Anthropic Adds Invisible Watermarks to Claude Text and Images Under EU AI Transparency Rules

Anthropic has begun marking content produced by new Claude models with invisible, machine-readable signals, according to The Verge AI, TechCrunch AI and Interestingengineering. The company says generated text carries embedded watermarks, while supported files receive digitally signed provenance metadata. The changes apply to models launched on or after August 2, 2026, and Anthropic plans to apply the approach worldwide rather than limiting it to European users.

The company described the system in an updated Claude support page, the publications report. The markings are designed for software and online platforms to detect, not for people to see while reading or viewing content. The Verge AI identifies images among the covered outputs, while TechCrunch AI reports that Anthropic also intends to extend watermarking support to older models.

The policy is tied to Article 50 of the EU AI Act and the European Union’s Code of Practice for AI-generated content, according to Interestingengineering, TechCrunch AI and Seangoedecke. The rules require providers to make synthetic or manipulated content detectable by systems, creating a regulatory push for provenance signals across text, images and other media.

Anthropic’s global rollout gives the EU requirement reach beyond the region, although the sources differ in how they describe the exact legal timing. TechCrunch AI says the transparency code took effect on August 2, while Seangoedecke describes enforcement as beginning in August 2026. The shared point is clear: regulatory compliance is driving a model-level change in how Claude outputs are produced and identified.

The watermark is a provenance signal embedded during generation, while file metadata provides a separate record for supported documents and images, according to The Verge AI, Explainx and Phrasly. Explainx describes the text marking as imperceptible and distinguishes it from C2PA-style metadata attached to files. Phrasly’s explainer likewise characterizes AI text watermarks as hidden markers that specialized tools can detect without changing the visible wording.

That distinction matters for developers. A detector may inspect Claude text directly, while a platform handling an image or document may read its signed provenance metadata. The available reports don't establish that every file format or every downstream application will preserve the signal. They also don't establish that the markings prove who edited content after generation, so provenance remains tied to the supported signal and the systems able to read it.

Claude’s watermarking won't make AI authorship permanently visible in every version of a text. Seangoedecke argues that text watermarks are trivial to remove through rewriting or other transformations, while Explainx examines what survives editing and what does not. The existence of online pages advertising Claude and AI watermark removal, including Chatgptwatermarkremover and Gpt-watermark-remover, shows that attempts to strip hidden markers are already part of the surrounding market, though those sites don't verify Anthropic’s implementation.

The system therefore works best as a provenance aid for original outputs moving through cooperative platforms, rather than as an infallible authorship test. Copying text can preserve some embedded signal, but rewriting, translation, formatting changes or extraction into unsupported formats can affect detection. Claims from Msukhareva.substack that the fingerprint always remains go beyond the stronger evidence in the technical and news reports.

Developers using Claude APIs will need to account for marked outputs in workflows that publish, archive or inspect generated content. The Verge AI, TechCrunch AI and Explainx indicate that the marking occurs at the model level, which places the change inside generation rather than in a separate user-controlled export step. Platforms can use that signal for disclosure labels, moderation queues, provenance records or research into synthetic content.

Code and text copied into editors remain practical edge cases. Explainx discusses API output and code, while Chatgptwatermarkremover claims hidden characters can cause formatting problems, although its account concerns watermark-removal tools rather than independently verified Anthropic behavior. Anthropic’s support documentation will matter more than third-party cleanup sites as organizations decide how to preserve, inspect or disclose Claude output.

Anthropic’s rollout places provenance at the center of the next phase of AI compliance. The Verge AI and Interestingengineering report coverage for text and images, while TechCrunch AI says Anthropic will support older models, making compatibility and detection coverage the immediate issues for customers and platforms.

The policy also sets a precedent for how providers respond to the EU AI Act: a regional transparency obligation can produce a worldwide product change. Detection reliability, resistance to editing and the range of supported file types will determine whether watermarks become a useful platform standard or remain one signal among several. Anthropic’s support page and future technical details will provide the clearest evidence as deployment expands.