AI Watermark Remover
Strip character-level AI provenance marks from text you own — invisible Unicode, exotic spaces, bidi controls, tag characters, homoglyphs and Markdown/HTML attribution artefacts found in output from Claude, Gemini, GPT and open-LLM pipelines.
Coming soon: binary file metadata (PNG, PDF, DOCX) — stripping EXIF, C2PA manifests and provenance tags from files, not just text.
Not removable: statistical token-sampling watermarks such as SynthID live in word choice, not invisible characters, so they can't be stripped without rewriting the text itself.
Input
Cleaned
Original text · marks highlighted
Detection report
Paste or upload text to run a scan.
Why invisible characters end up in AI-generated text
Between the model, the chat interface, the clipboard and the Markdown renderer, text picks up codepoints that render as nothing: zero-width joiners, narrow no-break spaces, bidirectional overrides, Unicode tag characters and variation selectors. Some are cosmetic side effects of typographic post-processing. Others are deliberate provenance signals designed to survive copy-and-paste. Either way, parsers see what renderers hide — which is why a paragraph that looks identical to another can fail an equality check, break a diff, or slip past a search index.
What this cleaner removes, and what it cannot
Every rule here is deterministic and auditable: each detection shows the exact codepoint, a label, an occurrence count and the replacement applied, and you can download that audit as a plain-text report alongside the cleaned file. It will not touch statistical token-sampling watermarks such as SynthID — those live in word choice, not in characters, so they can't be stripped without rewriting the text. Binary file metadata (PNG, PDF, DOCX) is on the roadmap and coming soon.
Read the full guide to invisible Unicode in AI text, or browse the FAQ for privacy and safety details.