Anthropic to Watermark Claude AI Text Globally Under EU AI Act Transparency Code

As AI-generated content floods digital platforms, the question of who — or what — wrote any given piece of text is becoming one of the defining challenges of the information age. Anthropic's new marking system attempts to answer that question, but introduces complications that publishers, writers, and policymakers will need to reckon with carefully. The stakes extend far beyond European borders, and the implications touch every organisation producing or publishing content with AI assistance.

What Anthropic Committed to Under the EU AI Act

Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content as a provider of both generative AI models and generative AI systems. The European Commission counted approximately 190 signatories by the end of July 2026. Google, Meta, Microsoft, Mistral, and OpenAI joined Anthropic on the provider section of the code, which specifically covers machine-readable marking and detection.

Claude models launched in the EU on or after August 2, 2026 support machine-readable marking at launch. Models released before that date fall under a transition period, and Anthropic says it is working to add marking support to those earlier versions as well. No specific date has been attached to that rollout.

The marking applies to outputs from supported models across a wide range of platforms — including the API, the Claude applications, Claude Code, Cowork, and Tag — as well as versions accessed via AWS, Google Cloud, and Microsoft Foundry.

Understanding how these commitments sit within the broader framework of EU data and digital regulation is essential context for publishers and compliance teams navigating their obligations.

This stands in contrast to OpenAI's trajectory. In September 2024, OpenAI scrapped its own watermarking plans after an internal company survey found that nearly 30% of ChatGPT users would use the platform less if watermarking were added — a revealing indication of user resistance that Anthropic has chosen to push through regardless.

The Scope of the Commitment

The breadth of Anthropic's commitment is notable. Rather than limiting marking to a specific product or region, the obligation attaches to the model itself. Wherever a supported Claude model is accessed — whether through a consumer application, a cloud provider's infrastructure, or the raw API — the marking system applies. This has meaningful implications for enterprise customers who integrate Claude into their own products and workflows, since the watermark travels with the output rather than being applied at the point of publication.

---

How the Marking System Works — and Where It Falls Short

Text from supported Claude models receives an embedded watermark that Anthropic says does not change the meaning, quality, or readability of a response. Because the mark lives inside the text itself, it survives copying and pasting and may persist through some forms of editing. Generated image files in .SVG, .PNG, and .JPG formats carry signed metadata following the C2PA open standard, which records how each file was created and flags tampering.

However, Anthropic is transparent about the system's limitations. A detected mark indicates only that Claude may have processed the content — not that Claude authored it. Proofreading, translation, summarising, and file conversion may all trigger a mark even when the original ideas and words came entirely from a human writer. Conversely, content from Claude may carry no detectable mark if it was produced by an older model, was heavily edited afterward, or was simply too short to produce a clear signal.

A Genuine Evidentiary Gap

This creates a genuine evidentiary gap. A writer who drafts their own copy and runs it through Claude for a light cleanup ends up with marked text. So does a translator working from someone else's original article. Any policy or platform that treats a positive watermark detection as proof of AI authorship would be drawing a conclusion the technology cannot support.

For publishers and editorial teams, this is perhaps the most operationally significant point in the entire framework. The presence of a watermark is a signal — not a verdict.

Roger Montti's analysis of Article 50's four exemptions adds another layer of complexity. Systems that assist only with standard editing — without substantially altering the input or its meaning — can fall outside the marking requirement entirely. Published text that has undergone qualifying human review or editorial control may also be exempt from disclosure obligations when someone holds editorial responsibility for it. This means a Claude watermark can appear on copy that its publisher has no legal obligation to label.

What the C2PA Standard Means in Practice

The C2PA (Coalition for Content Provenance and Authenticity) standard used for image metadata is an open, industry-backed specification designed to create a verifiable chain of custody for digital content. When Anthropic embeds C2PA metadata in a generated image, it records not just that AI was involved, but structured information about the generation process. Compatible platforms and tools can read this metadata and surface it to users. Incompatible tools will simply ignore it — meaning the provenance signal disappears entirely depending on where the file ends up.

This is not a flaw unique to Anthropic's implementation. It reflects the current state of the broader ecosystem. C2PA adoption among content management systems, social platforms, and publishing tools remains uneven, and a provenance record that is stripped or ignored at the point of upload offers limited practical protection.

---

Detection Tools, Durability Concerns, and What Comes Next

The Unpublished Detection Mechanism

Perhaps the most significant unresolved issue is that Anthropic has not yet published its detection mechanism. The company has stated it will support detection by users and third parties and will publish technical documentation, but has not specified what that access will look like or when it will arrive.

Alex Cui, CTO and co-founder of AI detection company GPTZero, published a technical analysis arguing that text watermarks can be defeated. Cui noted that watermarks can be lost through intense paraphrasing and that free tools have already bypassed Google DeepMind's SynthID watermarking system. His tests ran against other systems rather than Anthropic's own version, because that version has not been made publicly available. GPTZero has argued since 2024 that watermarking does not eliminate the need for independent AI detection tools that score text by pattern rather than checking for embedded marks.

Jonas Geiping, who leads a machine learning safety group at the ELLIS Institute Tübingen and studies watermarking, offered a more measured view. Paraphrasing can remove a watermark, he said — but not every paraphrase will. Stripping a mark from a long document requires more than a light editing pass, because enough of the original phrasing must be replaced to break the signal. The practical takeaway is that casual or opportunistic removal is harder than critics sometimes suggest, while determined and systematic removal remains achievable.

Google's approach offers a point of comparison. The company keeps SynthID verification inside its own products and expanded that system to Google Search in May 2026. Anthropic's stated intention to open detection to third parties would represent a more transparent posture — if and when that access materialises. For a deeper look at how these kinds of systems fit within the broader risks and challenges AI presents for businesses, the governance questions around verification and accountability are closely related.

The Invisibility Problem

AI watermarks operate invisibly — only revealing themselves when someone knows exactly where and how to look. This creates an asymmetry of information that regulators, platforms, and individual users are not equally equipped to navigate. A sophisticated publisher with dedicated technical resources can potentially build detection into their content pipeline. An independent journalist or small editorial team almost certainly cannot — at least not until Anthropic publishes accessible tooling.

This gap matters because the EU AI Act's transparency obligations are premised on disclosure being practically achievable. If detection remains locked behind unpublished documentation, the accountability the code is designed to create exists on paper but not in practice.

You can find further background on what artificial intelligence is and how it works, which provides useful grounding for readers less familiar with the underlying technology shaping these regulatory debates.

---

What This Means for Publishers, Writers, and Digital Professionals

For anyone producing content with AI assistance, the implications are immediate and practical.

First, a positive watermark detection result is not evidence of AI authorship. Writers and editors who use Claude for any processing step — editing, translation, or summarisation — should be aware their work may carry a mark regardless of how much original human thought it contains. Internal style guides and editorial policies should reflect this explicitly.

Second, detection tools from Anthropic are not yet publicly available, meaning organisations cannot independently verify whether content has been flagged. Waiting for published technical documentation before building internal policies around watermark detection would be prudent. Building policy on an inaccessible verification mechanism creates compliance risk rather than reducing it.

Third, the exemptions built into Article 50 mean that editorial responsibility and qualifying human review can remove the disclosure obligation even when a watermark is present. Publishers investing in clear editorial workflows and oversight processes now will be better positioned as regulations and enforcement expectations evolve across markets beyond the EU.

Fourth — and this point deserves emphasis — the global application of the marking system means that organisations operating entirely outside the EU are nonetheless affected. If your team uses Claude via the API, AWS, Google Cloud, or Microsoft Foundry, and your models are among those with marking support, your outputs may carry watermarks regardless of where you are headquartered or where your audience is located. Regulatory compliance framed as a European concern is, in practice, a global operational reality.

For further reading on how content provenance standards are developing, the C2PA specification published by the Coalition for Content Provenance and Authenticity provides the technical foundation underlying Anthropic's image marking approach.