Detect and remove invisible AI watermarks

Scan text, URLs or Word, PDF and PowerPoint files for invisible characters, look-alike homoglyphs and typographic fingerprints, then get a clean copy back. Your text and files are processed entirely on your device.

URLs are fetched by our server to read the page (browsers block cross-site reads). The page content is public; nothing you paste in the Text or Files tabs is ever sent.

Drop files here or click to choose
Word (.docx), PDF, PowerPoint (.pptx), .txt, .md, .csv — parsed in your browser
Private by design. Your text never leaves this page.

What this tool can and cannot do

Detect and remove character-level fingerprints. Invisible Unicode characters (zero-width spaces, joiners, bidirectional controls, variation selectors, byte order marks), unusual spaces, look-alike homoglyphs (a Cyrillic "а" hiding in a Latin word) and typographic substitutions are real, verifiable marks. This tool lists them and hands you a clean copy, whether they come from pasted text, a web page or a document.

The statistical watermark is different. In August 2026 Anthropic confirmed that Claude embeds an imperceptible watermark in the word-choice pattern of the text itself, like Google's SynthID for Gemini. That mark is not a hidden character, so no client-side tool can read or remove it reliably: it can only be verified with the provider's key, and only fades under heavy rewriting or translation. Anthropic has announced a public detection API; when it ships we will connect to it here. Sites claiming they remove the statistical watermark today are really just paraphrasing your text with another model.

Honesty notes. A clean result means no hidden characters, not proof a human wrote it. PDF text extraction can drop zero-width characters that carry no glyph, so a PDF scan is best-effort. Homoglyph fixing only touches look-alikes inside mixed-script words, so genuine Cyrillic or Greek text is left alone.