It depends on the detector. In a July 2026 test by Epoch AI, Pangram and GPTZero flagged none of 495 human-written passages, while Originality.ai flagged 19, including 16 of the 165 passages of fiction. None of the three can show how your book was made.
A detector's score is not proof, and nobody should decide anything about your book without other evidence.
This guide covers what researchers measured from 2023 to 2026, what a small error rate adds up to, which writers detectors have treated unfairly, and what to keep in case your book is flagged.
Why do AI detectors have such a bad reputation?
The reputation dates from 2023. That year Weber-Wulff and seven colleagues tested 14 detectors on 54 documents, and every one scored below 80% accuracy. GPTZero's false accusation rate in their test was 50%. The authors concluded: "Detection tools for AI-generated text do fail, they are neither accurate nor reliable".
OpenAI withdrew its own classifier in July 2023 because of its low accuracy. On OpenAI's own test set it had labelled 9% of human writing as AI. That month Ars Technica reported that GPTZero and ZeroGPT both scored part of the US Constitution as AI-generated.
If you've heard that detectors are junk, that's where it comes from, and for the tools of 2023 it was fair.
Have the best AI detectors become more accurate?
On human prose that nobody edited with AI, yes. The makers say so. GPTZero claims 99.9% accuracy, and Pangram says its version 4, announced in July 2026, flags human writing "only 0.0041% of the time, or roughly one false positive for every 24,000 documents".
Independent testers found the best detectors rarely wrong on human writing too.
| Test | Human writing tested | Pangram | GPTZero | Originality.ai |
|---|---|---|---|---|
| Epoch AI, July 2026 | 165 fiction passages from before 2022, about 500 words each | None flagged | None flagged | 16 flagged |
| Epoch AI, July 2026 | 495 passages of fiction, blogs and science | None flagged | None flagged | 19 flagged (3.8%) |
| Jabarian and Imas, September 2025 | 1,000 excerpts from novels published before 2000 | 0% | 0.45% | 0.25% |
| Russell, Karpinska and Iyyer, 2025 | 300 nonfiction articles under 1,000 words from eight US publications | 2.0% | 0.7% | Not tested |
So the belief that a good detector's flag means AI has real evidence behind it, for some detectors and some writing.
Look at the writing, though. Epoch's passages predate 2022, and Jabarian and Imas used novels from before 2000. A manuscript you finish this year can differ from both in age and in whether anyone edited it with an AI tool.
Does it matter which detector checks your book?
A great deal. In Epoch's test, Originality.ai flagged about one fiction passage in ten, and all 19 of its false positives, Epoch noted, "were flags of writing that predates ChatGPT." In Jabarian and Imas's test, the same maker's detector flagged 0.25% of old novels, though the texts differed and the detector version may have too.
In May 2026 the Authors Guild tested five detectors on ten of its own articles written before 2023. Pangram scored all ten at 0% AI. Originality.ai scored eight at 0% and two at 1%, and Grammarly scored eight at 0% and the other two at 7% and 9%. ZeroGPT scored them from about 5% to 76%, and Sidekicker.ai from 71% to 100%. It was a small test, and the Guild concluded that an accusation, or the use of these tools, shouldn't be a publisher's only reason to cancel a contract or pull a book.
The FTC challenged one maker's accuracy claim. In 2025 it settled allegations that Workado had advertised its AI detector as 98 percent accurate, and it said independent testing put the detector's accuracy on general content at 53 percent. The FTC's complaint alleges the detector's model was trained or fine-tuned to classify only academic writing.
If someone tells you a detector flagged your book, ask which detector, and which version.
What does a 1% error rate mean at scale?
Vanderbilt University did the sum in 2023, using the 1% false positive rate Turnitin claimed. "Vanderbilt submitted 75,000 papers to Turnitin in 2022." Had the detector been available then, the university wrote, around 750 of those papers could have been labelled as containing some AI-generated text when they didn't. Vanderbilt disabled the detector. The 750 is a projection from the maker's rate, not a count.
But the best detectors flagged no one, you say. In Epoch's test they didn't, and a test can only show a rate as small as its sample allows. From 0 flags in 495 passages, Epoch's own statistics put the true rate somewhere between 0% and 0.54%, at a 90% interval.
At the top of that range, a publisher checking 1,000 manuscripts by writers who used no AI would flag about five. At Pangram's claimed one in 24,000, it would most likely flag none. That claim comes from the maker's own test on a million documents, and the independent tests above used between 300 and about 2,000 human passages. The low number is still the maker's word.
A flag also says less than its error rate suggests. In one illustrative example from Schnell, a detector that catches 95% of AI essays and wrongly flags 5% of human ones checks 10,000 essays, 1% of them AI-generated. About 95 AI essays and about 495 human essays get flagged, so only about 16% of the flags are right.
Does editing your manuscript change the score?
In both directions. Karr and colleagues at Notre Dame ran light AI edits on published academic abstracts. Pangram flagged 64% to 80% of the lightly edited abstracts, and GPTZero flagged 38% to 49%.
Bloomberg Businessweek reported in 2024 that using Grammarly to improve an essay or make it sound academic turned work that passed as 100% human into work scored 100% AI, while its spelling and grammar suggestions made only a marginal difference. So a writer who uses an AI tool to polish their own sentences can score as AI.
The tools sold to hide AI use work the other way. After Karr's team ran AI-generated rewrites through a humanizer called Undetectable AI, "fewer than 4% of AI-labeled rewrites remain flagged." Studies disagree on how well the best detectors resist these tools. Jabarian and Imas found Pangram still caught text run through a different humanizer, StealthGPT, while GPTZero missed about half of it or more.
Has anyone tested AI detectors on novels?
Only on old ones. Epoch took its fiction from before 2022 so that "it is highly unlikely that any samples contain text from a language model", and Jabarian and Imas used novels from before 2000. That's a sound way to be sure the human writing is human, and it means neither test covered a novel written since ChatGPT.
The research for this guide found no independent test of current detectors on human fiction written after 2022 with known origin, or on novels edited with Grammarly or similar tools. Epoch says its own test "did not test the effectiveness of paraphrasing tools, manual editing, or other approaches".
The detectors also change under the tests. Epoch notes that detectors "can be updated by their makers without notice". Pangram published its version 4 results two weeks after Epoch published its test of version 3.3.2, and says it doesn't know how long a training set of writing from before 2022 will stay sufficient.
What do detector makers say a score means?
Several say it isn't proof. Originality.ai says its scores shouldn't be the only basis for disciplinary action, because even a low false positive rate is too high to rely on for that. Sapling says no current detector, its own included, should be used as a standalone check, and Grammarly says the same of its detector. GPTZero's FAQ says its results shouldn't be used to punish students, and Turnitin said in 2023 that the sentences it marks should start a conversation, not reach a conclusion.
The Authors Guild's guidance to publishers, from October 2026, says no publisher or agent should rely only on a detector to decide whether a book contains AI-generated text. It also says publishers shouldn't trust a detector's self-reported error rate, and that plain or formulaic prose isn't, by itself, a sign of AI use.
Are AI detectors biased against non-native English writers?
The detectors tested in 2023 were. For Pangram and GPTZero today, the evidence that they aren't comes only from their makers.
In 2023, Liang and colleagues at Stanford ran seven detectors over 91 TOEFL essays by non-native English speakers and 88 essays by American eighth-graders. On the American essays the authors found the detectors "near-perfect". On the TOEFL essays they averaged a 61.3% false positive rate, and at least one detector flagged 89 of the 91.
The authors traced the error to word choice. They argue that writers in a second language tend to use a narrower vocabulary and simpler sentences, and that detectors score that predictability as machine-like. When the researchers had ChatGPT improve the word choices in the same essays to sound more like a native speaker's, the false positive rate fell to 11.6%. They warned against using detectors "in evaluative or educational settings, particularly when assessing the work of non-native English speakers". Their samples were small, and the detectors were 2023 tools.
Later independent tests still find a gap. Al Ali, Helcl and Libovický ran Liang's essays through a commercial detector, Plagramme, in 2026. It flagged 23.1% of the TOEFL essays and none of the American ones. Stowe and colleagues tested 16 open and academic detectors on 41,743 American student essays and found English learners' essays more likely to be flagged by most of them, though "the magnitude of this effect is typically small."
Pangram reports a 0.012% false positive rate across 25,021 texts by English learners, Liang's essays included, and GPTZero reports no false positives on 10,402 student essays and on all three of Liang's sets. Both are the makers' own tests. The one independent re-run of Liang's essays, above, used a different detector.
Race is a separate question, and the evidence is thinner and points two ways. A 2024 Common Sense Media survey asked American teenagers whether a teacher had flagged their schoolwork as AI-generated when it wasn't. It found that "Black teens are more than twice as likely as White or Latino teens to say that teachers flagged their schoolwork as being created by generative AI when it was not (20% vs. 7% and 10%, respectively)." It can't separate a detector's error from a teacher's judgment, and it doesn't say which flags came from software.
Stowe's study measured how detectors scored essays by race. It found Black students' essays less likely to be flagged, Asian and Pacific Islander students' essays more likely, and non-White English learners' essays flagged more often than White English learners'. Those were open and academic research detectors, not commercial products. The research for this guide found no measurement of a commercial detector by race, and Innovation Village found no published test that runs African fiction known to be human-written through current detectors.
So the bias the evidence supports is against non-native English writing, and it varies by detector. If English is your second language, the Authors Guild's guidance tells publishers that a detector's flag on non-native English prose is weak evidence.
What can you show if your book is flagged?
Your drafts. In May 2026 a professor ran a Commonwealth Short Story Prize regional winner's story through Pangram, Brittle Paper reported, and a post said Pangram had flagged it as completely AI-generated. In June the Commonwealth Foundation said it was satisfied AI was not used to write the winning stories, after reviewing working drafts, time-stamped documents and notes. It added that detectors can be useful indicators but can't give conclusive evidence on their own.
In October, the Associated Press reported that, according to The Atlantic, Pangram had scored the Caine Prize winner Simbarashe Kundizeza's story "The Blue Fighter" as AI-generated, except for an excerpt of another author's poem. Kundizeza denied using AI. In November, Innovation Village reported, the prize concluded the story was human-written after interviewing him and the judges and having a forensic specialist analyze the metadata of several drafts. That review didn't test the detector.
You shouldn't have to prove anything. The Authors Guild says publishers shouldn't require authors to prove they wrote their books, and that an editor who suspects undisclosed AI use should first contact the author. It also fully supports authors who choose to preserve their drafts in case they're later accused. So keep them: old versions of the manuscript, your notes, files with their original dates, and any version history your writing app saves. Proselon saves a dated version history as you write, along with a record of which words you wrote and which are AI-generated.
Drafts are evidence, not proof. The research found no source that calls any one kind of record conclusive, and GPTZero says of its own writing-replay report that a large paste is a reason for a conversation, not proof.
What does Proselon record about where words came from?
Proselon saves a version of your project after two minutes without a change, and at least every five minutes while you keep typing with Proselon in front. Each version is listed with the time it was saved, and you can read any page as it was then.
It also keeps a record, page by page, of where each character came from, which the attribution page describes in full. Text typed or pasted in the editor is marked Human. Text that appears in a page while a co-writer is working in your project, without coming through the editor, is marked AI. Everything else is marked Unknown, including imported manuscripts and any words already in a page before Proselon first saw it. Proselon doesn't read your prose to decide any of this, and it keeps the record whether or not the colours are showing.
To see it, click the fingerprint button in the path bar above a page, and Proselon colours each word by where it came from. The bar at the bottom of the editor shows the split for the open page or folder. If a flag reaches you, send your dated drafts and your version history.
Proselon