Anthropic said Aug. 11 that newly launched Claude models will add machine-readable marks to AI-generated text and supported files, extending a European Union transparency requirement across its products worldwide. The policy gives businesses a new provenance signal for Claude output, but Anthropic cautioned that the mark is not definitive proof that Claude created a piece of content.
The company detailed the change in an updated support document after signing the European Union’s voluntary Code of Practice on Transparency of AI-Generated Content. TechCrunch independently reported the rollout Tuesday.
What Claude will mark
Claude models launched in the European Union on or after Aug. 2 will support the marking system at release, Anthropic said. Generated text will contain an imperceptible watermark embedded at the model level. Supported image files, including PNG, JPG and SVG formats, will carry signed provenance metadata using the C2PA standard.
Anthropic said the text watermark will travel when content is copied and pasted and may remain detectable after some editing. Because the marking happens within the model, the company said it will apply across supported access points, including Claude, Claude Code, Claude Cowork, Claude Tag and the Claude Platform API. Embedded text marks are also intended to cover supported Claude models offered through Amazon Web Services, Google Cloud and Microsoft Foundry. Availability of signed file metadata may differ by platform.
The company is still transitioning models introduced before Aug. 2. It also said detection tools for customers and third parties are forthcoming, leaving organizations without a broadly available way to check the new signal immediately.
A provenance signal, not an authorship verdict
Anthropic’s documentation draws a sharp boundary around what the technology can establish. A detected mark may show that content passed through Claude, but it does not prove that Claude originated every idea or sentence. Proofreading, translation and summarization can produce marked text even when a person supplied the underlying material.
The reverse is also true: An absent mark does not establish that content is human-made. Anthropic said detection may fail when text comes from an older or unsupported model, is too short, has been substantially rewritten or translated, or loses file metadata. The company therefore advises against treating the system as a stand-alone authorship test.
That distinction matters for publishers, educators, employers and software vendors. Policies that automatically punish people when a detector finds a watermark could misclassify legitimate editing or translation work. Procurement and compliance teams will need to decide how the signal fits with human review, audit logs and disclosure rules.
EU rules are reshaping global product design
The European Commission says Article 50 transparency obligations under the AI Act took effect Aug. 2. Providers must make generated or manipulated audio, images, video and text detectable in a machine-readable format when technically feasible. The code offers companies a voluntary framework for demonstrating compliance, although the underlying legal obligations are mandatory.
Anthropic’s worldwide implementation shows how a regional rule can influence the default architecture of a global AI service. Enterprise customers outside Europe will encounter the same embedded signal even when their immediate use is not governed by the EU requirement.
For executives, the operational question is broader than whether content carries a mark. Organizations using Claude in customer communications, code, research or internal workflows should map where marked output enters their systems, preserve relevant provenance data and avoid equating detection with misconduct. Anthropic said enterprise developers remain responsible for assessing their own Article 50 obligations.