AI Detector
Ask the AI whether a text reads like model output. You get a verdict in words, likely AI, likely human, mixed or uncertain, and the passages that led there, quoted from your own text so you can check them. There is no percentage, because there is no probability to report.
Nothing leaves this tab until you choose Check text. The text is then sent to the server once, read by the AI model, and not kept. Editing the text clears the verdict.
Paste text to check. Nothing is sent until you choose Check text.
What a verdict is, and is not
- No percentage. A score reads as the probability that a machine wrote the text, and nothing can know that. The verdict is a reading of register: how the prose is written, not who wrote it.
- Quoted evidence, verified. Every signal quotes a passage. The server checks that each quotation exists word for word in your text and drops any that does not; a decided verdict with no surviving quotation is rejected rather than shown.
- Nothing is kept. The text is sent once, read once, and the verdict is returned once. It is not stored, listed or retrievable, and it is never used to train a model.
- Limits. Short texts, code, tables and lists come back uncertain. Visitors can check up to 3,000 characters a few times a day per network. Signed-in accounts can check up to 20,000 characters, and each check spends the words it reads from the same balance a rewrite uses. A failed check spends nothing. A verdict is never proof of authorship.
The signals it reads
Every signal names one kind of evidence and quotes the passage that shows it. There are nine named kinds, plus one for anything else.
Usually points to a model
- Stock phrasing
- Set phrases that could open or close a text on almost any subject, such as a line about how important something is.
- Uniform rhythm
- Sentence after sentence of the same length and build, with no short one to break the pattern.
- Hedging
- Qualifiers stacked on a claim until it no longer commits to anything.
- Generic claim
- A statement that would be true of almost anything, with no name, example or figure behind it.
- Listicle structure
- Parallel points set side by side where an argument would connect them.
- Over-signposting
- Sentences that announce what the text is about to say, or recap what it has just said.
Usually points to a person
- Personal detail
- Something only the writer would know: what happened, to whom, and what it cost.
- Irregular rhythm
- Sentences of uneven length, fragments and asides that a template would smooth away.
- Specific reference
- A precise name, place, date or source that ties the text to one situation.
Either way
- Other
- Evidence that fits none of the nine names. It is quoted like the rest and can point either way.
Each signal also says which way its passage points, to a model or to a person. The grouping above is how each kind usually leans, not a rule: the direction is decided passage by passage.
The verdict follows the balance. Several independent signals pointing to a model give “Reads like AI-generated text”; personal, specific or irregular signals give “Reads like a person writing”; both kinds clearly present give “Reads like a mix”. A text under about 40 words, a text that is not prose, or too little evidence gives “Not enough to tell”, always at low confidence.
The habits themselves, with examples, are in The nine tells that make AI text obvious. Why the verdict carries no score is in What an AI detector score actually measures.
An example verdict
Illustrative: the sample and its verdict were written for this page to show the format. They are not the output of a real check, and your text gets its own reading.
Effective communication is the cornerstone of any successful team. Clear, consistent updates help everyone stay aligned, foster trust and drive meaningful results. Our own weekly update fell apart in March, when Marta went on leave and nobody else knew where the client notes lived. We spent two afternoons rebuilding them from old email threads. Now every update ends with one line: where the notes are.
The opening reads like stock workplace prose, while the account of the lost notes reads like someone describing what happened to their own team. Both kinds of signal are clearly present.
- Stock phrasingPoints to a model
“Effective communication is the cornerstone of any successful team.”
A general opening line that could start an article on almost any workplace topic.
- Generic claimPoints to a model
“help everyone stay aligned, foster trust and drive meaningful results”
Three broad benefits in a row, none tied to a name, an example or an outcome.
- Personal detailPoints to a person
“when Marta went on leave and nobody else knew where the client notes lived”
A first-hand account of what went wrong on one team, of the kind a template would not supply.
Every quotation above appears word for word in the sample, as the server requires of a real verdict. A mix describes the writing, not its origin: it does not mean part of the text was generated. As the limits above say, no verdict proves who wrote a text, this one included.
Frequently asked questions
Why does the detector give a verdict instead of a percentage?
A percentage reads as the probability that a machine wrote the text, and no detector can know that: the research on detector scores shows they are calibrated on samples that do not resemble your text, and they mislabel non-native writers most. So this tool returns four readings, likely AI, likely human, mixed or uncertain, with a confidence in words, and it must quote the passages that led there. A quotation you can check is worth more than a number you cannot.
What is sent to the server, and what is kept?
When you choose Check text, the text in the box is sent to the GenPolish server, which asks the AI model for a verdict and verifies that every quoted passage exists word for word in what you sent. Passages that do not are dropped, and a verdict without any surviving evidence is rejected. The verdict is returned once and stored nowhere: close the tab and it is gone, and nothing you paste is used to train a model.
How many checks can I run, and how long can the text be?
Visitors can check texts of up to 3,000 characters, a few times a day per network, separately from the free rewrites. Signed-in accounts can check up to 20,000 characters, and each check spends the words it reads from the same balance a rewrite uses: a 1,000-word text costs 1,000 words. Nothing is spent when a check fails, times out or returns no verdict. Very short texts, code, tables or lists come back uncertain rather than guessed at, and a verdict is a reading of register, never proof of who wrote something.
Related tools
Related reading
Want to rewrite what reads like a model?
Open the editor to review each change on its own. Your checked text is not transferred.