● Covered by 2 sources · 2 reportsMedium impact2 neutral
AI Watermark Removal Tools Emerge, Efficacy Unverified Against Text Watermarks
Key points
- Anthropic enabled invisible watermarks in Claude's AI-generated text in August 2026.
- Google's Gemini has watermarked text since 2024.
- AI text watermarking embeds statistical marks by influencing word selection.
- A market for AI watermark removal tools has emerged, including 'watermarks-remover'.
- Efficacy of removal tools is unverified as watermark details and detectors are not public.
Emergence of Watermark Removal Tools
Following Anthropic's announcement in August 2026 that it had switched on invisible watermarks in all text generated by its Claude models, a market for AI watermark removal tools has quickly developed. This includes a GitHub project, newly registered web tools, and existing AI detection evasion services.
One notable tool is 'watermarks-remover' by software developer Guillaume Meyer. Initially designed for Claude, it now claims to support watermarks from OpenAI, Gemini, SynthID-Text, and open-weight models using Kirchenbauer-style marks. Meyer's social media post about the tool gained over 2 million views. Other repositories include 'claude-watermark-cleaner', 'remove-ai-watermarks', and 'noai-watermark'.
The Mechanism of AI Text Watermarking
AI text watermarking works by embedding hidden statistical marks within generated text. This is achieved by subtly influencing the selection of words during the text generation process. Unlike image watermarks, text watermarks do not rely on pixels or metadata, but rather on the choices made between words.
When an AI model generates text, it selects from a shortlist of possible words. Watermarking involves biasing these choices in a way that creates a statistical pattern, which can later be detected without altering the text's readability or surviving copy-and-paste operations.
Current Implementations of Watermarking
Google has implemented text watermarking for its Gemini app and web experience since 2024. Anthropic's new Claude models began marking text at the model level in August 2026, with earlier models also being updated to include this feature. These marks are designed to be invisible and persistent across copying.
Unverified Efficacy of Removal Tools
Despite the emergence of these removal tools, their effectiveness against text watermarks cannot currently be verified. Anthropic has not published details on how its watermarking method works, nor has it released a corresponding detector. Without this information, it is impossible to confirm whether a 'cleaned' document still contains the watermark.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub guillaumemeyer/watermarks-remover GitHub mikiane/claude-watermark-cleaner GitHub wiltodelta/remove-ai-watermarks GitHub mertizci/noai-watermark arXiv 2301.10226How outlets covered it
AI text watermarking embeds hidden statistical marks within generated text by subtly influencing the selection of words during generation. This method allows for the identification of AI-generated content without altering the text's readability, addressing the challenge of verifying AI output. The technique is already in use by major AI models like Google's Gemini and Anthropic's Claude.
A market for tools claiming to remove AI watermarks has appeared following Anthropic's announcement of invisible watermarks in Claude's output. However, the efficacy of these tools against text watermarks cannot be verified because Anthropic has not released details on its watermarking method or a corresponding detector.