TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Anthropic says new Claude models covered by its EU commitments will embed machine-readable watermarks in text and attach provenance metadata to supported files worldwide. Its published guidance describes no user opt-out, while older models and public detection tools remain works in progress.

Anthropic will embed machine-readable watermarks in text generated by new Claude models and attach signed provenance records to supported files, applying the marks worldwide across Claude products. The company’s published guidance describes no user opt-out, a policy that could affect anyone using Claude to draft, edit, translate or format material intended for publication.

The policy covers Claude models launched in the EU on or after August 2, 2026. Anthropic says marks will appear across Claude, Claude Code, Claude Cowork, Claude Tag and the Claude API, as well as supported models accessed through AWS, Google Cloud and Microsoft Foundry. Older models are not all covered yet.

For text, Claude will weave what Anthropic calls an imperceptible pattern into the generated wording. Because the mark exists in the text rather than as a visible label, it can remain when material is copied and pasted and may survive some editing. Anthropic says the method does not change the output’s meaning, quality or readability, though it has not released evidence allowing outsiders to test that claim.

Supported files, including SVG, PNG and JPEG files, can receive digitally signed provenance metadata based on the C2PA open standard. That record can signal that Claude processed the file and whether it was altered afterward. File marking may vary by platform because not every service or file type supports the required metadata.

AI Assistance Gains a Trace

The change matters because a detected mark may identify Claude’s involvement without showing its extent. A document written entirely by a person could still carry the signal after Claude proofreads, translates, summarizes or reformats it. Employers, publishers and schools may need policies that distinguish AI assistance from AI authorship.

The worldwide scope also means users outside Europe will encounter a policy prompted by EU transparency rules. Developers using Claude in customer-facing products may need to review how the marks interact with their own disclosures, records and automated workflows. Anthropic says those developers must determine their own Article 50 obligations.

EU Rules Drive Global Marking

Anthropic linked the move to its signing of the EU AI Act Article 50(2) Code of Practice for transparency in AI-generated content. Under the company’s stated implementation, new models receive marking support at launch, while previously released models are being handled on a later schedule allowed by the rules.

The system uses two separate signals: embedded text watermarks and signed file metadata. The former is intended to travel with copied language; the latter stays attached to a file unless software, conversion or resaving strips it. Neither method is a visible disclosure to a human reader.

“New models will mark AI-generated content from day one.”

— Anthropic, in its Claude marking guidance

Detection Details Remain Pending

Anthropic has not yet published the technical design or detection mechanism for its text watermark. It is also unclear when users and independent third parties will receive detection access, what error rates the system produces, or whether it performs consistently across languages and writing styles. The company says technical documentation is forthcoming.

A detected mark is not conclusive proof that Claude wrote the underlying material. Anthropic says a signal means content may have been processed by Claude. Heavy editing, paraphrasing, translation or mixing with other text can weaken detection, while short passages may provide too little material for a reliable result. Missing file metadata also does not prove that Claude was uninvolved.

Older Models and Detectors Await

Anthropic is working to add marking to models released before August 2 and to provide tools that let users and third parties check supported content. The next milestones will be publication of the detection documentation, a clearer schedule for older models and details on platform-specific file support.

Users and organizations should watch for Anthropic’s guidance on which models are actively marking outputs and how detection results should be interpreted. Until those details arrive, a watermark should be treated as a limited provenance signal, not definitive proof of authorship.

Source: Anthropic

Key Questions

Can Claude users disable the new watermarks?

Anthropic’s published policy describes model-level marking across supported Claude services and provides no opt-out setting. The marks are designed to remain regardless of which covered Claude product generates the text.

Does every Claude response already contain a watermark?

No. The confirmed launch requirement applies to models introduced in the EU on or after August 2, 2026. Anthropic says support for older Claude models is still being added.

Can a watermark prove Claude wrote a document?

No. Anthropic says detection only signals that content may have been processed by Claude. Proofreading, translation or formatting could produce a marked output even when a person supplied the original material.

Can editing remove a Claude watermark?

Anthropic says a text mark may survive some editing and copy-and-paste operations. Heavy rewriting, translation or combining the passage with other material can make the mark harder or impossible to detect.

Source: Anthropic