Initializing, please wait a moment

Paste text or choose a .txt file, click Scan for Signals, then read the evidence table. The summary is always signals, not a verdict.

Paste text or pick a file, then click Scan for Signals. Everything runs in your browser - nothing is uploaded.

Why there is no percentage score here. No tool can reliably determine whether text was written by an LLM: statistical detectors false-positive on real human writing, and OpenAI withdrew its own classifier for low accuracy. Gemini can watermark text statistically (SynthID-Text), but only Google can check it - ChatGPT and Claude output carries no readable watermark in plain text. What this page CAN do honestly is show you concrete, checkable artifacts - invisible characters, lookalike letters, punctuation and repetition patterns - and tell you what each usually means. Humans and word processors produce many of these too, so treat every row as a lead, never as proof.

Example result - a pasted chatbot answer typically looks like this:

4 machine-writing signals found (example)
SignalCountWhat it usually means
Em-dashes7 (4.1 per 1,000 chars)Word-processor or LLM-default punctuation; a plain keyboard types a hyphen
Stock phrases3 ("delve into", "it is important to note", "in conclusion")Chatbot-flavored phrasing; humans use these too
Sentence-length uniformitymean 18.2 words, deviation 3.1Very even rhythm; human writing usually varies more
Zero-width space2 (first at character 214)Invisible character - a paste/pipeline artifact worth investigating

Summary: signals, not a verdict - no tool can reliably determine LLM authorship.

AI Text Detector


Scan pasted text or a .txt file for machine-writing signals: invisible Unicode characters, lookalike letters hidden inside Latin words, a smart-punctuation profile, chatbot stock phrases, and repetition statistics. Every signal is an evidence row - what was found, and what it usually means.

This page deliberately gives you no percentage score. No tool can reliably determine whether text came from an LLM: statistical detectors false-positive on real human writing, and OpenAI withdrew its own classifier for low accuracy. Gemini can watermark text statistically, but only Google can check it, and ChatGPT and Claude output carries no readable watermark in plain text. What a byte-level scan CAN show honestly is concrete artifacts - a zero-width space at character 214 is a fact you can verify, not a probability.

Typical uses: a teacher reading a submission that feels off, an editor screening a draft, or anyone whose paste picked up strange invisible characters. Each caveat is printed next to its signal, because humans and word processors trip most of these too.

The analysis runs in your browser - nothing is uploaded and the page works offline after it loads. Checking an image instead? Use the AI watermark detector, which reads C2PA Content Credentials and generator metadata from image files.

← Back to Utility Tools

Related tools:

Tags: #utility

Related guides:

Loading reviews...

Frequently Asked Questions

Why does this page not give an AI percentage like other detectors?

Because the number would not be honest. Statistical AI-text detectors false-positive on real human writing - non-native English writers get flagged disproportionately - and OpenAI withdrew its own classifier for low accuracy. This page reports concrete, checkable artifacts instead: each row is a fact you can verify in the text yourself.

Do ChatGPT, Claude, or Gemini watermark their text?

ChatGPT and Claude output carries no readable watermark in plain text. Gemini can watermark text statistically over its word choices (SynthID-Text), but that mark can only be checked by Google - no browser tool can read it, including this one.

What do invisible characters in text mean?

Zero-width spaces, word joiners, and directional marks do not appear on screen but survive copy-paste. They show up when text moved through machine pipelines, rich-text editors, or deliberate marking schemes. Finding them tells you the text has a processing history worth asking about - not who or what wrote it.

If the scan finds em-dashes and stock phrases, is the text AI-generated?

Not necessarily. Word processors insert smart punctuation automatically, and plenty of humans write "in conclusion" and "delve into". These signals lean machine-flavored when they stack up densely, which is why each row shows a count and a caveat instead of a verdict.

Is my text uploaded anywhere?

No. The analysis is plain JavaScript running in your browser; the page works the same if you go offline after it loads. Nothing about the text or the result leaves your device.

Can it check Word documents or PDFs?

Not directly - it reads pasted text and plain .txt or .md files. For a Word document or PDF, copy the text out and paste it in. Document-level metadata is a different trail: for images, the AI watermark detector reads provenance marks from the file itself.