AI Text Detector
Paste any passage and see how machine-written it reads, with the reasoning shown. Runs entirely in your browser.
reads machine-written borderline reads human Sentence by sentence this is a weaker signal than the score — one sentence is a very small sample.
What pushed the score
Each bar is how hard that group of signals pushed, and which way: right argued for machine, left argued for human, and longer means it weighed more. These are the model’s own workings, not a second opinion bolted on afterwards.
Measured
How accurate is this, really
The model is tested the hard way: trained on eight text generators and then scored on a ninth it has never seen, which is what happens every time someone pastes something here. Across nine such rounds:
The 13% is the figure to hold onto. Roughly one human passage in eight is wrongly flagged, and those misses are not random — formal, technical and second-language writing take the worst of it, because a narrower idiom range looks statistically like a machine. Use the score to decide what to read more closely. Never use it to accuse anyone, and never as grounds for a grade, a rejection or a disciplinary decision.
How it works
The detector is a small statistical model running on your device. Nothing you paste is uploaded, stored or logged, which you can verify in your browser’s network tab.
It reads two things. The shape of the writing: how much sentence lengths vary, how wide the vocabulary is, how often contractions and first-person detail appear, and how densely stock phrasing shows up. Then word choice, from 1,200 learned patterns that separate machine prose from human prose.
Shape is weighted more heavily, deliberately. Word lists go stale as models change; how a passage is structured is far more stable.
How accurate it is
It was trained on 8,316 balanced passages from nine text generators and five kinds of writing, with human and machine samples drawn from the same sources so it cannot cheat by recognising a topic.
It was then tested the hard way: trained on eight generators and scored on a ninth it had never seen, repeated for all nine. That averaged 85% accuracy, catching 83% of machine-written passages while wrongly flagging 13% of human ones.
Where it gets things wrong
Roughly one human passage in eight is flagged, and not evenly. Formal academic prose, technical documentation, templated writing and English as a second language all score too high, because a narrow idiom range and even sentence rhythm are exactly what the model reads as machine-like. Under about 120 words, treat any result loosely.
There is no watermark in ordinary AI text. Every detector is inferring from style, and style is not proof of authorship. A high score is a reason to read something more closely. It is not evidence, and it should never decide a grade, a hire or a disciplinary case.
Frequently asked questions
Is the AI Text Detector free?
Yes, completely, with no account and no usage limit. There is no server doing the work, so there is nothing to pay for.
Do you store the text I paste?
No. It never leaves your browser — no upload, no API call, no logging.
How accurate is it?
About 85% on generators it has never seen, catching 83% of machine-written passages and wrongly flagging 13% of human ones. Useful, nowhere near good enough to accuse anyone with.
Why does it flag my own writing?
Most likely because it is formal, evenly paced, or written in English as a second language — the biggest source of false positives in every detector. The breakdown under the score shows which signals pushed it.
Can it tell which AI wrote something?
No. It estimates whether a passage reads as machine-written, not which model produced it.
How much text does it need?
Fifty words to return anything, about 200 before the result is worth much. The signals it reads are averages, and averages over four sentences are noise.
Does it work on languages other than English?
Not well. Both halves of the model were trained on English, so results elsewhere are unreliable.