Highlights update as you type and stay in your browser. The two buttons send your text once to an AI model for review. You get 20 checks and 5 rewrites a day.
phrase or punctuation habit sentence flagged by the AI review
This tool measures writing patterns. It cannot tell who wrote a text, and its score should not be used to grade, hire or discipline anyone.
What it looks for
Language models fall back on the same moves over and over. Editors who clean up AI drafts have catalogued them, from Wikipedia's Signs of AI writing page to open-source checklists such as humanizer and stop-slop. This tool checks for them at two levels.
| Pattern | Example | What the rewrite does |
|---|---|---|
| Binary contrast | "It's not a tool, it's a mindset." | States the claim directly |
| Stock vocabulary | delve, tapestry, seamless, pivotal, landscape | Uses the plain word |
| Connector openers | Moreover, Furthermore, Ultimately | Cuts them |
| Colon or question reveal | "The best part: it learns." "The result? Chaos." | Says the point in one sentence |
| Rule of three | "fast, simple, and reliable" | Keeps the items that carry meaning |
| Dramatic fragments | "Big mistake." "That's it." | Folds them into the sentence they belong to |
| Em dashes as pivots | "It worked — until it didn't." | Commas, periods or parentheses |
| Tidy structure | Every paragraph one point, same length; a summary line at the end | Varies paragraph length, ends on the last concrete point |
| Vague claims | "Many people find this helps." | Narrows the wording and asks you for the missing detail |
The yellow highlights come from a rule set that runs in your browser. It counts phrases, punctuation and sentence-length variation, so it reacts the moment you paste. The red highlights come from an AI review that reads the whole text. That review catches what rules miss, such as a casual tone that sounds staged or a closing question to "everyone" that nobody asked for.
How the score works
The score runs from 0 to 100. It blends the AI review (80%) with the browser rules (20%). Below 30 the text reads as a person's; above 60 it carries many machine habits. Before the AI review runs, the number next to the word count is the rule score alone, which is rougher.
We tested the AI review on 136 texts we gathered: 50 written by people (answers from the HC3 dataset and Hacker News comments from 2014 to 2016) and 86 written by models (ChatGPT, GPT-4 and Mistral samples from public datasets, new samples from Kimi K3, MiMo, GLM and MiniMax, and 14 texts already run through a humanizer). At a cutoff of 50 it marked 75 of the 86 AI texts and 3 of the 50 human ones. The three human texts it got wrong were two encyclopedia-style history answers and a job post. Formal, dry writing by people is where it errs most, so read a high score on that kind of text with care.
We have no Claude samples in the test yet, and the test set is small. These numbers describe our data, not every text you might paste.
How the rewrite works
The rewrite starts from the review: the flagged sentences and the document-level habits go to the model along with your text and a list of rules (keep every fact, number and name; stay within 15% of the original length; add nothing that isn't there). The model writes three versions. Each one is reviewed again, and the version with the lowest score wins, as long as it hasn't drifted too far from your length. Then the process runs a second round on the winner. This keep-the-best loop follows the approach in Adversarial Paraphrasing (NeurIPS 2025), with our own review as the guide.
In our test of 12 AI-written samples, a single rewrite lowered the average score from 85 to 72. The three-version loop over two rounds took it from 86 to 54, and 5 of the 12 ended below 50. What kept the others high was vagueness. A rewrite can't fix that without inventing facts, so the tool lists the sentences where your own detail would do the most: which app, what number, whose name.
A rewrite takes 15 to 60 seconds. Read it before you use it. The model can soften a claim or drop a nuance you cared about.
What it can't do
- Prove authorship. A careful person can trip every pattern here, and a model told to avoid them can pass.
- Promise a result on Turnitin, GPTZero or any other detector. We don't test against them.
- Add your experience. The best edit is still a concrete detail from your own life or work.
FAQ
Does this tell me whether a person or an AI wrote a text?
No. It looks for writing habits that language models overuse, such as stock phrases, "it's not X, it's Y" sentences and paragraphs that all make one tidy point. People write that way too, and a model can be told to avoid it. A high score means the text has many of those habits, not that a model wrote it.
Can I use the score to grade students or screen job applicants?
Please don't. On our test set it marked 3 of 50 human texts as likely AI, mostly formal explainers and a job post. That error rate is too high for any decision about a person. Use it on your own drafts.
Will the rewrite get past Turnitin, GPTZero or other detectors?
We don't test against them and we don't promise it. The rewrite targets the patterns our own review flags. In our test of 12 AI-written samples, the average score dropped from 86 to 54 after rewriting, and 5 of 12 ended below 50. The rest still read as generic, because the fix for that is your own detail.
Why does it ask me to add details?
The strongest signal left after a rewrite is vagueness: claims with no name, number, date or example. A rewrite can't invent those without making things up, so the tool points to the sentences where your own specifics would help most.
Is my text stored?
We don't save your text. It is sent once to the DeepSeek Flash model through Tencent Cloud TokenHub to be reviewed and rewritten, and the response comes back to your browser. We keep a daily counter per IP address for the free limit, and that counter expires after about a day.
How many checks do I get?
20 checks and 5 rewrites per day, free, with no account. Each text can be 40 to 1,200 words. The highlights under the text box run in your browser and have no limit.
Does it work for languages other than English?
The browser highlights only know English patterns. The AI review and rewrite also accept other languages, but we calibrated them on English text only, so treat scores for other languages as rough.
Updates
- : First release. Browser highlights, AI review, and a rewrite that keeps the best of three versions over two rounds.
More tools
- textGrain, explained: what OpenAI's invisible text watermark can and cannot prove.
- llms.txt Checker: check your llms.txt against the spec.
- Claude Skill Validator: paste a SKILL.md and check it against the Agent Skills spec.