Detect AI-generated text, expose its traces, and clean them

How the detector works

Text passes through four independent layers. The first inspects the encoding itself: tight em dashes, curly quotation marks, non-breaking spaces and zero-width characters. These are the strongest signals, because an ordinary keyboard produces almost none of them.

The second layer reads document structure: leftover Markdown, hierarchical headings, bullet lists, and assistant-style openings and closings addressed to a conversation partner. The third layer searches a dedicated formulaic-phrase lexicon for each of the five supported languages.

The fourth layer is statistical. It measures sentence-length burstiness, lexical diversity and paragraph uniformity. Human writing is irregular by nature; a language model produces a far smoother distribution.

Why hidden characters give text away instantly

When you copy from a chat interface you do not copy letters alone. Characters you cannot see travel with them: zero-width spaces, non-breaking spaces, word joiners, byte-order marks. They survive pasting into Word, Google Docs and assignment submission systems.

The most serious case is Unicode tag characters in the U+E0000–U+E007F range. An entire message can be hidden with them inside text that looks completely ordinary — a watermark, a tracking identifier, or instructions aimed at another model that will read the text later. This tool decodes those messages and shows them to you.

Every character is reported with its line and column, its type, and a risk rating, alongside a visual map of the whole text — and you can strip them all with one click.

Can it really identify the model?

It can estimate, not prove. Each model has measurable stylistic habits: OpenAI models overuse tight em dashes and words like “pivotal”, “seamless” and “delve”; Gemini builds bullets that open with a bold lead-in and uses emoji heavily; Claude leans on explicit hedging, longer sentences and almost no emoji; DeepSeek and Qwen impose rigid ordering and sometimes leave full-width Chinese punctuation behind.

These habits weaken as soon as a human edits the text, and they shift between model versions. That is why the result is shown as a relative distribution with an explicit confidence level, never as a single definitive name.

Responsible use

This is a diagnostic tool, not a cheating tool. A score alone cannot justify accusing a student or an employee: polished human writing can score high, and well-edited machine text can slip through. Any academic or professional decision should rest on multiple lines of evidence and a conversation with the author.

The clean-up function has a clear legitimate purpose: removing dangerous characters and watermarks from text you own, and editing a draft you produced with AI assistance so that it carries your voice. Cleaning does not make copied text original, and it does not remove any disclosure obligation you may have.