How It Works
How Claude's Watermark Works
Claude embeds an invisible watermark in every piece of text it generates. Here's how it works, why it matters, and how to remove it.
What is Claude's watermark?
Claude uses SynthID-Text, a watermarking technology from Google DeepMind, to embed an invisible signal in generated text. Unlike traditional watermarks that add visible marks, SynthID-Text works by subtly biasing the choice of tokens (words and word fragments) during generation.
How it works technically
During text generation, Claude uses a cryptographic key to bias token selection. At each step, the model favors tokens that match a hidden pattern. This creates a statistical signal that persists across the entire text.
- →Invisible: No visible changes to the text. Words and sentences read normally.
- →Survives copy-paste: The signal is embedded in the text itself, not in metadata. Copy-pasting preserves it.
- →Survives light editing: Changing a few words doesn't break the pattern. You need to rewrite at the sentence level.
- →Detectable: Anthropic's upcoming detection API will be able to identify this pattern.
Image watermarks: C2PA metadata
When Claude generates images, it embeds C2PA metadata (Coalition for Content Provenance and Authenticity) that says "Made with Claude." This is a different type of watermark — it's metadata attached to the file, not a pattern in the image itself.
Why the watermark matters
The watermark serves legitimate purposes — AI transparency and trust. But it also creates problems:
- →Freelancers: Clients may reject content flagged as AI-generated, even if it's original and valuable.
- →Businesses: Internal documents and marketing copy may be flagged by AI detectors.
- →Developers: Code and documentation may be flagged even though AI assistance is standard practice.
- →Researchers: Academic work using AI assistance may face scrutiny.
How to remove Claude's watermark
ClaudeRewrite removes the watermark using a 3-layer approach:
Layer A: Invisible Unicode scrub
Claude sometimes embeds invisible Unicode characters (zero-width spaces, word joiners, soft hyphens) alongside the statistical watermark. ClaudeRewrite identifies and removes these characters, ensuring no hidden formatting remains.
Characters removed: U+200B (zero-width space), U+200C (zero-width non-joiner), U+200D (zero-width joiner), U+FEFF (BOM), U+00AD (soft hyphen), U+2060 (word joiner), U+180E (Mongolian vowel separator)
Layer B: Statistical watermark removal
The core layer. ClaudeRewrite rewrites your text using Gemini 2.5 Flash, a different AI model with a different tokenizer. This breaks the statistical pattern that SynthID-Text creates, while preserving your meaning, facts, and numbers.
Why Gemini? Rewriting with Claude could re-apply the same watermark. Using a different model ensures the pattern is broken without re-application.
Layer C: Metadata stripping
For images and files, ClaudeRewrite strips C2PA "Made with Claude" metadata, along with EXIF, XMP, ICC, and IPTC profiles. This ensures no file-level evidence of AI generation remains.
Supported formats: JPG, PNG, WebP, GIF, TIFF
Step-by-step guide
Paste your text
Go to removeclaudewatermark.com and paste your Claude-generated text into the tool. You can also drag and drop .txt, .md, .pdf, or .docx files.
Click Clean Text
ClaudeRewrite automatically applies all 3 layers. No configuration needed.
Copy the clean output
The rewritten text preserves your meaning while removing the watermark. Copy it and use it wherever you need.
Frequently asked questions
Does the watermark survive copy-paste?
Yes. SynthID-Text is designed to survive copy-paste and light editing. To remove it, you need to rewrite text at the sentence and clause level using a different AI model.
Can Claude detect its own watermark?
Anthropic has announced a watermark detection API that will be available soon. Once launched, anyone will be able to check if text was generated by Claude. ClaudeRewrite removes the watermark before detection.
Is it legal to remove the watermark?
Yes. Removing a watermark from content you created is legal in most jurisdictions. The EU AI Act requires AI providers to label AI-generated content, but does not prohibit users from removing those labels from their own content.
Why use a different model to rewrite?
Rewriting Claude output with Claude can re-apply the same watermark. ClaudeRewrite uses Gemini 2.5 Flash, a different model with a different tokenizer, ensuring the pattern is broken without re-application.