AI text detector
Two readings of how much a text looks machine-written.
0 words, 0 sentences.
The text is processed by a language model through our server. We do not store it and do not train on it, but it is better not to paste anything sensitive here. Free gives you 5 uses a day and 3,000 characters per request.
This is not proof. No tool, paid ones included, can reliably tell who wrote a finished text, and the model gets it wrong too. Carefully written human prose scores high; a model asked for a casual voice scores low. Do not use the result to accuse a student or an employee: it is a hint about where to look, not a verdict.
- Runs on our server
- Free, no sign-up
More about AI text detector
Two views of the same text
The tool gives you two independent readings. A quick analysis runs in your browser as you type, counting visible marks: how much sentence length varies, how often phrases typical of models appear, how the punctuation behaves. On top of that, a language model gives its own opinion with reasons and a counter-argument.
The browser analysis weights structure most heavily. A person alternates a clipped sentence with a long one; a model holds a steady rhythm. Phrase lists can only add suspicion, never clear a text, because current models have largely stopped using the obvious clichés.
Both readings are shown with their reasoning, so you can see what drove the number instead of taking a bare score on trust.
Frequently asked questions
Can I use this as evidence?
No, and no other tool can either, paid ones included. Careful human writing scores high, and a model asked for a casual voice scores low. Treat the result as a hint about where to look, never as a verdict about a student or an employee.
Why do the two readings disagree?
They look at different things. The browser analysis measures surface statistics; the model reads for meaning and for the kind of detail a person tends to include. Disagreement is itself informative: it usually means the text is genuinely borderline.
How long should the text be?
At least a hundred and twenty words for the local score to mean anything. Below that, both readings say so rather than pretending to certainty.
Does an edited AI text get caught?
Often not. A few minutes of human editing removes most of the marks both readings rely on, which is one more reason not to treat any score as conclusive.