AI Writing Pattern Detector

The number you get is a density of habits, not a guess about who wrote the text. The detector counts phrases and punctuation patterns that language models produce far more often than most people do, scales each count by the length of your text, and adds the results into a score from 0 to 100. A 600-word essay with three stock transitions scores lower than an 80-word paragraph with the same three, because the habit is spread thinner across the page.

Every pattern family has its own ceiling, so one tic cannot flood the result. Buzzwords and dashes can add up to 20 points each. Stock transitions, hedging phrases and filler openers each top out at 12. Uniform sentence length can add 15, contrast clichés and lists of three 10 each, and ellipses, markdown and emoji bullets 5 each. A text has to show several different habits at once to reach the top band.

Treat the bands as an editing signal. A careful human writer who likes em dashes and words such as robust or landscape can land in the middle band without any machine involvement. A model draft that someone has already edited, with the openers cut and the sentence lengths varied, can sit comfortably in the low band. The score tells you how much of the familiar chatbot texture is on the page, and the findings table underneath tells you exactly where it comes from.

The checks fall into two layers. The phrase layer uses a separate, hand-built list for each of the 21 supported languages, so the Turkish list catches Turkish habits rather than translated English ones. In English it looks for:

• Stock transitions: sentence glue such as moreover, in conclusion or at the end of the day.
• Hedging phrases: throat-clearing like it is important to note or keep in mind that.
• Filler openers: scene-setting such as in today's fast-paced world, let's dive in or look no further.
• Buzzwords: the vocabulary that surged once chatbots arrived, including delve, tapestry, testament to, seamless and pivotal.
• Contrast clichés: the "it's not just X, it's Y" and "not only X but also Y" construction.

Studies of millions of scientific abstracts found that words like delves, showcasing and underscores jumped in frequency right after ChatGPT was released, far beyond normal language drift. That excess vocabulary is why the buzzword family carries the highest cap.

The structure layer ignores vocabulary and works on any text:

• Em and en dashes: every em dash counts. An en dash counts only when it is spaced like a separator (word – word), so ranges such as 1990–2000 or 9 – 17 and compounds like Paris–London are left alone.
• Ellipses: three dots or the single … character used as a trailing pause.
• Lists of three: the rhythmic "fast, simple, and reliable" triad, scored by the share of sentences that contain one.
• Uniform sentence length: with 8 or more sentences, a length variation under 30% adds points.
• Markdown and emoji: **bold** markers, # headings and lines that start with an emoji, which usually mean the text was pasted straight out of a chat window.

Turkish and Korean get one extra family: formal verb endings repeated sentence after sentence, a habit that is very typical of model output in those two languages.

After you press Analyze, the first card shows the score with its verdict and five numbers: the phrase list that was used, the word count, the sentence count, the average words per sentence and the sentence length variation. That last figure is the standard deviation of sentence lengths divided by the average, shown as a percentage. People write in bursts, a long winding sentence followed by a short one, so their variation tends to run high. Model output often keeps every sentence within a few words of the same length, and that flat rhythm shows up as a low percentage.

The findings table lists each pattern family that fired, sorted by how many points it added:

• Hits is the raw count.
• Density is hits per 100 words. Phrase families are measured against content words, so small words like the and of do not dilute the result. For lists of three and formal verb endings it is the share of sentences instead, and for uniform sentence length it is the variation percentage.
• Points is that family's contribution to the score.

You can save the score card as an image and export the findings table as CSV or an image.

Below the table, your original text appears with every hit highlighted in the color of the swatch next to its family. Hover over a highlight to see which family it belongs to.

The last card is a cleaned copy. It replaces dashes and ellipses with a full stop or a comma, depending on whether the next word starts with a capital letter, and removes stock openers at the start of sentences. If you tick the checkbox, it also strips ** markers, # headings and leading emoji. You can copy it or download it as a text file.

Treat the cleaned copy as a starting point. A dash swapped for a comma sometimes needs a better fix, and nothing in it touches buzzwords, hedges in the middle of a sentence or sentence rhythm. Those need a human edit.

There is no language model behind this detector. Every check is a fixed rule: a phrase list, a regular expression or a bit of arithmetic on sentence lengths. That design has practical consequences. The analysis runs in your browser the moment you press Analyze, nothing is uploaded to a server, and there is no usage cost to pass on, so the tool is free with no sign up and no word cap beyond the 40-word minimum. You can check a confidential report, an unpublished manuscript or a student essay without the text ever leaving your device.

Rules are also inspectable. A model-based detector hands you a probability and no reasoning. Here every point traces back to a highlighted phrase or a measured statistic, so you can disagree with a specific flag instead of arguing with a black box.

Language detection works in two steps. The writing system narrows the candidates first: Cyrillic points to Russian, Hangul to Korean, kana to Japanese, and Chinese characters are sorted into Simplified or Traditional by characters that exist in only one of the two standards. When several languages share a script, as the Latin-alphabet languages do, the detector counts common words such as articles and prepositions for each candidate and picks the clear winner. If the vote is tied or too thin, it falls back to the language of the page you are on, as long as that language uses the same script. You can always pick the language by hand from the dropdown.

For a language without a phrase list, choose Other (structure only). Only the dash, ellipsis, list, sentence length, markdown and emoji checks run, and a note under the score says that word-level patterns were not counted. Japanese, Chinese and Thai are written without spaces between words, so their word counts are estimated from characters and the lists-of-three check is skipped for them.

If the score is high and the draft is yours to fix, work down the findings table from the top, since it is sorted by points. A few edits usually do most of the work:

1. Cut the openers. Most stock transitions and filler openers can simply be deleted. A paragraph that begins with in today's fast-paced world usually improves by starting at its second sentence.
2. Replace buzzwords with the specific thing. Instead of "a robust solution that streamlines workflows", say what it does: "it imports the CSV and flags duplicate rows". Concrete nouns and verbs carry more meaning than pivotal or seamless.
3. Vary sentence length on purpose. Read the text aloud. Merge two medium sentences into one long one, then follow it with something short. Run the check again and watch the variation figure rise.
4. Break the reflexive triad. Not every list needs three items. Two strong examples, or four honest ones, read less mechanical than three that were chosen for rhythm.
5. Drop the hedges. it is important to note that the deadline is Friday says nothing more than "the deadline is Friday".
6. Rethink the dashes. Some dashes deserve a comma, some a full stop and some a rewrite. The cleaned copy handles them mechanically. If you want to pick one replacement style for a whole document, an em dash remover gives you that choice.

Cutting openers and hedges shortens a text more than people expect. If you are writing to a length target, a word counter confirms the edited draft still fits.

The goal is clearer writing, not beating a detector. A draft that has been through these edits is better for readers whatever its score, and a reader who never sees a score will still notice the difference.

No tool can prove who wrote a text, and this one does not try. Every pattern it counts is something people did long before chatbots: academics hedge, marketers love innovative and seamless, and plenty of good writers use em dashes heavily. Models learned these habits from human text in the first place. A high score means the habits are dense, nothing more.

False positives cluster in predictable places:

• Formal and academic prose leans on transitions like furthermore and consequently.
• Corporate writing is full of leverage, robust and streamline.
• Second-language writers are often taught stock connectives as the correct way to structure an essay, so a careful non-native student can outscore a native speaker who writes loosely.
• Dash-heavy languages follow different punctuation norms. Spanish, French and Russian use dashes for dialogue and as everyday punctuation, so a dash hit in those languages is weaker evidence than in English.

The reverse happens too. Someone who prompts a model for plain, varied prose, or edits its output for ten minutes, can produce text that lands in the low band.

For these reasons the score should never be the basis for accusing a student, rejecting a freelancer or flagging a colleague. Use it the way the findings table is designed to be used: to see which habits a text has and decide whether each one helps the reader.

Short texts are refused for the same reason. Below 40 words a single stock phrase becomes a huge density, and the score would swing on one word choice. The minimum keeps the numbers meaningful.