Claude Will Watermark Its Writing as EU AI Act Rules Take Effect

By Suad Seferi · Aug 12, 2026

Text written by Claude is about to carry something readers will not be able to see. Anthropic is introducing machine-readable watermarks into text generated by its Claude AI models as new transparency requirements under the European Union's AI Act begin to apply. The watermark will be embedded directly into generated text rather than displayed as a visible label. Anthropic says the signal should travel with the text when it is copied and pasted and may survive some editing, allowing compatible software to identify content as originating from Claude. Generated files will be handled differently. Where supported, Anthropic plans to attach digitally signed provenance information, including the C2PA standard used to record information about the origin and history of digital content. The changes apply at the model level and are expected to cover Claude across its different products, including the Claude API, Claude Code and other Anthropic interfaces. Reporting on the announcement says the marking will also travel with supported Claude models when accessed through cloud platforms such as Amazon Web Services, Google Cloud and Microsoft Foundry. Anthropic is rolling out the system globally rather than limiting it to European users. The reason it is happening now is European law. The EU now requires AI-generated content to be detectable Article 50 of the EU AI Act began applying on August 2, 2026. Among its transparency requirements, providers of AI systems that generate synthetic text, images, audio or video must ensure their outputs are marked in a machine-readable format and can be detected as artificially generated or manipulated. The law does not prescribe one universal watermarking technology. It requires providers to use solutions that are effective, interoperable, robust and reliable as far as technically feasible. Anthropic's answer for text is an invisible watermark. Exactly how that watermark works has not yet been publicly detailed. The company says it is developing tools that will allow users and third parties to detect the marks and plans to provide additional technical documentation. That means Claude-generated prose could contain a statistical or structural signal indicating its origin without changing what the reader sees on the page. The watermark is not supposed to alter the meaning or readability of Claude's response. And it is not the same thing as placing “Written by AI” at the bottom of an article. Existing Claude models get more time The August 2 deadline does not mean every existing Claude output suddenly carries a watermark. European Commission guidance provides a limited transitional period for AI systems that were already on the market before August 2. For the marking and detection requirement under Article 50(2), those systems have until December 2, 2026 to comply. New systems entering the European market after the rules became applicable do not receive the same window. That distinction explains why Anthropic can commit to marking new Claude models from launch while continuing to adapt existing products. Axios reported that Anthropic intends to use the marking system worldwide for relevant models, despite the requirement originating in the European Union. If other major AI providers take the same approach, an EU transparency rule could end up changing how AI-generated content is handled far beyond Europe. Watermarks are not proof There is an obvious limitation. A watermark can help establish that text was probably generated by a particular system. Its absence cannot reliably establish the opposite. Anthropic itself has acknowledged that its marking techniques are not infallible. Text can be substantially rewritten, translated or combined with other material. File metadata can disappear as content moves through software or online platforms. That makes the technology different from a traditional signature attached permanently to a document. It is better understood as another piece of provenance evidence. For journalists, universities, businesses and online platforms, that distinction will matter. A detection tool may be able to identify Claude's watermark in a document, but it cannot automatically determine how much human work went into the final version or whether the underlying information is accurate. Nor does the AI Act treat every use of generated text in the same way. The rules for publishers go further Article 50 contains a separate requirement for organisations using AI-generated text to inform the public about matters of public interest. In those circumstances, the deployer may have to disclose that the text was artificially generated or manipulated. There is an important exception. The disclosure requirement does not apply in the same way when AI-generated material has gone through human review or editorial control and a person or legal entity takes editorial responsibility for the publication. That distinction is particularly relevant to newsrooms. Using an AI system somewhere in the reporting or editing process is not automatically equivalent to publishing an untouched AI-generated article. The European rules recognise editorial responsibility and human review as part of that equation. The Commission's July guidelines make the same distinction: providers are responsible for machine-readable marking, while deployers have separate disclosure responsibilities for certain public-facing uses of generated content. For media organisations across the Balkans that serve audiences in the EU or work with European partners, those details are likely to become increasingly important. Claude Code is part of the change too Anthropic's decision is also relevant outside publishing. The company says watermarking operates at the model level, which means Claude-generated text can carry the signal regardless of whether it comes from the main Claude interface, the API or products such as Claude Code. That does not necessarily mean every line of source code will display a visible A…