Most AI detectors hand you a percentage and no reasons. This scanner lists the actual traces: the invisible characters riding in the text, the punctuation a chat window leaves behind, the forty-odd phrases models overuse, and the structural habits that make generated prose read as generated. Free, unlimited, and nothing leaves your browser.
Last updated September 2026
Paste text and click Scan. You get a style score plus four lists: hidden characters (by code point), typographic tells, phrase markers with examples from your text, and structural habits. The score is a read on style, not a verdict on authorship. Click through to clean the mechanical parts.
Scans are free and unlimited. The Humanizer, one click away, counts toward the daily free cleans.
| Group | What is checked | How certain |
|---|---|---|
| Hidden characters | Zero-width spaces and joiners, word joiners, byte order marks, soft hyphens, directional marks, tag characters, non-breaking and typographic spaces, Cyrillic and Greek look-alikes | Certain. They are there or they are not, and you see the code point. |
| Typographic tells | Em and en dashes, curly quotes, the ellipsis character, non-breaking spaces | Strong hint. Chat windows and word processors both produce them. |
| Phrase markers | delve into, a testament to, rich tapestry, in today's fast-paced world, plays a pivotal role, harness the power of, navigate the complexities, chatbot openers, and about thirty-five more, each weighted | Pattern. People use them too, models use them far more. |
| Structure | Chatbot openers and closers, summary openers, Markdown residue, colon-led lists, rule-of-three triplets, repeated sentence starts, uniform paragraph lengths, low sentence-length variation | Pattern. Tidy generation leaves tidy shapes. |
Four signals are weighted and combined: sentence-length burstiness (human prose varies more), phrase-marker density per hundred words, lexical diversity measured with a moving window, and em dash density. Texts under 30 words are not scored; texts under 100 are scored with reduced confidence. The result lands in one of three tiers: reads human, some tells, strong tells. It is deliberately a coarse gauge, because a precise-looking number would be a lie.
Whether a model wrote the text. Style is shared. A careful writer with a fondness for triplets and semicolons scores like a model, and a model asked to write plainly scores like a person. The one signal that is close to certain, a deployed statistical watermark, is invisible to everyone but the vendor holding the key. Read the watermark explainer for that boundary.
Paste the text
Anything from a sentence to a full article. Under 30 words is listed but not scored.
Scan
Four groups are checked in your browser. Each finding shows a label, a count and, for phrases, an example from your text.
Read the score as a style read
Hidden characters are certain. Everything else is a habit that people share with models. Use the list, not the number.
Fix what is mechanical
Send the text to the Humanizer for rule-based cleanup, then edit the remaining flagged phrases yourself.
It is a signature scanner. It lists the concrete traces AI-assisted text tends to carry, hidden characters, typographic tells, phrase markers and structural habits, and turns them into a style score. It does not claim to know who wrote the text. Detectors that give a verdict with no explanation produce false positives on human writing; this one shows you exactly what it found so you can judge.
Four groups. Hidden characters: zero-width spaces, joiners, directional marks, tag characters, odd spaces and look-alike letters, which are certain and shown by code point. Typography: em dashes, curly quotes, ellipsis characters, non-breaking spaces. Phrases: more than forty markers such as delve into, a testament to and in today's fast-paced world. Structure: chatbot openers and closers, summary openers, Markdown residue, rule-of-three lists, repeated sentence starts and low sentence-length variation.
No. SynthID-Text and the model-level mark Anthropic describes for Claude live in word choice and can only be read with the vendor's key. No public tool can detect them, and this one does not claim to. It reads style, which is a weaker but honest signal.
No. It means the text uses habits that are common in model output. People write with em dashes, triplets and formal transitions too. Hidden characters are the one group that is certain, and even those only prove the characters are present, not where they came from.
No. The scan runs in your browser and is free and unlimited. Nothing is sent to a server or a model.
Click through to the Humanizer. It applies rule-based fixes for the mechanical signatures, straight quotes, no dashes, filler phrases collapsed, openers and closers cut, hidden characters removed, and lists what is left for you to edit by hand.
Rule-based cleanup with every change shown.
The hidden-character cleaner on its own.
Why style is the only signal outside the vendor.
Files and text, layer by layer.
Pricing
No card needed to start. Upgrade anytime.