For organisations using Claude in products, workflows or content operations, the change separates a provider’s technical marking from a deployer’s potential disclosure duties. It also adds a concrete example to the wider practice of AI Trust— determining when a synthetic-content signal exists, what it conveys and who remains accountable for publication decisions. These considerations form part of broader AI Act compliance and governance guidelines for AI providers, deployers and downstream providers.
Anthropic says Claude models launched in the EU on or after 2 August 2026 will support marking from launch, while support for earlier models is in progress. The company says the measures apply to supported output across the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag, and wherever Claude is offered worldwide.
Claude combines two marking methods. Generated text receives an imperceptible embedded watermark. Supported files, including PNG, JPG and SVG, receive signed provenance metadata based on the C2PA open standard. Anthropic says the text watermark does not change a response’s meaning, quality or readability and may persist after copying, pasting or some editing. The metadata can indicate that a file was processed by Claude and whether it has been tampered with.
The company also sets out important limits. A detected mark indicates that material may have been processed by Claude, rather than proving full provenance. Absence of a mark does not show that content was not AI-generated: editing, paraphrasing, translation, short passages, metadata stripping and unsupported surfaces can reduce detectability. This distinction matters for detection of AI-generated content, deepfakes and other manipulated content.
Under Article 50, providers of generative AI systems must ensure output is marked in a machine-readable format and detectable as artificially generated or manipulated. The regulation calls for solutions that are effective, interoperable, robust and reliable as far as technically feasible, with limited exceptions such as assistive functions or standard editing that does not substantially alter input data.
The regulation also creates disclosure duties for deployers. Organisations deploying systems that produce deepfakes must disclose that the content is artificially generated or manipulated. Organisations publishing AI-generated text to inform the public on matters of public interest have disclosure duties too, subject to Article 50 conditions, including an exemption where content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.
A watermark or provenance record can therefore be useful evidence in a content pipeline, but it is not automatically a visible disclosure for an audience. Organisations should connect such controls with their AI privacy and data-governance practices and assign ownership of publishing decisions. This includes considering GDPR transparency requirements, information obligations and additional transparency obligations that may apply to particular professional activities.
The European Commission published the voluntary transparency code on 31 July 2026. It reported that about 190 organisations had signed it by the end of July, including Anthropic among the 82 signatories to the provider-focused section. Two task forces are scheduled to begin in September to exchange implementation practices and advance the technical state of the art.
The code provides practical guidelines for AI providers, deployers and general-purpose AI models, while the AI Office and EU institutions continue developing implementation materials. Organisations should distinguish voluntary code commitments from binding AI Act obligations and monitor any related documentation, including guidance relevant to notified bodies and national market surveillance authorities.
For enterprises deploying Claude or comparable systems, the immediate question is not only whether output carries a technical mark. They need to identify where generative AI is used, determine which output is distributed in the EU, and define who assesses and carries out any required disclosure. These are also questions of accountability, access control and auditability in governance for agentic AI.
Anthropic has said that further documentation on detection mechanisms is forthcoming. Organisations should monitor that guidance, Commission updates to the code and task-force outputs, and the retention of signals when content is edited, converted or republished. Teams responsible for synthetic media should also consider the verification and governance issues discussed in Nemko Digital’s coverage of deepfake-related TrustOps.
Organisations should also track how marking requirements interact with generative AI systems, interactive AI systems and other transparency measures, including any implementation guidance, official journal updates or materials issued by the AI Office. Where systems are integrated into larger products, downstream providers may need documentation showing how technical marks, disclosures and human oversight are handled.
The rollout makes content-origin signals available in supported Claude workflows. It does not remove the need for each organisation to assess its role under the EU AI Act or determine when Article 50 transparency duties apply. For legal entities and natural persons involved in publication, that assessment should form part of a documented compliance process rather than rely on a watermark alone.