novels AI-written. GPTZero, Turnitin and Pangram all say a score is not proof, and the benchmark most vendors cite doesn't test novel prose at all.
We ran two AI detectors on fiction: the results
Checked 14 November 2026: we took 60 passages of human fiction, ten from each of six public-domain novels on Project Gutenberg (Pride and Prejudice, Jane Eyre, Frankenstein, Dracula, The Hound of the Baskervilles, Treasure Island), whole paragraphs from evenly spaced points, 200 to 300 words each. Then we had two models write 12 scenes in the same genres ("a proud heroine refuses a proposal at a country ball", "a detective examines footprints on a moor at night"…), six each from Llama 3.1 8B and Mistral Small 3.1 24B on Cloudflare Workers AI. Both sets went through two open-source detectors anyone can download: OpenAI's RoBERTa detector and ChatGPT-RoBERTa. A passage counted as flagged when the detector gave it more than a 50% chance of being AI. Then we cut every text to its first 60 words and ran it again.
| Detector | Human passages flagged (~240 words) | AI scenes caught (~250 words) | Human flagged at 60 words | AI caught at 60 words |
|---|---|---|---|---|
| RoBERTa (OpenAI) | 1 of 60 | 8 of 12 (Llama 6/6, Mistral 2/6) | 7 of 60 | 4 of 12 |
| ChatGPT-RoBERTa | 0 of 60 | 10 of 12 (Llama 6/6, Mistral 4/6) | 1 of 60 | 5 of 12 |
Source: KudoNovel's run, 14 November 2026. Every text and score is in the run's data file.
The model matters: Llama's ballroom and moor scenes scored over 99% AI, while one of Mistral's, a ballroom scene that opens "In the resplendent ballroom of Lord Harrington's estate", scored 0.2% on the RoBERTa detector. Length matters both ways: at 60 words the RoBERTa detector's false flags on Austen, Brontë and Stoker went from one to seven, and it caught half as many AI scenes. And the one full-length false flag, a stretch of Van Helsing talking to Dr Seward in Dracula, scored 50.6%, barely over the line.
What this run doesn't show: classic novels may sit in these detectors' training data as human text, which flatters their false-positive rate, and modern human fiction wasn't tested, and the two models were small open ones, not ChatGPT or Claude. The commercial tools below are better trained; RAID shows the same two weak spots, other models and other genres.
What the benchmarks say about AI detectors on fiction
The study vendors quote most is RAID (Dugan et al., ACL 2024): over 6 million generations, 11 models, 12 detectors. It has no novel prose. Its closest domains are plot summaries of books, poems and Reddit posts. Here is how each detector did there, set so that it wrongly accuses only 5% of human texts (Table 14 of the paper):
| Detector | Reddit posts | Book plot summaries | Poetry | News, for comparison |
|---|---|---|---|---|
| Originality | 85.0% | 90.4% | 75.1% | 88.4% |
| Binoculars (open source) | 79.4% | 83.7% | 81.0% | 80.7% |
| Winston | 68.9% | 70.7% | 64.3% | 72.4% |
| ZeroGPT | 65.1% | 73.3% | 50.0% | 72.2% |
| GPTZero | 57.0% | 71.4% | 69.5% | 58.1% |
Share of AI texts caught at a 5% false-positive rate, RAID, 2024.
So even tuned to accuse one human in twenty, the best tool on poetry, Binoculars, still missed about one AI poem in five, and GPTZero missed 43% of the AI Reddit posts. And a tool isn't tuned per genre in real use. With one threshold for everything, the paper's Table 8 shows the false-positive rate swinging by domain: GPTZero and Winston each wrongly flagged 13% of the human book summaries while holding a 5% average, and ZeroGPT couldn't get down to 5% at all (16.9% overall, 48% on Wikipedia).
Writers working in a second language carry the most risk. Stanford researchers (Liang et al., 2023) found seven detectors flagged 61.22% of TOEFL essays by non-native English writers as AI on average, and 89 of the 91 essays were flagged by at least one.
What the detectors claim about themselves
| Detector | Its own accuracy claim | What holds it back on a novel |
|---|---|---|
| GPTZero | "99% accuracy rate and a 1% false positive rate" (page dated 14 Oct 2026) | Its own FAQ answers "No" to whether a detector can prove someone cheated; in RAID it caught 57% of AI Reddit posts at a 5% false-positive rate |
| Pangram | False positives of 0.009% on short stories, 0.003% on books, 0.05% on poems (15 Sep 2025) | Its own figures, on its own data; its CEO told NPR it can make mistakes, most of all on short text |
| Turnitin | Under 1% false positives per document with over 20% AI writing; about 4% per highlighted sentence | Needs 300 words; an academic tool, tuned on student writing |
| Originality.ai | Best overall in RAID's main table (85.0% caught) | 13% false positives on human Wikipedia text at RAID's single threshold; 75.1% on poetry |
| ZeroGPT | No published false-positive rate found | Scored ten pre-2022 Authors Guild articles anywhere from 5.3% to 76.3% AI |
| OpenAI's classifier | Withdrawn 20 July 2023 for "low rate of accuracy" | Caught 26% of AI text and flagged 9% of human text, by OpenAI's own figures (Search Engine Land) |
The Authors Guild ran its own check on 26 May 2026: ten of its news articles from 2022 or earlier, written before ChatGPT, through five tools. Pangram scored all ten 0% AI, Originality 0–1%, Grammarly 0–9%, ZeroGPT 5.3–76.3%, and Sidekicker.ai flagged every one, 71% to 100%.
Do publishers and magazines use AI detectors?
Some decide without one. Hachette pulled Mia Ballard's horror novel Shy Girl in March 2026 after an investigation concluded large portions were AI-generated, following Reddit threads and a YouTube video (AV Club, 20 March 2026). Ballard said an associate who edited the self-published version may have used AI. Hachette now asks authors to disclose any AI use.
Clarkesworld "will not consider any submissions translated, written, developed, or assisted by these tools" and may ban anyone who tries, but its editor, Neil Clarke, told NPR (31 October 2026) he doesn't use detectors himself. In the same report, ghostwriter Sally McGraw lost a book-proposal job after several detectors flagged a sample she says she wrote.
The Authors Guild's line is plain: "Publishers should never cancel a contract or pull a book on the basis of accusations or use of these tools alone." Amazon's rule is about disclosure, not detection; our KDP AI content policy guide quotes it, and which story sites allow AI fiction covers Wattpad, AO3 and Royal Road.
Can a human-written story be flagged as AI?
Yes. Our run flagged a passage of Dracula; RAID's single-threshold test wrongly flagged 13% of human book summaries on two commercial tools; the Authors Guild's pre-ChatGPT articles hit 100% on one. Short samples, plain or formal prose, and writing by non-native speakers raise the odds, because many detectors score how predictable each next word is. University of Pennsylvania's Chris Callison-Burch told NPR that models also find their own training data predictable, which is how well-known human text gets flagged.
Are romance novels flagged as AI more often?
No test we found measures romance on its own. In our run the two romances held up at full length (0 of 20 passages flagged for Pride and Prejudice and Jane Eyre) and lost one passage of Austen at 60 words; Treasure Island lost three. What raises the risk is the sample, not the genre: a blurb, a short excerpt, or a love scene pasted alone gives a detector little text and a lot of familiar phrasing.
What to do if your writing is flagged as AI
- Ask which tool and what score. A 50% score is a guess; a 300-word excerpt is below what Turnitin itself accepts.
- Show your process. A student at UC Davis was cleared in 2023 after he handed over his document history (Authors Alliance, 20 July 2026). Drafts, dated versions and notes beat any score.
- Quote the vendors back. GPTZero and Turnitin both say their scores are one signal, not proof.
- Get help. The Authors Guild asks members who have been falsely accused to contact it for legal assistance, and its line on publishers (above) is worth quoting to yours.
- Don't run your chapters through a "humanizer": it rewrites your voice to dodge a test that can't prove anything either way.
For the prose itself, how to write fiction that doesn't sound like AI goes through the stock phrases readers spot.
Where KudoNovel fits
In KudoNovel you type each paragraph yourself, a line or a rough passage, and Expand writes it out around your words and events, keeping them. Every change the AI makes is a version, so the paragraph as you typed it is still there in Versions, beside the expanded one and every steer after it: a record of what you wrote and what the AI added, the kind of history that cleared the UC Davis student. The story bible keeps names and facts straight as the book grows, so you're not pasting chapters into a chatbot to remember them.
It's a browser app; you get one free credit to try it, then one-off credit packs with no subscription. Write your next chapter in KudoNovel, and if you publish on KDP, disclose what the AI generated, as Amazon asks.