Vale CMS Need more than a rule playground? Explore

AI Writing Detector

Paste any text to see the phrases and habits that make it read as AI-written. Each one is highlighted where it sits, with the reason and how to fix it. Nothing leaves your browser.

Your text
Paste an essay, an email, a blog post…

Runs in your browser. Your text is never uploaded or stored on a server.

What this AI detector looks for

Legalese is how contracts read. Modelese is how unedited model output reads. Language models have habits. They reach for the same vocabulary (“delve,” “tapestry,” “robust”), the same sentence shapes (“it’s not just X, it’s Y”), and the same ways of opening and closing (“In today’s fast-paced world,” “I hope this helps!”). Readers have learned to spot these, and text that uses them gets trusted less, whoever wrote it. This page finds those habits. It doesn’t try to say who wrote the text, because no tool can do that reliably.

It checks your text against 128 of those patterns and shows you each match. The patterns come from ai-tells, an open-source collection maintained in public, so you can read exactly what is being checked. Beside them run the nine rules of Modelese, each found by comparing what people and models wrote in answer to the same questions.

What the result means

The headline number is how many tells were found, and the rate is tells per 1,000 words so that long and short texts compare. As a rule of thumb:

  • None found: the text avoids the usual patterns.
  • Light (under 5 per 1,000 words): a few stock phrases, normal for anyone.
  • Noticeable (5 to 15): enough that a reader may wonder.
  • Heavy (over 15): reads like unedited model output.

These bands are a reading aid, not a calibrated verdict. The page reports what is on the page. It can’t know who put it there, and people who write formally, or in a second language, use many of these phrases naturally.

The overall read

Beside the phrases, the page measures the text as a whole: how much sentence lengths vary, how often words and word pairs repeat, how expected each word is, and how often common words and punctuation marks appear. Each measure has a weight, and the weighted sum is placed against where human writing lands. It needs at least fifty words.

  • Reads like most human writing: 19 human texts in 20 score here.
  • Has some of its habits: 1 human text in 20 scores this high.
  • Reads like typical model output: 1 human text in 100 scores this high.

The weights were fitted to public collections of paired human and model writing, and to issues, pull requests, and commit messages written before 2022. Text from a model or a kind of writing unlike those is scored less reliably. On the model text in those collections, most of it from older and smaller models, it separates the two well. On text from a current frontier model it mostly does not: those models have been trained away from these habits, and their writing lands in the human range. So a low score doesn’t mean a person wrote the text, only that it doesn’t read as generated. The three lines under the result name what moved the score most, so you can see why it landed where it did.

Why not a percentage?

Detectors that return “87% AI” are guessing from statistics you can’t see, and they are wrong often enough that students have been falsely accused of cheating. A list of specific phrases is something you can verify and act on. If the goal is better writing, the fix is the same either way: replace the stock phrase with what you actually mean.

Check a Word document, PDF, or Markdown file

Drag a file onto the editor, or choose Open. The file is read in your browser, so it is never uploaded. Files up to 25 MB are accepted.

FormatFilesWhat is read
Plain text.txt .textRead as it is.
Markdown.md .mdx .markdown .qmdRead as it is. Code blocks, URLs, and markup are skipped.
Word document.docxParagraphs, headings, and table cells. Tracked deletions are left out.
PDF.pdfThe text layer, with hyphenated line breaks rejoined. A scan needs OCR first.
HTML.html .htm .xhtmlThe page’s text. Scripts, styles, navigation, and code are left out.
reStructuredText.rstProse and titles. Directives and literal blocks are left out.
AsciiDoc.adoc .asciidocProse and titles. Listing blocks and attributes are left out.
Org.orgProse and headings. Source blocks and keywords are left out.
LaTeX.texThe document body. Math, figures, citations, and commands are left out.
Subtitles.srt .vttThe captions, without their numbers and timings.
Jupyter notebook.ipynbThe Markdown cells. Code and output are left out.

Check every draft automatically

The same patterns are available as a package for Vale, which runs them in your editor and on every pull request. If a coding agent writes your docs or commit messages, Vale can check its output before you read it.

Questions

How does this AI detector work?
It looks for Modelese: the habits that make writing read as generated. It searches your text for the patterns language models overuse, such as stock phrases (“delve into,” “rich tapestry”), formulas (“it’s not just X, it’s Y”), filler openings and closings, and marketing-style headings. Each match is highlighted with a suggested fix. It also measures the text as a whole, such as how much sentence lengths vary and how often words repeat, and places that against human writing. It describes how the text reads. It does not, and cannot, tell you who wrote it.
Can it tell me for certain whether text was written by AI?
No, and no detector can. People wrote these phrases long before language models existed, and models learned them from people. A high count means the text reads as machine-written to many readers. It is not proof of how it was produced, and it should never be used on its own to accuse someone.
Can I check a Word document or PDF for AI writing?
Yes. Drag the file onto the editor or choose Open. Word (.docx), PDF, Markdown, HTML, and several other formats are read in your browser, so an essay or report never leaves your device.
Is this AI detector free?
Yes. There is no sign-up, no word limit, and no daily cap. The patterns come from an open-source project, and the check runs on your own device.
Is my text stored or used to train AI?
No. The check runs entirely in your browser. Your text is not uploaded, logged, or shared, and no AI model reads it.
Does it catch text from the newest AI models?
Often not, and nothing reliably does. Newer models have been trained away from the stock phrases older ones overused, and in our tests the text of a current frontier model reads, by these measures, like human writing. What this page finds is the writing readers react to, whoever or whatever produced it. Text without those habits passes, because it doesn’t read as generated.
How is this different from other AI detectors?
Most return a probability that text is AI-generated, produced by a model you cannot inspect. This one shows you the exact phrases it found and why each was flagged, so you can check every result yourself and fix the ones that matter. It is built for editing your own drafts, not for grading someone else’s.
How do I make AI-assisted writing sound less like AI?
Replace each highlighted phrase with something specific. Cut the opening that restates the question and the closing that summarizes what you just said. State claims directly instead of hedging them, and delete words like “robust,” “seamless,” and “comprehensive” unless you can say what they mean here.