Free tool · no accountExperimental

Where your writing reads like a template

Paste a page and see the phrasing a reader reacts to: stock openings, sentences built to one shape, uniform paragraph starts, claims with nothing to check them against. Every mark is a span of your own text with the rule that produced it.

Or try

Not an AI detector. It does not judge who wrote the text and does not imply it — it finds specific editorial patterns and shows you where they are.

Why not an AI detector

Because they do not work well enough to be published on somebody else's writing. The classifiers are wrong most often on second-language writers and on formal registers — the two groups least able to argue with a verdict — and the cost of that error lands on the author, not on the tool.

And a verdict is not what anybody actually needs. "An assistant wrote this" changes nothing about the draft. "This sentence shape appears four times in nine hundred words, here are all four" is an edit you can make in a minute.

So this tool answers the second question and refuses the first. Every finding is a span of your own text, produced by a written rule, and any one of them can be dismissed — a deliberate three-fold repetition is a rhetorical device, and the tool has no way of knowing that you meant it.

How to read the number

Three things, and the third is the one people skip.

It is a density, not a grade
Weighted findings per 1 000 words, so a short landing page and a long report can sit side by side. High is not bad: a legal page is deliberately uniform, and a manifesto repeats on purpose.
Every point traces to a mark
The breakdown beside the number is the number. Disagree with a family and you can subtract it yourself; dismiss a finding and it stops counting.
Four rules are not the whole of writing
These detectors see recurrence and abstraction. They cannot see whether the argument is right, whether the order is wrong, or whether the piece is worth publishing at all.

What happens to your text

It is analysed inside the request and is not stored. There is no copy on our side to delete, no result page to keep out of search, and no retention window to explain — the reading you are looking at exists in this browser tab.

No model sees it either. Every detector here is deterministic: phrase lists, sentence skeletons and word counts, which is also why the reading is the same every time you run it.

What we keep is one row per reading with no text in it: a hash of the document, the score, the word count and which rules produced it. That is what the per-address allowance is counted from and what tells us whether a threshold change moved the readings.

When the tool declines

codes, not sentences

Every state is a code from the server and a sentence written here, in all three languages. None of them pretends to be a result.

Not enough text
The reading is a density per 1 000 words, so at this length one phrase would swing it between bands. Add text rather than trusting a number that moves on a comma.text_too_short
Too much text
Four thousand words is the ceiling for one reading. Split the page and read the halves — density is normalised, so the two numbers are comparable.text_too_long
Nothing to read
The field was empty, or held no letters.text_empty
This language is not analysed
This site publishes in five languages, and the analysis runs in three of them: English, Ukrainian and Spanish. This text is not one of the three, so no detector ran on it. If the language was detected wrongly, pick it by hand below.language_not_supported
That language is not ready
English, Ukrainian and Spanish are analysed today. Each needs its own phrase lists and its own thresholds, written in that language rather than translated; French and Polish follow once each has them.unsupported_language
Two languages in one text
Neither language is dominant enough to read. The lexicons do not carry across languages, so a blended score would be a number with nothing behind it. Split the text by language.unsupported_mix
Could not tell the language
Too few words to decide. Add text, or pick the language by hand below.language_undetermined
No readings left in this window
They come back as the oldest one ages out. Nothing is queued and nothing is charged.quota_exceeded
The tool is resting
It cannot account for its own usage right now, so it is not taking runs. A tool that has lost its accounting would spend everything it is asked for.tool_resting

Questions people ask

Is this an AI detector?

No. PatternScope does not judge whether a text was written by a person or a model, and does not imply it either. It finds specific editorial patterns — template phrasing, repeated sentence shapes, uniform openings, abstract claims — and shows where each one is. “Written by AI” classifiers are wrong most often on second-language writers and on formal registers, and the cost of that error lands on the author, not the tool.

Does a high score mean the text is bad?

No. Pattern density is an editorial signal, not a grade. A legal page is deliberately uniform and a three-fold anaphora is a rhetorical device, and both read as dense. The number says where to look; whether it is a fault is the author's call, and any finding can be dismissed.

What happens to the text I paste in?

It is analysed inside the request and is not stored — there is no copy on our side, no result page and no retention window. No model sees it either: every detector is deterministic. What we keep is one row per reading with no text in it, which is what the per-address allowance is counted from.

Which languages are analysed, and why not more?

English, Ukrainian and Spanish. Each detector carries its own list of stock phrases, written by a native speaker rather than translated: “у зв'язку з тим, що” and “in connection with the fact that” are two different habits in two different languages, and a translated list measures the translation. French and Polish follow once each has its own evaluation set — the same bar the first three cleared.

What is the score actually counting?

A weighted density of findings per 1 000 words: a light finding counts 1, a medium 3, a heavy 6. Every point traces to one finding with an offset in the text — the breakdown beside the number is the number. One template repeated forty times scores four times and reports the rest as suppressed, so a single habit cannot take the whole reading.

Asked out loud

spoken, not typed

The same worries in the words people use speaking to an assistant rather than typing into a box. Each question keeps the circumstance it came from — the occasion is what makes it a different question, and what makes it answerable.

My editor says my draft sounds like ChatGPT wrote it. Is she right?

Who wrote it is a question this tool does not answer and will not. What it answers is what your editor actually heard: every template phrase, repeated sentence shape and uniform opening, each with its exact place in the text. The impression usually rests on about a dozen fragments, and they arrive as a list.

rewriting an article after review
Five people on my team write landing pages and they all sound the same. How do I show that?

Run a few pages and compare the breakdown rather than the number: if one family dominates in all of them — uniform openings, say — that is a shared template, not a shared voice. Density is per 1 000 words, so a short landing page and a long report can sit side by side.

running a content team
My text is in Ukrainian and half these tools simply do not work on it. Does this one?

Yes, and not by translating English rules. The Ukrainian stock-phrase list is written in Ukrainian and shares no entries with the English one, and its thresholds were derived separately rather than inherited.

writing in two languages daily
I need to show my boss the edits are not a matter of taste. What backs that up?

Every finding is a span of your own text rather than a paraphrase: the rule that produced it is named, and the evidence is exactly the characters highlighted. Not “this reads better” but “this phrase occurs four times in 900 words, here they are”.

defending an edit to a manager

The words on this page

AI slop
Text that is grammatical, on topic and says nothing — the texture this tool measures.
Near-duplicate content
Pages that differ in words and not in substance, which is repetition at the scale of a site.
Editorial guidelines
The written rules a publication holds its own prose to, and what a reading like this is evidence for.
Goodhart's law
A measure that becomes a target stops measuring. Worth reading before optimising this number.