Anthropic will mark Claude text worldwide from its 2026 EU launches. A removal market already exists.

Anthropic has disclosed that Claude models launched in the European Union on or after 2 August 2026 will mark generated text at launch, and that the same marks will travel worldwide rather than only in Europe ^{[1] [2]}. That turns output provenance from a policy commitment into a shipped feature of the platform, and a paid removal tool for it appeared on 12 August, before the marking scheme even has a published specification ^{[13] [5]}.

The mechanism is two separate systems. Text gets what Anthropic calls an imperceptible watermark applied during generation, while supported files such as PNG, JPG and SVG assets get signed provenance metadata based on the C2PA standard ^[3]. Anthropic says the text signal can survive copying, pasting and some editing because it is woven into the generated language at the model level, not attached as metadata ^[4]. The stated scope covers Claude, Claude Code, Claude Cowork, Claude Tag and the Claude API, including access through Amazon Web Services, Google Cloud and Microsoft Foundry ^[2]. If you build on Claude through a cloud reseller, this is your surface too.

The origin is Article 50 of the EU AI Act, which requires providers of generative systems to make synthetic text, audio, image and video output machine-readable and detectable, and which names watermarks, metadata, cryptographic provenance and fingerprints as acceptable methods ^[10]. The same article exempts systems performing standard editing or changes that do not substantially alter the user's input or its meaning ^[11]. Anthropic elected not to run separate output behaviour for Europe, which is how a European transparency rule becomes the default for Claude customers elsewhere ^[12].

The interesting part is the honesty in Anthropic's own caveats. Heavy editing, paraphrasing, translation or combining a passage with other writing can make a mark undetectable, and short passages may not carry enough information for a reliable signal ^[6]. File provenance can vanish when someone converts or re-saves a file, or takes a screenshot ^[7]. Anthropic also says the planned detector will indicate only that material may have been processed by Claude, and explicitly warns that a positive result will not establish the original author or the content's full history ^[8]. A document can therefore carry a Claude mark because the model proofread, translated, summarised or reformatted human writing, which is an attribution problem for publishers and for anyone running Claude inside a production pipeline ^[9].

Against that, the removal tool is thin. Brian Roemmele, founder of ReadMultiplex, published a Mac-only release called AI Watermarks Cleaner, introduced on X and paired with LM Studio for local model execution ^{[13] [14]}. The workflow is local rewriting rather than stripping metadata or invisible Unicode: separate visible formatting signals from statistical patterns, then have a local model rewrite the passage ^[15]. Instructions and download sit behind ReadMultiplex's paid membership, with Windows and Linux versions promised ^[16]. Roemmele describes the marks as fingerprints capable of tracking output back to a user or prompt, which goes beyond what Anthropic documents ^[17]. And because there is no public spec, no public detector and no reproducible before-and-after results, nobody can currently verify that the cleaner removes anything ^{[5] [18] [19]}.

Watch for the technical specification and detector, since both the compliance value and the removal claims are unmeasurable without them ^[5]. Watch whether the mark attaches to Claude output that only edited human text, which is exactly the case Article 50 exempts ^{[9] [11]}. And watch how AWS, Google Cloud and Microsoft Foundry expose or suppress this for their own customers ^[2].