Blab AI
Anthropic to Embed Watermarks in Claude Outputs to Meet EU AI Act Transparency Rules
On 2 August 2026 the European Union’s Artificial Intelligence Act (AIA) will take effect, obligating providers of general‑purpose AI systems to mark any content they generate or edit. Anthropic, the maker of the Claude family of large language models, has released guidance explaining how its new models will satisfy the law’s transparency requirements.
According to the guidance, every Claude model launched on or after the August 2 deadline will embed an invisible watermark in text and attach provenance metadata to image files. The text watermark is created by subtly biasing the token‑selection process during generation. When a piece of text is later examined, the statistical pattern left by the bias can be detected, revealing that Claude produced it. Anthropic asserts that the watermark does not alter the meaning or quality of the output.
The company notes that text fragments shorter than 200 tokens are exempt from watermarking, because the token‑level signal would be too weak to be reliably detected. It also warns that direct quotes from a user’s prompt may be watermarked inadvertently, and that the absence of a watermark does not prove that content was not AI‑generated.
For images, Anthropic will attach a provenance certificate to files in SVG, PNG and JPG formats. The certificate follows the Coalition for Content Provenance and Authenticity (C2PA) standard, creating a cryptographically signed record of the file’s origin that is invalidated if the image is altered after creation. While the guidance does not mention image watermarking—another requirement of the Code of Practice—the company indicates that it plans to add that feature in the future.
Anthropic also says it will extend watermarking and provenance support to existing Claude models by 2 December 2026, as required by the AIA. The guidance clarifies that the marking will apply worldwide, not only to users in the EU.
Anthropic is not the first AI provider to announce compliance measures. OpenAI and Google had published similar watermarking strategies earlier in the year. Because the EU law applies extraterritorially, any provider that offers AI services to EU users must meet the transparency obligations. The AIA’s transparency rules apply to high‑risk and general‑purpose AI systems. While the law does not impose fines for non‑compliance, it requires that providers offer tools for detecting watermarks. Anthropic says it will release detection tools in forthcoming technical documentation.
The guidance is part of Anthropic’s broader effort to align its products with the EU’s regulatory framework. The company’s public‑benefit corporation status and its focus on safety have positioned it as a prominent player in the generative‑AI market.
In summary, Anthropic’s new Claude models will embed invisible watermarks in text and attach C2PA provenance certificates to images. The measures will be applied globally from 2 August 2026, with support for older models added by 2 December 2026. Anthropic will provide detection tools later in the year, fulfilling the EU AI Act’s transparency requirements.
According to the guidance, every Claude model launched on or after the August 2 deadline will embed an invisible watermark in text and attach provenance metadata to image files. The text watermark is created by subtly biasing the token‑selection process during generation. When a piece of text is later examined, the statistical pattern left by the bias can be detected, revealing that Claude produced it. Anthropic asserts that the watermark does not alter the meaning or quality of the output.
The company notes that text fragments shorter than 200 tokens are exempt from watermarking, because the token‑level signal would be too weak to be reliably detected. It also warns that direct quotes from a user’s prompt may be watermarked inadvertently, and that the absence of a watermark does not prove that content was not AI‑generated.
For images, Anthropic will attach a provenance certificate to files in SVG, PNG and JPG formats. The certificate follows the Coalition for Content Provenance and Authenticity (C2PA) standard, creating a cryptographically signed record of the file’s origin that is invalidated if the image is altered after creation. While the guidance does not mention image watermarking—another requirement of the Code of Practice—the company indicates that it plans to add that feature in the future.
Anthropic also says it will extend watermarking and provenance support to existing Claude models by 2 December 2026, as required by the AIA. The guidance clarifies that the marking will apply worldwide, not only to users in the EU.
Anthropic is not the first AI provider to announce compliance measures. OpenAI and Google had published similar watermarking strategies earlier in the year. Because the EU law applies extraterritorially, any provider that offers AI services to EU users must meet the transparency obligations. The AIA’s transparency rules apply to high‑risk and general‑purpose AI systems. While the law does not impose fines for non‑compliance, it requires that providers offer tools for detecting watermarks. Anthropic says it will release detection tools in forthcoming technical documentation.
The guidance is part of Anthropic’s broader effort to align its products with the EU’s regulatory framework. The company’s public‑benefit corporation status and its focus on safety have positioned it as a prominent player in the generative‑AI market.
In summary, Anthropic’s new Claude models will embed invisible watermarks in text and attach C2PA provenance certificates to images. The measures will be applied globally from 2 August 2026, with support for older models added by 2 December 2026. Anthropic will provide detection tools later in the year, fulfilling the EU AI Act’s transparency requirements.
