The Complete Guide to Removing AI Watermarks from Text, Images & Files

What Are AI Watermarks?

AI watermarks are invisible signals embedded in content generated by AI models. They exist in three forms:

• Invisible Unicode characters — hidden characters like zero-width spaces embedded in text
• Statistical patterns — biased token choices that create detectable word patterns
• Metadata credentials — C2PA "Made with Claude" tags in images and files

According to Anthropic's announcement, Claude uses SynthID-Text from Google DeepMind. The EU AI Act requires providers to label AI-generated content, driving watermarking across all major platforms.

How Each AI Provider Watermarks Content

Claude (Anthropic)

Claude uses SynthID-Text, which biases token choices during generation. When selecting between equally valid words, Claude slightly favors tokens that create a detectable statistical pattern. This watermark:

• Survives copy-paste and light editing
• Is invisible to human readers
• Will be detectable via Anthropic's upcoming detection API
• Also adds C2PA metadata to generated images

ChatGPT (OpenAI)

OpenAI has implemented statistical watermarking that embeds patterns in generated text. The watermark is designed to be robust to common text transformations.

Gemini (Google)

Google uses SynthID, the same technology Claude uses for text watermarking. Gemini also adds metadata to generated images.

DeepSeek

DeepSeek implements pattern-based watermarking similar to other providers, though less is known about its specific implementation.

How to Remove AI Watermarks

Step 1: Remove Invisible Unicode Characters

The simplest layer of watermarking uses invisible Unicode characters. These can be stripped deterministically without changing any visible text. Tools like ClaudeRewrite handle this automatically.

Step 2: Rewrite Text to Break Statistical Patterns

Statistical watermarks are embedded in word choices. To remove them, you need to rewrite the text at the sentence and clause level. The key is to use a different AI model than the one that generated the original text — rewriting Claude output with Claude can re-apply the same watermark.

ClaudeRewrite uses Gemini 2.5 Flash to rewrite Claude output, ensuring the statistical pattern is broken without re-applying the watermark.

Step 3: Strip Metadata from Files

For images and files, you need to strip C2PA, EXIF, XMP, and other metadata. ClaudeRewrite handles this automatically for supported image formats (JPG, PNG, WebP, GIF, TIFF).

Why Manual Editing Often Fails

Many people try to remove watermarks by manually editing the text. This approach has several problems:

• The watermark survives light editing — you need significant changes to break the pattern
• Manual editing is time-consuming and error-prone
• You may inadvertently introduce errors or change the meaning
• For images, manual metadata stripping requires specialized tools

The Best Approach: Automated Watermark Removal

Automated tools like ClaudeRewrite handle all three layers:

1. Strip invisible Unicode characters (deterministic, lossless)
2. Rewrite text using a different model to break statistical patterns
3. Strip metadata from images and files

This approach is faster, more reliable, and preserves the original meaning and facts.

References

• Anthropic: Making Claude's compliance with the EU AI Act
• Google DeepMind: SynthID
• C2PA: Content Credentials
• EU AI Act