Rhythm & burstiness
Variation in sentence and paragraph length. People write in bursts — a fragment, then a long winding clause. Default model output is far more even.
AI detector / AI content checker
AI Detector
Check any text for the writing patterns typical of ChatGPT, Claude, Gemini and other models. Every signal behind the score is shown, and nothing you paste leaves your browser.
Client-side · deterministic · no account
Click to upload or drop a document
.txt · .docx · .pdf · .rtf — parsed locally
0 words
0 characters
Measuring text patterns…
Estimated AI-pattern score—%
—%Sentences scoring high
—%Mid-range
—%Sentences scoring low
Each bar shows how far one signal leans toward machine-typical writing. These are stylistic measurements, not evidence of authorship — a careful human writer can score high on all of them.
Want this to read less like a machine?
Rework the cadence and vocabulary with 0Humanizer.
This text shows few of the patterns this tool looks for. Nothing here needs changing on our account.
Sentence-level analysis
Scores high
Mid-range
Scores low
Read this before you trust a number. No AI detector — this one included — can prove who wrote a text. It measures writing style, not origin. Formal, academic and non-native English writing scores higher regardless of author, and edited AI text scores lower. Never use a detection score as the basis for an accusation or for any decision about someone's grades, job or reputation.
Method
How this detector works
Nine stylistic measurements, combined into one weighted score. No model, no black box, no randomness — the same text always returns the same number.
Vocabulary fingerprint
Density of roughly 70 phrases and connectives that language models over-produce — "furthermore", "delve", "it is important to note" — plus lexical diversity adjusted for text length.
Voice markers
Contractions, first and second person, passive constructions, punctuation range, and how often sentences start the same way. Assistant-default prose is short on all of these.
Where it goes wrong — worth knowing
Formal academic and legal writing scores high even when entirely human-written, because that register genuinely is even, impersonal and connective-heavy.
Writing by non-native English speakers scores higher on average. This is a documented bias in every detector, not a quirk of this one.
AI text edited by a person often lands mid-range or low. A low score is not evidence that a human wrote something.
Under about 120 words there is not enough signal to say much, so the tool labels the result low-confidence rather than pretending otherwise.
Next step
Scored high? Here's what actually helps.
A high score means the writing reads as generic — which is a writing problem worth fixing on its own terms, detector or no detector.
Put something specific in
A number, a date, a name, something that happened to you. Generic prose is what scores high; detail breaks the pattern and makes the piece better to read.
Vary the rhythm
Mix short sentences with long ones. Machine prose is metronomic. Break it. Then let a sentence run on a little further than the reader expects, the way people actually talk.
Cut the filler connectives
"Furthermore", "moreover", "it is important to note". They add nothing and they are the strongest single signal here. Delete them and the sentence usually improves.
FAQ
AI detector questions
How accurate is this AI detector?
Accurate enough to tell obviously generic machine prose from lively personal writing, and not much more than that. It measures style, not origin. Every detector on the market — including paid ones with confident-looking dashboards — has the same ceiling, because the underlying problem is genuinely hard and getting harder as models improve.
Can I use these results as evidence?
No. Please don't. Students have been failed and staff disciplined on detector output that was wrong, and the harm falls hardest on non-native speakers and on people who write formally. If you are assessing someone's work, talk to them about it — a five-minute conversation about how they wrote something tells you more than any score here.
Why does it flag text I wrote myself?
Because formal writing genuinely shares its statistical shape with model output: even sentence lengths, few contractions, no first person, heavy connectives. The tool cannot tell a careful human from a careful machine, and it doesn't claim to. Open the signal breakdown — it will usually show exactly which measurement pushed the score up.
Does a low score mean my text will pass other detectors?
Not necessarily. Other tools use different signals — some use trained classifiers rather than rules — so scores don't transfer. We can't promise a result on a third-party system and won't pretend to.
What does the score actually mean?
It's a weighted average of nine stylistic measurements, each scaled 0 to 1. Roughly: under 35% the writing has clear human texture, 35–65% is genuinely ambiguous, above 65% it reads as machine-typical. The same text always returns the same number — there is no randomness in the calculation.
What file formats can I upload?
.txt, .docx, .pdf and .rtf, parsed in your browser. Legacy .doc isn't supported — save as .docx first. Scanned PDFs without a text layer won't work, since there is nothing to read.
Is my text stored or used for training?
No. The whole analysis is JavaScript running on your machine. Nothing is sent to a server, so there is nothing to store, log or train on. You can confirm it by opening your browser's network tab and running a scan.