The question keeps coming up, and from two completely different directions. Some authors used AI tools and want to know whether it shows. Others never touched one, yet still fear their text will be rejected because it “looks like AI.” Both fears are reasonable, because more publishers, agents and contests now run submissions through an AI-content scanner before a human reads them.
The bad news is that there is no reliable, definitive test that will say once and for all “a human wrote this.” The good news is that once you understand what these detectors actually measure and how unreliable they can be, you stop fearing them and start doing the one thing that genuinely helps: writing so the text sounds like you.
Short answer: AI detectors do not detect the “truth” about authorship, they detect the statistical shape of prose, and they get it wrong often, especially on clean, smooth, slightly formal text. A “this looks like AI” verdict is not proof. Instead of hunting for a magic detector, scan the text yourself, find the passages that read machine-like, and rewrite them in your own voice.
What “looks like AI” really means
A detector does not understand your story. It does not know whether your protagonist is convincing or whether the ending makes sense. It measures the statistical shape of the prose: how predictable each next word is, how even the sentence rhythm is, how often you reach for the same constructions, how little surprise there is across a passage. Machine-generated text tends to be smooth, evenly paced and low in surprise, so the detector scores that smoothness as a signal.
Two things follow. First, the detector judges style, not honesty. That is why your own careful, correct prose can score badly even though you wrote every word. Second, the things people take for “tells of AI text” (the models’ favourite words, even lists, formulaic transitions) are stylistic observations, not a hard test. Sometimes they are right, sometimes misleading, and they shift with every model release. Treat them as circumstantial clues, not a verdict.
Why detectors get it wrong more often than you think
This is not a cautious hypothesis, it is the result of a peer-reviewed study. A team at Stanford (Liang et al., 2023, the journal Patterns from Cell Press) ran seven popular AI detectors over essays written by people for whom English is not a first language. The result should cool anyone who treats these tools as an oracle: the detectors falsely flagged those human essays as AI-generated 61.3% of the time on average. Worse, 89 of 91 essays, that is 97.8%, were flagged as AI by at least one of the seven detectors. By comparison, on text from native English writers the false-alarm rate was many times lower.
The conclusion is twofold. First, these tools have a real, high false-positive rate. Second, they are biased: they hurt people who write more simply, more evenly, more “by the book” the most, which often means exactly those still learning the craft or writing in a language that is not their first. If a detector ever flags your text, remember that number. A single tool’s verdict is not proof, it is an estimate with a large margin of error.
How to check your own text sensibly
Since any single detector can be unreliable, the method matters, not faith in one tool.
- Scan the whole thing first for a baseline. A whole-book number tells you whether you have a local problem or an even one.
- Then go chapter by chapter. Machine-like passages usually cluster. A chapter drafted in a rush, or with a tool’s help, tends to stand out from the rest.
- Read the lowest-scoring sentences, not just the number. A good detector points you at specific lines. Those lines are your edit list, not the percentage at the top of the screen.
- Rewrite them in your own words. Break the even rhythm, cut the filler, swap the generic phrasing for something specific to your story. Then re-scan and check the passage now reads like the rest of the book.
- Do not race the detector. The goal is not to “beat” the tool, it is to make sure the text genuinely sounds like you before someone else decides whether it does.
What exactly do you change when a passage reads machine-like? Usually the same things you would fix in editing anyway: vary sentence length, drop the passive constructions and generic verbs, add a concrete detail only your character would know. That is not gaming a detector, it is simply better prose. We cover this separately in the piece on passive voice in prose.
A detector for novelists, not for essays
Most AI detectors you find online were built for essays, articles and homework. They are generic, often score per word, and treat a 90,000-word novel as one giant blob of text to paste in. They were not designed for fiction, for chapter structure, or for the way a novelist actually revises.
Writing tools have the opposite gap: the generation-first ones ship no AI detector at all, for obvious reasons. So most authors are stuck choosing between a tool that helps them write and a separate generic checker that was never meant for books. Vellam has the detector built in: it scores how machine-like the prose reads, chapter by chapter or across the whole book, and highlights the lowest-scoring sentences so you can rewrite them in your own words. It is an AI detector built specifically for writers, not a generic essay scanner, which is exactly why it works best for novelists. Alongside it runs a free, rule-based prose layer that helps you fix those sentences. This is reassurance for you, the author, not a gate someone else holds. For the wider picture of scanning before you submit, see the companion piece on checking your novel for AI before submitting.
Keep one thing in mind: the built-in detector is an estimate too, not an oracle. The score alone settles nothing. The value is not the percentage, it is the flagged sentences you can work on.
What not to do
Finally, three traps that are easy to fall into once you start fearing “looking like AI.”
Do not rewrite the whole book to chase one score
If a detector shows a high percentage but the text is yours, the problem is the tool, not the text. Read the flagged sentences and judge them yourself.
Do not trust a single detector
With a 61% false-alarm rate on human prose, one verdict settles nothing. If anything, compare a few and read the specific passages.
Do not confuse smoothness with guilt
A clean, even sentence is not proof of a machine. Sometimes it is just a clean sentence. Change it because it reads impersonal, not because a tool lit up red.
Frequently asked questions
Are AI detectors reliable?
Only partly. A peer-reviewed 2023 Stanford study found popular detectors falsely flagged human essays as AI 61.3% of the time on average, with 97.8% of texts flagged by at least one of seven detectors. A single tool’s verdict is an estimate with a large margin of error, not proof of authorship.
How can you tell if a text reads as AI-written?
Detectors respond to the statistical shape of prose: high word predictability, even sentence rhythm, repeated constructions and little surprise. What people call “tells of AI text” are stylistic observations, not a hard test, and they change with every model release. Treat them as clues, and base your decision on whether a passage sounds like you.
Can my text be flagged as AI even though I wrote it myself?
Yes, and it happens often. Detectors hurt people who write more simply, more evenly and more formally the most, including those learning the craft and writing in a language that is not their first. A false flag does not mean you did something wrong, it means the tool is wrong.
How do I fix a passage that reads machine-like?
Usually the same way you would fix it in editing: vary sentence length, drop generic verbs and passive constructions, add a concrete detail from your own story and break the even rhythm. The goal is not to “beat the detector” but to make the sentence sound like the rest of your book.