Skip to content
Beyond Prompt AI Studio

Law & regulation

Anthropic marks AI text worldwide - and its own documentation explains exactly how the mark disappears

August 12, 2026 · 12 min read · Beyond Prompt AI Studio

EU AI ActClaudeComplianceLaw

Anthropic announced yesterday, 11 August 2026: every Claude model launched since 2 August 2026 embeds an imperceptible mark into generated text; image files get signed provenance metadata under the open C2PA standard. Explicitly framed as a response to the EU AI Act's Code of Practice on Transparency of AI-Generated Content - specifically the machine-readable marking duty under Article 50(2), which we broke down here in detail last week. Notably, Anthropic applies the marking worldwide, not just for EU users. This analysis didn't stop at the announcement - it read Anthropic's own support documentation, which openly states the everyday conditions under which the mark disappears. That's the part that matters for companies using Claude to produce their own content.

Key points at a glance

  • Since 2 August 2026, Anthropic invisibly marks generated text and signs image files with C2PA metadata - worldwide, not just in the EU, even though the measure is explicitly justified as a response to the EU AI Act.
  • Per Anthropic's own support documentation, the text mark disappears entirely under heavy editing, paraphrasing, translation, or short excerpts. Image files' C2PA metadata can be removed by re-saving, converting formats, or taking a screenshot.
  • Direct quote from the documentation: 'A detected mark is not proof of AI authorship.' Even when the mark survives, its presence alone says nothing about who is ultimately responsible for the content or whether it was altered afterward.
  • Older Claude models (launched before 2 August 2026) don't mark their output yet - Anthropic says it's 'working on' retrofitting them, with no committed date. Anyone using an older model currently gets no marking at all, regardless of their own need for it.
  • The Article 50(2) marking duty legally falls on the provider of the generative system (here: Anthropic), not automatically on the company that subsequently edits and publishes the generated text. Anthropic's own technical limits show why that distinction matters in practice: the normal editorial process - trimming, rephrasing, pasting into a document - is exactly the kind of editing that destroys the mark.
  • In practice that means: a company with its own labelling duty for published AI content (for instance under Article 50(4) for matters of public interest) can't rely on Anthropic's marking to cover that duty - deliberate, independent disclosure remains necessary.

What Anthropic announced yesterday

The mechanism in brief: Claude models introduced since 2 August 2026 embed a statistical signal into generated text that's imperceptible to readers. Per Anthropic, it survives copy-pasting and partially survives editing. For image files (including SVG, PNG, JPG), Anthropic instead uses signed provenance metadata under the open C2PA standard - the same standard camera makers and news agencies use to attribute the origin of digital content. Per the support documentation, coverage includes the Claude Platform (API), Claude itself, Claude Code, Claude Cowork, and access via AWS, Google Cloud, and Microsoft Foundry.

The justification is unambiguously regulatory: Anthropic explicitly cites the EU AI Act's Code of Practice on Transparency of AI-Generated Content - exactly the duty under Article 50(2) we broke down here last week, including the important nuance that it carries a transition period until 2 December 2026 for generative systems already on the market before 2 August 2026, but applies immediately to new systems. What's notable about Anthropic's implementation: the marking applies worldwide, not just to European users - a company fulfilling a regionally grounded duty globally, rather than building a geographic distinction into the product. That's a pattern that could pressure other vendors to follow suit, even for users outside the EU.

The part only the documentation reveals

The announcement itself reads like a settled compliance story. Anthropic's support documentation, one level below the press release, is markedly more cautious. It states plainly that no mark is detectable when content has been 'heavily edited, paraphrased, or translated, is too short, had file metadata stripped, or came from an unsupported surface'. For image files, the C2PA metadata can be removed by re-saving, converting formats, or simply taking a screenshot - not as a workaround, but as an inherent technical property of the method itself.

A second sentence from the same documentation is even more direct: 'A detected mark is not proof of AI authorship.' Even in the best case, where the mark survives, its mere presence says nothing about who is responsible for a piece of content or whether it was altered after creation. Anthropic itself warns in the documentation against treating a marking result as a verdict - more accurately described as a coarse filter, not evidence.

Why this is the real point for companies

This is where the loop closes with our analysis from last week. The marking duty under Article 50(2) legally falls on the provider of a generative AI system - in this case Anthropic, which fulfils its own duty with this announcement. For a company using Claude to produce content, though, that's only half the story. The normal editorial process - an AI draft gets trimmed, rephrased, dropped into a template, translated into German, blended with human-written text - is exactly the kind of editing that, per Anthropic's own documentation, makes the mark disappear.

That means the final published text a company actually puts online often carries no detectable mark in practice - regardless of whether Claude originally applied one. Concluding 'we use Claude, so we automatically meet our own labelling duties' confuses two different things: the provider's technical measure to fulfil its own duty, and the content-level disclosure duty that can fall on a company itself - for instance under Article 50(4), when publishing AI-generated or AI-assisted text on matters of public interest. That second duty doesn't take care of itself just because the vendor in use has fulfilled its own.

An additional, barely noticed gap: older models

One more detail from the documentation gets little mention in coverage so far: marking applies to models introduced since 2 August 2026. For older models, Anthropic says it's 'working on' a retrofit, without naming a firm date. A company that, for good reasons - cost considerations or established workflows - still uses an older Claude model currently gets no marking at all, regardless of its own need for it. Anyone leaning on 'Claude marks that automatically' in their own compliance planning should first check exactly which model version is actually in use.

What this means in practice

Nothing in this analysis argues against Anthropic's measure - quite the opposite: a worldwide rather than EU-only rollout, and an honest, publicly documented account of its own limits, are both notably transparent for a compliance measure of this kind. The point is not to treat your own responsibility as settled just because the vendor has settled its own.

  • Check your own labelling duties separately from Anthropic's marking: if you publish AI-generated or AI-assisted content on matters of public interest, Article 50(4) requires your own, deliberate disclosure - regardless of whether a technical mark is still detectable in the text.
  • Review your own editorial process against this question: does an AI draft get edited, translated, or embedded heavily enough that any mark is lost anyway? If so, your own disclosure shouldn't implicitly rely on the provider's technical measure.
  • Check which specific Claude model version is actually in use: older models not yet retrofitted currently carry no marking at all, regardless of Anthropic's announcement.
  • Never treat a marking detection result as proof, in either direction: a detected mark doesn't conclusively prove AI authorship, and a missing mark doesn't prove content wasn't AI-generated.

The real value of this analysis isn't a warning against Anthropic's measure - it's a clarification: a provider duty and a company duty can be triggered by the same regulation and still remain two separate things, each requiring its own fulfilment. Anyone treating the two as one and the same only notices the difference once a regulator or a customer asks precisely that question.

Frequently asked questions about Claude's watermarks and your own labelling duty

Does our company automatically meet its labelling duties by using Claude?

Not automatically. Anthropic's marking fulfils Anthropic's own duty as the provider of the generative system. Your own labelling duties - for instance under Article 50(4) for AI-generated content on matters of public interest - fall on the publishing company separately and remain in place, regardless of whether Anthropic's technical mark is still detectable in the published text.

Why does the mark disappear so easily?

Per Anthropic's own support documentation, a mark is no longer detectable after heavy editing, paraphrasing, translation, in short excerpts, or when file metadata has been stripped. That's not a workaround - it's an inherent technical property of the method, and it affects practically every normal editorial process in which an AI draft gets further processed.

Does the marking also apply to older Claude models we're already using?

Only to models introduced since 2 August 2026. For older models, Anthropic says it's working on a retrofit, with no committed date. If you're using an older model, factor that into your own compliance planning rather than assuming automatic marking.

Can we use a detected mark as proof that a text was AI-generated?

Only to a limited extent. Anthropic explicitly states in its own documentation that a detected mark is not proof of AI authorship - it only indicates Claude may have been involved, and says nothing about responsibility or subsequent alterations.

Want to know which of your own labelling duties apply to your AI-assisted content - regardless of your provider's marking?