Anthropic brings hidden markers to Claude outputs

Anthropic is introducing invisible, machine-readable markings into Claude-generated text and supported files, adding a new layer of provenance as artificial intelligence companies face tougher transparency requirements in Europe. The system is designed to follow generated material beyond the Claude interface, although the company acknowledges that substantial editing can defeat detection.

Claude models launched on or after August 2, 2026 are being designed to support the marking system from launch. Older models are not automatically covered yet, with Anthropic still working to extend support to products released before that date. The markings apply worldwide when a supported model is used, rather than being restricted to European users.

Text and files are treated differently. Text generated by supported Claude models contains an imperceptible watermark woven into the output at model level. Anthropic says the mechanism does not visibly alter the meaning, quality or readability of the response. Because the mark is integrated into the generated text itself, copying and pasting the material into another document should not automatically remove it.

The company has disclosed little about the technical construction of the text watermark. It has not identified the statistical or encoding method being used, nor published a public detector. Anthropic says detection tools for users and third parties are being developed, with fuller technical documentation to follow. That leaves independent researchers unable for now to assess the system’s false-positive rates, resistance to deliberate removal or reliability across different languages.

Anthropic says the watermark may remain detectable after some editing. Its limits become more significant when a passage is extensively rewritten. Heavy editing, paraphrasing, translation or combining Claude output with other writing can weaken or remove the detectable signal. Very short excerpts may also contain too little information for reliable detection.

That distinction is important for publishers, universities and organisations considering watermark detection as evidence of AI authorship. Finding a Claude mark would show that material may have been processed by Claude, but it would not establish that Claude originated every sentence or idea. A user could, for example, submit human-written material to Claude for proofreading, translation or summarisation and receive marked output.

Files use a separate provenance mechanism. Supported formats such as SVG, PNG and JPG can carry digitally signed metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, standard. The metadata can indicate that a file passed through Claude and can help reveal whether its provenance information has been altered.

C2PA metadata is not as durable as a watermark embedded within content. Converting a file into another format, re-saving it through software that discards metadata or taking a screenshot can remove the provenance record. Some online platforms may also strip metadata when files are uploaded or compressed. Anthropic therefore cautions that absence of a detectable mark cannot prove that material was created without Claude or another generative AI system.

The marking technology covers supported Claude output across the company’s consumer service, Claude Platform API, Claude Code, Claude Cowork and Claude Tag. Embedded text watermarks are also intended to operate when Claude is accessed through cloud partners including AWS, Google Cloud and Microsoft Foundry. File-provenance support can vary depending on the capabilities offered by individual platforms.

The move follows the European Union’s AI Act transparency provisions, which became applicable on August 2. Article 50 requires providers of generative AI systems to implement machine-readable measures enabling synthetic or manipulated content to be identified. Anthropic has signed the accompanying Code of Practice on Transparency of AI-Generated Content, a voluntary framework intended to help companies demonstrate compliance with those legal obligations.

The rules are accelerating an industry-wide shift towards provenance technologies rather than relying solely on conventional AI-content detectors, whose accuracy can vary significantly with editing and text length. Watermarking offers a direct signal inserted during generation, but research into language-model watermarks has repeatedly shown that robustness depends heavily on the amount of surviving original text and the transformations applied afterwards.



Notice an issue?

Arabian Post strives to deliver the most accurate and reliable information to its readers. If you believe you have identified an error or inconsistency in this article, please don't hesitate to contact our editorial team at editor[at]thearabianpost[dot]com. We are committed to promptly addressing any concerns and ensuring the highest level of journalistic integrity.


Loading next story…
Social Media Auto Publish Powered By : XYZScripts.com