Anthropic will introduce machine-readable watermarks to all text generated by new Claude models from August 2, 2026, aiming to meet EU AI Act transparency rules and provide a technical signal for identifying AI-generated content worldwide
Anthropic has announced that, beginning August 2, 2026, all text generated by new Claude models will carry invisible, machine-readable watermarks. This change is designed to comply with Article 50 of the European Union's AI Act, which requires providers of generative AI systems to mark synthetic content for transparency. While the regulation applies within the EU, Anthropic will implement the watermarking system globally, affecting users in all regions and across major cloud platforms.
Watermarking Claude-Generated Text
The watermarking system will embed an imperceptible signal directly into the output text at the model level. According to Anthropic's technical documentation, the watermark is designed to persist through common user actions such as copying, pasting, and some forms of editing. The company states that the mark will be present in responses generated via the Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as through cloud providers including AWS, Google Cloud, and Microsoft Foundry. The approach aims to provide platforms, organizations, and third parties with a technical means to assess whether a given text was produced by a Claude model.
Anthropic has committed to publishing technical details and detection tools for its watermarking method. These tools are intended to help users and external parties verify the presence of a Claude watermark in supported content. However, the company cautions that the watermark should not be treated as definitive proof of authorship, since Claude models can process and modify human-written material, and heavy editing or short passages may weaken or remove the signal.
Digital Provenance for Files
For supported files, including common image formats, Anthropic will use a separate system based on the Coalition for Content Provenance and Authenticity (C2PA) standard. This approach attaches signed provenance metadata to generated files, recording information about the file's origin and any subsequent modifications. A valid C2PA record can indicate that a file was processed by Claude and whether its provenance data has been altered. However, file metadata can be lost during format conversion, screenshotting, or re-saving, making it less robust than text watermarking for some workflows.
The distinction between text watermarks and file provenance reflects different technical challenges. While a text watermark can survive copying and some editing, file metadata is more vulnerable to loss through ordinary processing. Anthropic notes that older Claude models and unsupported file types will not carry the new markings, and the absence of a watermark does not guarantee human authorship.
Regulatory and Technical Context
The introduction of watermarking follows Anthropic's signing of the European Union's Code of Practice on AI-generated content. The company's decision to apply the system worldwide, rather than limiting it to the EU, reflects the growing expectation for transparency in generative AI outputs. For businesses, developers, and content platforms, the watermark provides an additional technical signal for tracking and managing AI-generated material, potentially aiding compliance and content moderation efforts.
Anthropic's announcement does not specify the exact technical method used for watermarking, nor does it provide independent evaluation of the system's robustness against deliberate removal or adversarial editing. The company plans to release further technical documentation as the system matures. As with other watermarking and provenance schemes, the effectiveness of the approach will depend on adoption, detection accuracy, and the ability to withstand attempts to evade or strip the marks.
Anthropic's watermarking system is scheduled to take effect on August 2, 2026, coinciding with the enforcement of the relevant provisions of the EU AI Act. The company has not disclosed the number of Claude model versions affected, but states that all new models released from that date will include the marking system by default. The technical documentation and detection tools are expected to be published in advance of the rollout to allow for integration and testing by external stakeholders.
Watermarking in generative AI refers to the practice of embedding a hidden, machine-detectable signal within generated content to indicate its synthetic origin. Unlike visible labels or disclaimers, invisible watermarks are designed to persist through common user actions and can be detected using specialized tools. However, watermarking is not foolproof: marks can be weakened or lost through editing, reformatting, or adversarial manipulation, and the absence of a watermark does not guarantee human authorship. Watermarking is one of several technical approaches being explored to support transparency, provenance, and accountability in the deployment of large language models and other generative AI systems.
