AI words checker
Paste a text and get a count of AI writing tells by category: vocabulary, connectors, sentence patterns, punctuation, formatting and rhythm, each occurrence highlighted. It counts markers; it does not decide who wrote the text.
Nothing is sent to a server: the text stays in your browser, the tool runs offline once the page is loaded. Ctrl+Enter runs it.
- Six categories: vocabulary, connectors, patterns, punctuation, formatting, rhythm; each with a count, a rate per 1,000 words and a level (low, medium, high)
- This is not an AI detector: it never says a text was written by AI and never gives a percentage; it counts markers that human writers also use
- Rates are compared to fixed thresholds published on this page, not to a hidden model
- Rhythm needs at least eight sentences: it measures how regular sentence lengths are (coefficient of variation)
- Everything runs in your browser; the report can be copied or downloaded
What the checker counts
Since 2023 readers have learned to spot a small set of words and shapes that large language models over-use: delve, tapestry, testament to, landscape, crucial, moreover, in conclusion, the "it's not X, it's Y" contrast, the triad of adjectives, the em dash every other sentence, the bullet list under every heading, sentences that all run to the same length. None of these proves anything: good human writers use every one of them. But if you edit AI drafts, or your own drafts, knowing where the markers cluster is useful. This tool counts them and shows where they are.
| Category | What is counted | Medium from | High from |
|---|---|---|---|
| Vocabulary | About 115 words and phrases (delve, tapestry, testament, landscape, crucial, pivotal, robust, seamless, leverage, foster, unlock, embark, realm, beacon, meticulous, and so on) | 5 per 1,000 words | 15 per 1,000 words |
| Connectors | Moreover, furthermore, additionally, in conclusion, it is worth noting, ultimately, in today's world and about 25 others, wherever they appear | 4 per 1,000 | 10 per 1,000 |
| Patterns | "it's not X, it's Y", "not only X but also Y", "whether you are X or Y", "let's dive in", "here's the thing", "the reality is", rhetorical question followed by an answer, "from X to Y", "in a world where", assistant closings | 1 per 1,000 | 3 per 1,000 |
| Punctuation | Em dashes and spaced en dashes used as pauses (semicolons and ellipses are listed but not scored) | 3 per 1,000 | 10 per 1,000 |
| Formatting | Share of lines that are list items, heading-like lines, bold spans or emoji-led lines | 20 % of lines | 50 % of lines |
| Rhythm | Coefficient of variation of sentence length, from eight sentences up; low variation means very regular sentences | cv below 0.50 | cv below 0.35 |
How to read the result
The headline gives the total number of tells and the rate per 1,000 words. The table gives each category with its count, rate and level. The highlighted text shows every occurrence in the color of its category. A high level in one category means the text leans on that habit more than most edited prose; it does not mean the text is machine-written. Editorial prose from a careful human often scores medium on connectors and low elsewhere; a first AI draft often scores high on vocabulary and patterns and medium on formatting. Use the highlights as an editing map: replace, cut, vary.
Why this is not an AI detector
AI detectors claim to tell who wrote a text from its style. Their published false-positive rates are high, they penalize non-native writers, and there is no way to check their verdicts. This page makes no such claim. It counts a fixed list of markers, published above, with fixed thresholds, and shows the occurrences so you can judge. The only technical way to know that a text came from a given model is a statistical watermark checked with the vendor's key; the Claude watermark detector covers that case with its limits.
Fixing what you found
To replace the em dashes it flagged, use the em dash remover. To strip the Markdown lists and bold lead-ins, the AI text cleaner does it in one pass. Vocabulary and patterns are yours to rewrite; the highlighted view is meant to make that quick.
Privacy
The count runs in your browser. Your text is not sent, stored or logged.
Unmarker is an independent product, not affiliated with or endorsed by Anthropic.
Questions people ask
Is this an AI detector?
No. It counts markers that language models over-use and that human writers also use. It never says a text was written by AI and never gives a percentage. The list of markers and the thresholds are on this page.
Which words are counted?
About 115 vocabulary markers (delve, tapestry, testament, landscape, crucial, pivotal, robust, seamless, leverage, foster, unlock, embark, realm, and so on), 32 connectors (moreover, furthermore, in conclusion, it is worth noting), and nine sentence patterns such as it's not X, it's Y. On the French page the lists are French.
What does the level mean?
Low, medium or high compares the rate per 1,000 words to fixed thresholds published on the page. High means the text uses that habit more than most edited prose; it is an editing signal, not a verdict.
Why does rhythm say n/a?
Sentence rhythm needs at least eight sentences to be meaningful. Below that the tool does not compute it.
Can I export the report?
Yes. Copy result copies the report as text; Download saves it as a .txt file. Nothing is sent to a server.