Anthropic begins labeling all of Claude's output globally as EU transparency regulations come into force.

      Anthropic has started adding machine-readable markers to the text and images generated by its Claude models, implementing them globally instead of just in Europe, following its adoption of a European Union code of practice related to new transparency regulations that became effective on August 2, 2026. The company confirmed in a support document that it agreed to the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-generated Content as a developer of both generative AI models and systems.

      This move positions Anthropic as one of the first major AI developers to concretely outline its plans to adhere to a regulation that will eventually mandate most AI-generated content circulating in the EU to be recognized as such. The responsibilities outlined go beyond a single company, dividing them between two categories.

      Providers, which are the companies that develop and supply AI models, must ensure that the outputs of their systems are identifiable as artificial. Deployers, the larger group of organizations that utilize those systems to create and publish content, have separate obligations to disclose this information. This distinction clarifies who is accountable and what aspects they are responsible for.

      Regarding how Anthropic marks Claude's outputs, the company uses two methods. For text, it embeds an unobtrusive watermark that it claims does not impact the meaning, quality, or readability of the response, and this watermark persists when the text is copied, pasted, or lightly edited. For images, it adds signed provenance metadata to compatible formats like .png, .jpg, and .svg, in accordance with the open C2PA standard implemented by the Coalition for Content Provenance and Authenticity, which records the source of the files and indicates if they have been altered. This system has also been adopted by companies like Adobe, the BBC, and several camera manufacturers to track the origin of images.

      Anthropic stated that Claude models released in the EU on or after August 2, 2026, will feature this marking at launch, while those introduced before that date are being updated to support it. This marking will be incorporated into all of the company’s products, including its developer API, Claude apps, Claude Code, and its enterprise deployments, and it will be applied globally, not just for European users.

      Anthropic included an important caveat: a detected mark indicates only that content “may have been processed by Claude,” and a lack of a mark does not mean that the content was not AI-generated, as outputs from older models or extensively edited texts might not carry marks. The company plans to release additional technical documentation for third parties to detect its markers.

      A machine-readable mark differs from a visible label. The watermark and metadata are signals interpretable by detection tools, not notices visible to the reader. Converting this into an easily understandable disclosure for human comprehension is another step, primarily the responsibility of the organizations that publish the content instead of the model provider.

      In comparison to other AI firms, Anthropic is not the first significant developer to start marking content. In May 2026, OpenAI joined the C2PA coalition and collaborated with Google to incorporate the SynthID watermark into its image outputs, showcasing a tool that enables users to confirm if an image was generated by its models. SynthID, developed by Google DeepMind, embeds invisible markers into images designed to remain intact through screenshots, resizing, and compression. Although C2PA had garnered over 6,000 members and affiliates by early 2026, wider industry adoption has been inconsistent.

      What distinguishes Anthropic's initiative is its focus on watermarking text, the primary output format of Claude, which is the most challenging to mark effectively, and its framing of this work as compliance with a specific legal requirement rather than simply a voluntary provenance initiative.

      The obligations derive from Article 50 of the AI Act, a risk-based regulation governing artificial intelligence in the EU, which came into force in 2024 and is being implemented over several years. The transparency provisions of Article 50 became applicable on August 2, 2026. Paragraph 2 specifically addresses providers, mandating that the outputs of generative AI systems be "marked in a machine-readable format and detectable as artificially generated or manipulated," requiring that technical solutions used are "effective, interoperable, robust and reliable as far as this is technically feasible."

      There are narrow exceptions to this requirement, including situations where an AI system only provides assistive or minor editing functions and does not significantly alter the input. Paragraph 4 pertains to deployers, thereby extending responsibilities to organizations utilizing AI tools. Any entity deploying a system that produces a deepfake—such as a synthetic or manipulated image, audio, or video mimicking real people, objects, or events—must reveal that the content is artificially generated. Deployers sharing AI-generated text to inform the public about matters of public interest must also disclose this unless the content has undergone human review and a person or organization has editorial responsibility for it. Material that is distinctly artistic, creative, satirical, or fictional is