AI Watermark Removers Spread, but Their Claims Remain Unverified

· Original article ↗

Summary

A wave of tools claims to remove Claude’s invisible AI watermark, but Anthropic has released no public detector to verify those claims. Many tools strip metadata, while integrations into agent workflows raise software supply-chain concerns.

Key points

  • Tools including watermarks-remover claim to remove Claude’s invisible watermark, but without a public Anthropic detector their effectiveness cannot be independently checked.
  • Many tools remove hidden characters or file metadata; Anthropic says its text watermark is woven into word choice and may disappear through heavy rewriting.
  • The developer of watermarks-remover says the project currently removes metadata, not the actual text watermark.
  • An independent code review found that one popular text cleaner left a common hidden-payload technique intact.
  • The article flags tools installed into agent workflows as a potential software supply-chain exposure; an optional scoring setup clones a research repository and downloads a roughly 220MB artifact.
  • Anthropic says a detected mark indicates content was processed by Claude, not necessarily written by it, and that editing, paraphrasing, or translation can remove a mark.

Article Details

Event Type
Emergence of AI watermark-removal and detection-evasion tools following Anthropic's disclosure of Claude text watermarking.
Impact
The effectiveness of claims to remove Claude's text watermark cannot be independently verified because Anthropic has not released a public detector. The article says many tools remove hidden characters or metadata rather than the watermark itself, and that integrating unvetted tools into agent workflows creates a potential supply-chain risk; it reports no confirmed compromise.

People

Vendors

Products

ClaudeA market for removing Claude’s new invisible AI watermarks has emerged, led by open-source projects, web tools, and commercial evasion services claiming to clean marked text.Claude CodeMarking is applied at the model level, so it appears across the API, claude.ai, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry.Claude CoworkMarking is applied at the model level, so it appears across the API, claude.ai, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry.Claude TagMarking is applied at the model level, so it appears across the API, claude.ai, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry.GeminiIt began as a Claude-only agent skill and now advertises coverage of Claude, Gemini and SynthID-Text, OpenAI provenance surfaces, and open-weight models using Kirchenbauer-style marks.GPTZeroAnother, Human Writes, advertises bypassing Turnitin and GPTZero on essays and assignments, claims to strip Claude’s watermark, and carries a footer telling users it must only be used in compliance with academicSynthID-TextIt began as a Claude-only agent skill and now advertises coverage of Claude, Gemini and SynthID-Text, OpenAI provenance surfaces, and open-weight models using Kirchenbauer-style marks.TurnitinAnother, Human Writes, advertises bypassing Turnitin and GPTZero on essays and assignments, claims to strip Claude’s watermark, and carries a footer telling users it must only be used in compliance with academic

Tools

Related Articles