AI News

Anthropic Rolls Out Invisible Watermarks on Claude Output Worldwide — Text Marks and Signed File Provenance, Starting With EU Models

By The AI Post·2026-08-12·📖 4 min read

Anthropic has begun marking AI-generated content across Claude — invisible watermarks woven into generated text plus digitally signed provenance metadata attached to generated files — worldwide, with Claude models launched in the EU on or after August 2, 2026 supporting the marks from day one. The rollout is tied to the company's EU AI Act transparency obligations.

What happened

According to Anthropic's official help-center documentation, marking applies to output from supported Claude models across the Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, wherever Claude is offered — and embedded watermarks also apply when supported models are accessed through AWS, Google Cloud or Microsoft Foundry. Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, which is why EU-launched models carry the marks from launch.

Why it matters

Platforms and publishers are scrambling for ways to identify machine-written content as AI output floods the web. Anthropic's approach is notable because the watermark is applied at the model level: it is part of the generated text itself, survives copy-and-paste, and may persist through light editing — so Claude output remains verifiable even after it leaves the app.

How Claude marks content

Embedded watermarks in text: an imperceptible mark woven into generated text that does not change meaning, quality or readability. It travels with the text when copied elsewhere and may persist through some editing.
Signed provenance metadata on files: when Claude generates a supported file type (.svg, .png, .jpg), it attaches metadata following the C2PA open standard, signaling that the file was processed by Claude and letting you detect tampering.
Detection tooling: Anthropic says it is working to enable users and third parties to detect both the watermarks and the provenance metadata.

Limitations Anthropic acknowledges

A detected mark is a signal, not proof of authorship — people often use Claude to proofread, translate or summarize content, so the underlying ideas may originate elsewhere.
Marked content can be modified after Claude processes it.
The absence of a mark does not mean content was not AI-generated: older models, heavy editing or paraphrasing, very short passages, or stripped file metadata can all remove or prevent marks.
Not every platform or feature supports every marking type, and signed provenance may not be supported on every cloud partner platform.

What we don't know yet

Anthropic has not detailed the detection tools' accuracy, how marks hold up against deliberate stripping, or when pre-August 2 models get retroactive marking (it says it is working on that). For context, Google already offers SynthID watermarking for AI images and audio and extended text watermarking to Gemini API customers in 2025; OpenAI has explored but not broadly shipped text watermarking.

Verdict

Watermarking itself is not new — the commitment is. Anthropic is making provenance a default property of Claude output in every product and region, enforced at the model level rather than by an app-layer check. Expect the industry to follow, and expect the cat-and-mouse over stripping marks to follow too.

Sources

Primary source

Anthropic / Claude Help Center — How Claude marks AI-generated content

Additional reporting

Get it built for your business

Using Claude or another AI model in your workflow? We help businesses pick the right AI stack — and document how AI-generated content flows through it. Message us on WhatsApp.

Chat on WhatsApp

The AI Post

Independent AI & technology publication — research and verification driven. Read the editorial policy.