Check whether writing came from a person or a model. Scored passage by passage, on your own device.
The text detector is not published yet.
At least 50 words · any number of files · PDF, Word and text · nothing uploaded
Reading the text…
0decision line 65100
There is no upload and no queue. Nothing downloads because you visited. Paste the writing or hand over a file, and the analyser loads into that tab only once you ask for a result.
One document or a folder of them. Any length past fifty words, and PDF, Word and plain text are all read here in the browser, so a marked essay never leaves the machine it was opened on.
Text is cut at sentence boundaries into blocks of about two hundred words, because that is what fits the analysis window and because a passage is something you can actually go and read.
The document comes back tinted block by block, with an overall figure weighted by length. A part-written essay shows as exactly that rather than as one flat average.
Documents do not sort themselves into written and generated. The common case is neither: a draft a person wrote and a model tidied, or an outline a model produced and a person filled in. The readout separates the three because they call for different conversations.
The opening, the middle and the close all land in the same band. Consistency across a whole document is the shape generated writing makes, because it was all made the same way.
A few blocks sit well over the line while the rest stay low. That is a document somebody wrote and then had rewritten, or one where a section was pasted in.
Every passage stays at the bottom of the scale. The writing behaves like a person's, which is the absence of a signal rather than proof of authorship.
Not every use of a model is a problem. These are the situations where somebody needs to know, and where the answer changes what happens next.
An assignment where the question is whether the student can defend the argument.
A covering letter identical in shape to forty others in the same pile.
Copy filed under a byline, on a deadline nobody had time to research.
Hundreds of reviews for a product nobody bought, written in an afternoon.
A methods section that reads well and describes work that was never done.
A submission citing cases that sound right and do not exist.
A thousand pages of copy aimed at search results rather than readers.
A message that answers everything and commits to nothing.
A slush pile that filled up faster than anybody can read it.
Identical arguments submitted a thousand times to look like public opinion.
A recommendation whose warmth is entirely generic.
Instructions describing a version of the software that never shipped.
Every passage gets a likelihood from 0 to 100, and the headline figure is those passage scores weighted by length. The five bands and the decision line at 65 are the ones the picture detector uses, so a 61 means the same kind of thing here as it does on a photograph.
Read the passage strip as well as the number. A document at 55 made of passages scoring 20 and 95 is a different finding from one where every passage scores 55, and only the first of those is worth a conversation about a specific paragraph.
| Score | Band | What it means | What to do next |
|---|---|---|---|
| 0 – 20 | No AI signal | Nothing in the writing leans towards a language model. | Reads like writing by a person. Worth knowing where it came from all the same. |
| 20 – 45 | Probably written by a person | A weak reading, of the kind careful editing alone can produce. | Take it as human-written unless something outside the text says otherwise. |
| 45 – 65 | Inconclusive | The passage sits between the two populations the analysis separates. | Do not carry this into a decision in either direction. Ask for a draft or an outline. |
| 65 – 90 | Likely AI-written | Over the decision line, where generated prose usually lands. | Over the line, but near enough that heavy editing could have lifted it. Check the passages. |
| 90 – 100 | Strong AI signal | As firm as this analysis gets on a block of text. | Work on the basis that a model wrote most of this, and ask for the drafting history. |
The bands are fixed and published, so a 72 next term means what a 72 means today.
None of it is proof. It is a measurement of how writing is put together, and people who write plainly and to a formula will always score higher than people who do not.
It reads how the writing is put together, not what it says. Generated prose has habits that survive editing: sentence lengths that cluster instead of varying, a preference for hedged and balanced constructions, transitions arriving on schedule, and a vocabulary that stays in the middle of the register from beginning to end. Those are statistical properties of the text, and they can be measured.
The analysis was trained on a public benchmark built for exactly this comparison, covering many generators rather than one, and it reads about two hundred words at a time. That window is why the document is split into passages, and it turns out to be the more useful shape of answer anyway.
Nothing about a document's origin, file name, metadata or formatting is consulted. Two files with identical words score identically, whatever produced them.
This measures a register, not an author. Writing that is deliberately plain, formulaic or heavily edited scores higher whoever produced it, and that is the single most important thing to hold on to before a result changes anything for a real person.
On paragraphs of a hundred words or more the separation is wide: human writing lands in the twenties and thirties, generated writing in the nineties, and very little arrives in between.
Below fifty words the separation collapses, which is why fifty words is a floor rather than a suggestion. On one and two sentence inputs the same analysis returned figures in the sixties and eighties for writing a person plainly produced, and a tool that accuses somebody on the strength of a sentence is worse than one that declines.
A false positive here has a cost nothing else on this site carries: a student or a writer accused of something they did not do. That is the reason for the fifty-word floor, the passage strip instead of one number, and the middle band that says inconclusive rather than guessing.
A false negative is the easier failure to live with and the easier one to cause. Asking a model to write in a specific voice, then editing the result, moves a passage down the scale, and there is no honest way to claim otherwise.
Read the result as a reason to ask a question, never as an answer. The question that settles it is almost always about drafts, notes and the ability to talk about the argument.
Drafting history beats any detector. Version history in a document, notes and outlines, a browser's own revision trail, and a five-minute conversation about the argument will each tell you more than a score will. For coursework, asking for the working is both more reliable and fairer than asking a model about a model.
There is no paid tier of this AI content detector holding back the part you need. The passage strip, the per-block scores, the caveats and the file reading are all in the free version, because there is no other version.
The document comes back tinted block by block, so a part-written essay shows as exactly that rather than as one flat average.
The figure, the band it falls in, and what that band means for what you do next. The same five bands as every other tool here.
Add a folder of them at once. Files are parsed in the browser, so a submitted essay is never uploaded to be read, which is the whole point on this page.
Under fifty words it says so instead of scoring. That is where the analysis stops separating, and a confident number there would be a lie.
Somebody else's writing is somebody else's. It is read and scored in your own browser and never sent anywhere.
Results are kept in your own browser storage so you can come back to them, and clearing your site data removes them.
This is the part where most AI content detection tools ask you to trust a policy. There is no policy to trust here, because there is no transfer. It matters particularly on this page: the text people check is usually somebody else's, often submitted in confidence, and several well-known checkers keep what they are given and use it.
A measurement is the reliable path, but generated prose has tells a careful reader catches, and knowing them makes the score easier to interpret rather than harder.
Every one of these appears in ordinary careful writing, particularly under exam conditions or in a second language. Use them to decide what to check, not to reach a conclusion.
Most free AI content detector sites follow one pattern: take the text, keep it, return a percentage with no working, and say nothing about how much of the document was actually read.
| Capability | Original or AI | Typical free checker |
|---|---|---|
| Runs on your device, nothing transmitted | Yes | No Your text goes to a server |
| Text is not kept or reused | Yes Never received | No Often retained |
| Scores passage by passage | Yes About 200 words each | Partly Sometimes sentence-level |
| Published band thresholds | Yes Five bands, line at 65 | No |
| Declines text too short to judge | Yes Under 50 words | No Scores anything |
| Reads PDF and Word in the browser | Yes | Partly Uploaded to be parsed |
| States what raises a score innocently | Yes | No |
| No word cap on the free tier | Yes | No Usually a few hundred |
| Works without an account | Yes | Partly Often after a sign-up wall |
| Result kept only in your browser | Yes | No Stored server-side |
| Names the model that wrote a passage | No Nobody can do this reliably | Partly Frequently claimed |
| Proves who wrote a document | No No detector can | Partly Sometimes implied |
“Typical free checker” describes the pattern shared by the free web tools we have used, not one named product. Where a competitor does better on a row, that row is wrong and we would like to be told.
Paste the writing you were unsure about, or drop a folder of files in, and see which blocks the analysis reacted to. Free, nothing uploaded, no word cap.
Check for AI textNo sign-up. No upload. Fifty words minimum.