Anthropic has begun rolling out invisible, machine-readable watermarks for text generated by its Claude AI models. The move, announced via an updated support page on August 10, 2026, is part of the company’s effort to comply with transparency requirements under the European Union’s AI Act.

Here’s a clear, fact-based look at what this watermarking system does, how it works, and what its limitations are.

Why Now? The EU AI Act Driver

The watermarking initiative stems directly from Article 50 of the EU AI Act, which requires providers of generative AI systems to make synthetic content identifiable through technical means. Anthropic has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content as both a generative AI model provider and a generative AI systems provider.

The transparency obligations took effect on August 2, 2026. Non-compliance can trigger fines of up to €15 million or 3% of total global annual turnover, whichever is higher. While the law includes a transition period for models launched before August 2, new Claude models will support marking from day one upon release.

How the Text Watermark Works

According to Anthropic’s official documentation, when a supported Claude model generates text, it “weaves an imperceptible watermark directly into the text itself.” Key characteristics include:

• Invisible to human readers – You won’t see it during normal reading.

• No impact on quality – It does not change the meaning, quality, or readability of Claude’s response.

• Survives copying and pasting – Because the watermark is part of the text itself, it travels with the text when copied and pasted elsewhere.

• May persist through some editing – Anthropic notes the watermark “may persist through some editing,” though the company has not specified exactly how much editing is required to remove it.

• Applied at the model level – The watermark is present regardless of which Claude product or interface the text comes from.

For supported file types such as .svg, .png, or .jpg, Anthropic uses a separate system: signed provenance metadata following the Coalition for Content Provenance and Authenticity (C2PA) open standard.

Global Rollout Across All Claude Products

The watermarking applies worldwide, not just to users in the EU. Coverage includes:

• Claude Platform (API)

• Claude (consumer app)

• Claude Code

• Claude Cowork

• Claude Tag

The watermark also applies when supported Claude models are accessed through cloud partners including AWS, Google Cloud, and Microsoft Foundry.

Anthropic is also working to add marking support to Claude models released before August 2, 2026, during the EU AI Act’s transition period.

Detection Tools Are Coming

Anthropic has committed to supporting users and third parties in detecting Claude’s watermarks and provenance metadata, as the EU Code requires. The company says it will share technical details in forthcoming documentation.

What distinguishes Anthropic’s approach is the explicit commitment to third-party detection access – anyone with Anthropic’s detection tools will be able to verify whether text was generated by Claude. This stands in contrast to closed, proprietary detection systems where verification depends entirely on a single company.

Important Limitations: Signal, Not Proof

Anthropic has been careful to frame the watermark as a signal, not conclusive proof. The company’s support documentation outlines several important limitations:

False positives – A detected mark does not mean Claude was the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.

False negatives – The absence of a mark does not confirm that content was not AI-generated. Reasons a mark might be undetectable include:

• Content came from an older Claude model without marking support

• Heavy editing, paraphrasing, or translation

• Very short passages that provide too little material for reliable detection

• Metadata stripped through format conversion, screenshots, or re-saving

Anthropic explicitly states that “a detected watermark is a signal, not proof,” and that “unmarked content does not necessarily mean a human created it.”

What This Means for Users

For most Claude users, this change will be entirely invisible. The watermark does not affect response quality, readability, or the overall experience of using the model.

For content platforms, publishers, and organizations, the system provides a new technical signal for identifying AI-generated material that can survive copying, pasting, and some downstream workflows. However, because the watermark is not foolproof and can be removed through editing or translation, it should be treated as one data point among others rather than a definitive authentication mechanism.

Anthropic plans to publish more detailed technical documentation about the watermarking system and detection tools as they become available.

This article is based on Anthropic’s official support documentation and announcements made on August 10–11, 2026. All claims are drawn from published sources.