How I Removed Claude's Watermark from 10,000 Words in Under a Minute

The Test Setup

We wanted to push ClaudeRewrite to its limits, so we generated a 10,000-word technical report using Claude 3.5 Sonnet. The document covered AI safety frameworks, model architecture comparisons, and regulatory compliance — dense, terminology-heavy content that would stress any rewriting tool.

The report was pasted directly from Claude into ClaudeRewrite with no pre-processing. We tracked every metric: processing time, characters stripped, watermark detection before and after, and content fidelity.

Breaking Down the 3-Layer Process

Layer A: Unicode Stripping

ClaudeRewrite's first pass identified and removed 847 invisible Unicode characters embedded throughout the document. These included zero-width spaces (U+200B), zero-width non-joiners (U+200C), and various bidirectional override characters. This layer is deterministic and lossless — no visible content was affected.

Layer B: Statistical Watermark Removal

The document was then rewritten through Gemini 2.5 Flash, ClaudeRewrite's rewriting engine. This step breaks the statistical watermark by re-generating the content with a different model. Because Gemini uses a different tokenizer and sampling strategy than Claude, the SynthID-Text token bias pattern is completely disrupted.

The rewriting preserved all technical terminology, data points, and structural formatting. Key definitions, numbered lists, and table data remained intact.

Layer C: Metadata Cleanup

Since this was text-only input, there was no C2PA metadata to strip. However, ClaudeRewrite still performed a metadata scan and confirmed zero embedded credentials in the output.

Before and After Comparison

Before ClaudeRewrite

• Watermark detection score: 98.2% (flagged as AI-generated)
• 847 invisible Unicode characters embedded throughout
• Statistical token pattern consistent with Claude 3.5 Sonnet output
• File size: 42.3 KB (plain text)

After ClaudeRewrite

• Watermark detection score: 0.0% (passes all detection tools)
• Zero invisible Unicode characters
• Statistical pattern matches Gemini 2.5 Flash output (no Claude signature)
• File size: 41.8 KB (slightly smaller due to invisible char removal)

The Practical Limit: Chunking Longer Documents

While this test processed 10,000 words successfully, we found that the practical sweet spot for a single ClaudeRewrite pass is 2,000–3,000 words. Beyond that range, you may want to chunk your document into sections and process each independently.

For documents in the 3,000–5,000 word range, ClaudeRewrite handles them fine but processing time increases non-linearly. For 10,000+ word documents, we recommend splitting into 2,000-word chunks and processing sequentially. This approach gave us the most consistent results and the fastest overall throughput.

The chunking approach also makes it easier to review output quality — you can spot-check each section before combining them into the final document.

Content Fidelity Check

We ran the before and after documents through a diff tool and confirmed that the rewriting preserved 99.4% of the original meaning. The 0.6% variance came from natural paraphrasing (e.g., "utilizes" becoming "uses", minor sentence restructuring). All technical data points, citations, and numerical values were preserved exactly.

Key Takeaways

• ClaudeRewrite can process 10,000 words in under a minute
• The 3-layer process (Unicode strip, rewrite, metadata cleanup) is comprehensive
• Content fidelity remains high — meaning and data are fully preserved
• For best results on very long documents, chunk into 2,000–3,000 word sections
• The watermark detection score drops from ~98% to 0% after processing

References

• Anthropic: Text Watermarking Research — Details on Claude's watermarking approach
• Google DeepMind: SynthID — SynthID-Text algorithm documentation
• Kirchenbauer et al., "A Watermark for Large Language Models" — Foundational research on LLM text watermarking via token biasing
• C2PA: Content Credentials — Metadata standards for AI-generated content