Claude's Invisible Watermark and the Rise of Removal Tools

Anthropic has announced that texts generated by its AI model, Claude, will now feature an invisible watermark. This marking is reportedly linked to word choices. A few days later, the company released some technical details.

Consequently, GitHub and the web have been flooded with tools that claim to remove this watermark. The battle over AI watermarking has begun even before Anthropic has released its own detector, which is expected to be available via its API.

How the Watermarking Works

The concept is based on SynthID Text, a technology developed by Google DeepMind in 2024.

• When Claude generates text, it selects a word from multiple options.
• The watermark subtly alters these choices using a cryptographic key held by Anthropic.
• The text remains natural, but when analyzed with the correct key, the watermark can be detected.

Contrary to some reports, the watermark is not hidden in invisible characters but is embedded within the words themselves. This measure aligns with Article 50 of the AI Act, which came into effect on August 2, 2026, and imposes penalties of up to 15 million euros or 3% of global turnover. Anthropic, therefore, had little choice but to implement watermarking.

The Response is Organizing on GitHub

The most visible project is named watermarks-remover, created by Guillaume Meyer, the founder of Memo. It targets watermarks from Claude, Gemini, and other models. Its method is based on three concepts:

• Removal of Unicode characters hidden within the text.
• Stripping of C2PA metadata from files.
• Rewriting the content by another model to reduce the watermark's presence.

However, Guillaume Meyer admits that while metadata removal is effective, watermark deletion is considerably more complex. He questions the value of paying for premium subscriptions for less capable models to rewrite content.

Other projects have emerged, including claude-watermark-cleaner, noai-watermark, and commercial services like StealthGPT offering removal options. It's important to note that Anthropic has not yet released its detector, making independent verification currently impossible.

Independent Researchers Test Watermark Resistance

Simple edits, such as replacing hyphens or altering spelling, have no effect on the watermark. However, one method has shown promise: inserting Unicode variation selectors, essentially invisible characters, into 20 to 30% of the text. These do not remove the watermark itself but disrupt its analysis to the point of overwhelming the detector.

The detector then fails to identify the watermark.