Remove hidden marks from AI text
Hidden characters in text — and AI metadata in images, PDFs, Office documents, audio, and video. Find and clean them, right in your browser. Nothing is uploaded.
What it detects
- Zero-width and invisible format characters, soft hyphens
- Unicode tag characters and variation selectors
- Bidirectional control characters and control codes
- Unusual spaces and Unicode line separators
- Latin lookalike and fullwidth letters
- C2PA provenance metadata in images and media
- EXIF, XMP, and AI generator signatures in images
- PDF, Office, and audio/video container metadata
Frequently asked questions
- Which hidden characters does it detect?
- Zero-width and other invisible format characters, Unicode tag characters, variation selectors, bidirectional controls, soft hyphens, unusual spaces, control characters, and Latin lookalike or fullwidth letters — patterns reported across ChatGPT, Claude, Gemini, and other AI output.
- Does it remove statistical AI watermarks?
- No. Statistical watermarks from model providers — such as Anthropic's Claude text watermark or Google's SynthID — are embedded in word choice, not characters. Character cleaning does not remove them, and there are no public detectors yet.
- Is my text uploaded anywhere?
- No. Scanning and cleaning run entirely in your browser. Your text never leaves your device.
- What gets removed, and what gets normalized?
- Invisible and control marks are removed. Unusual spaces become regular spaces, Unicode line separators become standard line breaks, and lookalike or fullwidth letters become their ASCII equivalents.
- Will it break legitimate multilingual text?
- No. Characters with legitimate uses — ZWJ/ZWNJ, LRM/RLM, the ideographic space, emoji variation selectors — are kept. Lookalike letters are only replaced when the surrounding text is mostly Latin.