Anthropic updated a Claude Help Center article on August 10, 2026, setting out how it will place machine-readable marks in text and files generated by its models. The document confirms the company has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and that marking applies wherever Claude is offered rather than only inside the European Union.

The support page carries a plain administrative title, "How Claude marks AI-generated content," and sits in the privacy and legal section of Anthropic's help centre. Its content is less administrative than its placement suggests. According to Anthropic, the company signed the code as a provider of both generative AI models and generative AI systems, a dual classification that matters because Article 50 splits its duties between the two roles.

The commitment has a hard date attached. Anthropic states that Claude models launched on or after August 2, 2026 support marking at launch, with generated text carrying embedded watermarks and generated files carrying digitally signed provenance metadata where supported. That date is not an engineering milestone. It is the day transparency obligations under Regulation (EU) 2024/1689 became applicable, eight days before the support page was updated.

One wrinkle sits inside the document itself. The summary box at the top of the article refers to Claude models launched in the EU on or after August 2, 2026. The detailed section further down drops the geographic qualifier and refers simply to models launched on or after that date. The scope statement elsewhere in the same article resolves the ambiguity in the wider direction: according to Anthropic, "Marking will apply to output from supported models wherever Claude is offered, worldwide."

Two techniques, applied at different layers

Anthropic describes two complementary methods. Neither is new to the industry. What is notable is where each one is applied.

Embedded watermarks in text

The first technique weaves an imperceptible signal directly into generated text. According to Anthropic, the mark does not change the meaning, quality, or readability of a response, and remains invisible during normal reading.

The technical detail that carries the most consequence concerns persistence. Anthropic states that because the watermark forms part of the text, "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." The company adds that watermarking is applied at the model level, which means the signal is present no matter which Claude product or surface produced the text.

Model-level application is the load-bearing choice here. It removes any possibility of a product surface opting out, and it means a marketer pasting Claude output into a content management system, a brief, or an ad copy field carries the signal along without any deliberate step. It also means the mark is not tied to a region, an account tier, or a compliance setting.

Signed provenance metadata on files

The second technique covers files rather than prose. When Claude generates a supported file type, Anthropic says the output receives signed provenance metadata following the C2PA open standard, maintained by the Coalition for Content Provenance and Authenticity. The document names .svg, .png and .jpg as examples of supported types.

C2PA attaches cryptographically signed manifests describing an asset's creation and editing history. The standard is already widely deployed across the content supply chain. Cloudflare integrated Content Credentials into its image optimisation service in early 2025, preserving signatures through transformations. YouTube uses C2PA metadata alongside SynthID watermarks to apply automatic AI labels even when a creator does not select the disclosure option during upload. TikTok applies Content Credentials to assets produced through its Symphony creative suite.

According to Anthropic, the presence of a signed metadata label signals that a file was processed by Claude and allows detection of whether the file has been tampered with. The two techniques therefore answer different questions. The text watermark answers whether a passage came from a Claude model. The file manifest answers whether a file has been altered since Claude touched it.

Five products, three clouds, every region

The scope section of the document is specific about surfaces. Marking covers output from supported models across Claude Platform, the company's API, plus the Claude consumer application, Claude Code, Claude Cowork and Claude Tag. Embedded watermarks apply to all generated text. Provenance metadata applies wherever Claude supports file processing.

That product list has direct relevance to marketing operations. Claude Cowork, launched on January 12, 2026 as a research preview for Max subscribers, extended file automation from developers to general knowledge workers on macOS. Anthropic's own marketing operations team subsequently documented rebuilding its weekly metrics reporting and campaign infrastructure around the product. Claude's file creation capability, introduced in 2025, produces spreadsheets, documents and visualisation files. Every artefact leaving those workflows now falls inside the marking scope.

Cloud distribution is treated separately and unevenly. Anthropic states that embedded watermarks apply when supported Claude models are accessed through AWS, Google Cloud, or Microsoft Foundry. Signed provenance metadata carries a qualification: it may not be supported on every platform, depending on the features each platform offers. Enterprises running Claude through a hyperscaler therefore receive a consistent text signal and an inconsistent file signal, with the variance determined by the intermediary rather than by Anthropic.

The regional decision is the one that extends furthest beyond legal necessity. The AI Act governs systems placed on the European market. Nothing in the regulation reaches an API call originating in Singapore or São Paulo. Anthropic applied the marking globally regardless.

Models released before August 2 remain in transition

Older models are not covered at launch. According to Anthropic, the law includes a transition period for models launched before August 2, 2026, and the company is working to add marking support for those models as well. No completion date appears in the document.

That transition is not open-ended. An AI Omnibus grandfathering rule gives systems already on the market until December 2, 2026 to bring marking and detection into conformity, a deadline documented alongside the Commission's July guidelines. A separate interoperability requirement for watermark detection mechanisms under Measure 3.4 of the code falls on February 2, 2027.

The practical consequence for anyone auditing content is a coverage gap measured in months. Text produced by a Claude model released in, say, early 2026 carries no mark today, and may carry none until the December deadline. Content generated before August 2, 2026 does not require retroactive marking at all, though text generated before that date and published on or after it must still be labelled by whoever publishes it.

Detection has been promised, not delivered

The gap between marking and detection is the most consequential omission in the document. Anthropic states that it is working to enable users and other third parties to detect its embedded watermarks and provenance metadata, and that details on detection mechanisms will appear in forthcoming technical documentation.

No tool exists publicly at the time of the update. No date is given. The code Anthropic signed requires signatories to offer a free detection solution operating under a zero-retention rule, and Section 1 requires providers to apply at least two machine-readable marking layers and to watermark free-form text longer than 200 tokens. The two-layer requirement appears satisfied on files, which receive both a text-adjacent signal and a signed manifest. Plain text output receives one layer.

The 200-token threshold also carves out short outputs. A headline, a meta description, an ad callout or a product feed attribute produced through Claude may fall below the length at which a statistical text watermark can be recovered reliably. Anthropic acknowledges this directly, listing very short passages among the cases where content may carry no detectable mark.

What a detected mark does not prove

The limitations section is unusually blunt for a compliance page, and it undercuts the interpretation most likely to be applied downstream.

According to Anthropic, a detected mark provides a signal that content was processed by Claude but is not fully conclusive. Two failure modes are named. Claude may not be the original author, because the model is frequently used to proofread, translate, summarise or convert files, meaning marked output can carry underlying ideas, text or data that originated elsewhere. And content may have changed after processing, having been modified, excerpted or combined with other material.

The reverse error is documented with equal care. Absence of a mark does not establish that content was not AI-generated. Anthropic lists five conditions under which its own output would go undetected: generation by a model released before marking support existed; heavy editing, paraphrasing, translation or mixing into other writing; passages too short for a reliable signal; file metadata stripped through format conversion, re-saving or screenshots; and production through a platform, feature or file type where a marking type was not supported.

Read together, those two lists describe a probabilistic signal rather than a proof of authorship. The distinction matters most in the places where such signals get used as evidence. A false negative rate that includes "paraphrased" and "translated" is a wide gate. A false positive interpretation that treats "processed by Claude" as "written by Claude" mislabels every document the model has merely edited.

The deployer duty does not move

Nothing in the marking commitment shifts the visible labelling obligation. Article 50(2) governs providers of generative systems and requires machine-readable marking of synthetic output. Article 50(4) governs deployers and requires disclosure of deep fakes and of AI-generated text published to inform the public on matters of public interest, absent human editorial control.

That separation has already produced a settled pattern in advertising. Google shifted AI ad labelling liability entirely to advertisers through its July 2026 documentation, giving advertisers a labelling control across five products while confirming that AdSense publishers carry no equivalent setting and no compliance duty of their own. Display and Video 360 and Campaign Manager 360 gained a synthetic content attestation field that cannot be unset once written. Googlesigned the transparency code on July 24, 2026. Meta followed days later, five days before Article 50 became binding.

Penalties for non-compliance reach EUR 15,000,000 or 3 percent of total worldwide annual turnover, whichever is higher. Adherence to the code is voluntary. The obligations underneath it are not.

The free labelling icon set the Commission published in June 2026 covers the deployer side of that split, with separate marks for fully AI-generated and partially AI-modified content. Advertising sits outside the lighter disclosure regime available to artistic, satirical and fictional work, which means a persuasive commercial asset built with generative tools attracts full labelling rather than the narrower treatment.

Why the marketing industry has a stake

Three effects follow for agencies, brands and publishers.

The first is auditability of internal workflows. A model-level watermark means output produced inside an agency, then handed to a client, then published, carries a signal through the chain. Provenance disputes over creative deliverables acquire a technical dimension that did not previously exist, though the reliability caveats limit how far any such signal can be pushed as evidence.

The second is the reliability question around third-party detection services. Marketing teams increasingly encounter AI-detection claims from clients, platforms and freelance marketplaces. Anthropic's own documentation states that its mark cannot confirm authorship and that its absence proves nothing. Detection services built on top of these signals inherit both limitations, whatever confidence their interfaces project.

The third is audience trust, which has been measured. A Raptive study found that suspected AI content reduces reader trust by close to half, with a 14 percent decline in purchase consideration when readers believed content was AI-generated. The IAB published an AI Transparency and Disclosure Framework in January 2026 alongside research showing a 19-point gap between advertiser assumptions and Gen Z sentiment on AI advertising. Machine-readable marking makes the suspected-AI question answerable in more cases than before, which changes the risk profile of undisclosed generative production regardless of what any regulator requires.

For anyone deploying Claude inside a product, Anthropic states that operators carry their own obligation to assess what Article 50 requires of the products and services they build, and that the company will share technical guidance on its marking and detection approach as it becomes available. The document does not indicate when.

Timeline

Summary

Who: Anthropic, publisher of the Claude family of models, acting as a signatory to the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content in its capacity as a provider of both generative AI models and generative AI systems.

What: A published marking commitment covering two techniques. Imperceptible watermarks are embedded at the model level in all generated text, and signed provenance metadata following the C2PA standard is attached to supported generated file types including .svg, .png and .jpg. Coverage spans Claude Platform, the Claude application, Claude Code, Claude Cowork and Claude Tag, plus access through AWS, Google Cloud and Microsoft Foundry, where provenance metadata support varies by platform. Anthropic documents that a detected mark is not conclusive proof of authorship and that absence of a mark does not establish human origin.

When: The Claude Help Center article was updated on August 10, 2026. Models launched on or after August 2, 2026 support marking at launch, matching the date Article 50 transparency obligations became applicable. Models released before that date fall inside a transition period with no completion date stated, against a December 2, 2026 conformity deadline for systems already on the market and a February 2, 2027 deadline for watermark-detection interoperability.

Where: Marking applies wherever Claude is offered, worldwide, rather than only within the European Union or the European Economic Area.

Why: Article 50 of Regulation (EU) 2024/1689 requires providers of generative AI systems to mark synthetic output in machine-readable form, with non-compliance carrying penalties of EUR 15,000,000 or 3 percent of total worldwide annual turnover, whichever is higher. Anthropic states that the marking supports transparency about content origin and compliance with its legal obligations. For advertisers, agencies and publishers, the change makes a growing share of generative output technically identifiable while leaving the visible disclosure duty under Article 50(4) with the party that publishes.