Claude will secretly watermark AI-generated text worldwide

Anthropic is including invisible watermarks to textual content generated by the latest Claude models, giving AI-written content material a hidden sign that may observe it even after it’s copied out of the chatbot.

The change comes as Anthropic indicators the EU AI Act’s Article 50(2) Code of Follow on Transparency of AI-Generated Content material. In an replace printed on its privacy and legal support page, Anthropic says that new Claude fashions launched from August 2, 2026, will assist machine-readable marking from day one. Regardless of the EU regulation prompting the transfer, the markings will apply worldwide wherever Claude is obtainable.

The watermark is embedded instantly into generated textual content and is designed to be imperceptible. Anthropic says it received’t change the response’s which means, high quality, or readability. As a result of the marking is utilized on the mannequin stage, it’s going to work throughout Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag, in addition to supported cloud platforms.

Importantly, the mark isn’t tied to the place you employ Claude. Supported fashions accessed by means of AWS, Google Cloud, or Microsoft Foundry may even embed the watermark, the place supported.

The watermark will journey with the textual content when it’s copied and pasted elsewhere and should survive some enhancing. Nevertheless, Anthropic isn’t positioning the know-how as a definitive check of AI authorship. A detected mark signifies that content material could have been processed by Claude, fairly than proving that Claude initially wrote it. For instance, somebody may use AI to proofread, translate, summarize, or reformat their very own work.

The reverse additionally applies. A lacking mark doesn’t show that the textual content wasn’t generated or processed by AI. Heavy enhancing, paraphrasing, translation, brief passages, or combining Claude’s output with different writing could make the watermark undetectable.

Anthropic hasn’t but detailed how customers or third events will detect the marks, saying it’s going to publish technical steerage and detection mechanisms later.

The corporate can also be including signed provenance metadata to supported file sorts, together with PNG, JPG, and SVG. The metadata makes use of the Coalition for Content material Provenance and Authenticity (C2PA) commonplace and might point out {that a} file was processed by Claude.

Anthropic isn’t the primary main AI firm to signal the code. Google announced in July that it will signal the EU’s transparency code, whereas Meta additionally confirmed its participation. The code offers a framework for complying with Article 50’s necessities relating to the marking and detection of AI-generated content material, which took impact on August 2.

Present Claude fashions launched earlier than August 2 received’t instantly assist the brand new marking system. However the firm is working so as to add the function to these fashions throughout the transition interval supplied by the EU AI Act.