Anthropic introduces imperceptible watermark in texts generated by Claude
Anthropic has started inserting an imperceptible watermark in all texts generated by its latest Claude models. The change went into effect for models launched in the EU on August 2, 2026, and will be extended globally. This initiative follows the signing of the EU AI Act's Code of Practice on AI, which imposes stricter transparency standards.
Quick Answer
Anthropic has introduced an imperceptible watermark in texts generated by its Claude models starting August 2, 2026, in the EU. The mark is applied at the model level and persists through copy-pasting, but can be removed with substantial changes. The exact method remains secret, but it is believed to be based on a statistical signature.
Technical features of the watermark
The watermark is woven directly into the generated text, without altering its meaning, quality, or readability. Since it is an integral part of the text, it travels with it when copied and pasted elsewhere and can persist through some changes. Generated files receive an additional layer of protection in the form of signed metadata according to the C2PA standard, which acts as a digital manifesto to track the file's origin and any modifications.
Secret watermarking method
Anthropic has not revealed the exact method used to create the watermark. The supporting document describes the mark as model-level (the model is trained with it) and native to the text (it is not an external tool like a metadata generator). Documentation for detection and the precise technique have not yet been published.
Open-source projects for watermark removal
Despite the introduction of the watermark, open-source projects for its removal have already emerged. The project mikiane/claude-watermark-cleaner, with 106 stars on GitHub, cleans invisible Unicode characters and rewrites the text with a non-Claude model to disrupt the token pattern. A broader project, guillaumemeyer/watermarks-remover, with 4.6k stars on GitHub, removes Claude text marks, as well as C2PA and SynthID-class signals in various file formats such as PNG, JPEG, SVG, PDF, and DOCX.
Privacy implications
The reaction from privacy experts has been particularly intense, considering Anthropic's history. In March, the company removed a hidden tracker in Claude Code after researchers discovered it was labeling user location and proxy usage through undisclosed Unicode markers, the same silent marking technique now at the heart of the watermark plan.
Watermark usage and limitations
The watermark demonstrates that Claude contributed to text generation, but it does not prove that it wrote it entirely. For example, if you ask Claude to correct or translate a paragraph, the resulting text can still carry the signal. Anthropic admits that substantial changes can remove the watermark and that the absence of the mark does not prove that a text was written by a human.
Regulatory and future of watermarking
The U.S. COPIED Act bill proposal promotes the idea of a standardized method for watermarking AI-generated content, allowing platforms to track its origin. This becomes particularly relevant considering that Claude models are already under intense scrutiny for how they handle the texts they generate and manipulate.
Detection tools not yet available
Anthropic has not yet specified when it will publish the detection tools that will allow anyone to verify the presence of the watermark. Until then, complete removal of the mark cannot be guaranteed, and open-source projects continue to explore alternative methods to bypass this technology.
Impact on transparency and trust
The introduction of the watermark represents a significant step toward transparency in AI content generation. However, the secrecy of the method and the rapid appearance of tools for its removal raise questions about its long-term effectiveness. The community of researchers and developers will continue to closely monitor developments, seeking to balance the need for traceability with user privacy protection.
Evolution of the AI watermarking market
Anthropic's introduction of the watermark comes at a time of rapid technological evolution in the field of AI watermarking. Competitors like Google, Microsoft, and Meta are developing similar solutions for their language models, creating an ecosystem where different watermarking technologies could coexist or compete. This development could lead to more consolidated industry standards, but also to market fragmentation with different technical approaches.
Impact on content platforms
Content platforms like Twitter, Facebook, and YouTube are closely watching the evolution of AI watermarking. The ability to identify AI-generated content could influence moderation and authentication policies. For example, Twitter could use the watermark to identify bots generating false content, while YouTube could apply it to distinguish between human-created videos and AI-generated videos.
Legal and intellectual property challenges
The introduction of the watermark raises important legal questions regarding intellectual property. If a text generated by Claude is modified and the watermark is removed, who holds the rights to that content? Current laws are not clearly defined in this area, and new legislative frameworks may be needed to address these challenges. Technology companies and intellectual property lawyers are already beginning to explore these issues.
Reactions from the publishing sector
The publishing sector has welcomed watermarking with interest, seeing it as a potential tool to combat disinformation and false content. However, some publishers express concerns about the possibility that the watermark could be used to track and monitor the use of their content. This could raise privacy and security issues for both publishers and users.
Future of AI watermarking
The future of AI watermarking is still uncertain. On one hand, it could become an industry standard, improving transparency and trust in AI-generated content. On the other hand, the rapid evolution of watermark removal technologies could make this solution ineffective. The technological and regulatory community will need to collaborate to find a balance between traceability and privacy, ensuring that watermarking remains a useful tool without compromising user rights.
International collaboration
The global nature of the internet requires international collaboration to address the challenges of AI watermarking. Organizations like UNESCO and the OECD are exploring ways to promote international standards that can be adopted in different jurisdictions. This collaboration is essential to ensure that technological solutions are effective globally and do not create artificial barriers between markets.
User education and awareness
Finally, user education is fundamental to the success of AI watermarking. Users must understand what the watermark means, how it works, and what its implications are for privacy and security. Technology platforms and educational organizations have a crucial role in disseminating this information, helping users navigate a world where AI-generated content is increasingly present.
The introduction of Anthropic's watermark represents a significant step toward greater transparency in AI content generation. However, technical, legal, and regulatory challenges remain substantial. The technological, regulatory, and academic communities will need to collaborate to ensure that AI watermarking reaches its potential without compromising user rights and freedoms. The future of watermarking will depend on the ability to find a balance between traceability and privacy, ensuring that this technology remains a useful tool for all.
Editorial Note and Disclaimer
The guides and content published on GoYou are the result of independent research and analysis activities, for informational, educational, and in-depth purposes.
GoYou does not constitute a journalistic publication or an editorial product pursuant to Law No. 62/2001 and does not perform real-time information activities.
The GoYou project does not provide professional, technical, legal, or financial advice and disclaims all responsibility for the improper use of the information published.
In the Crypto sector, every investment involves risks: the reader is invited to always inform themselves autonomously before making any decision.