Anthropic to add invisible watermarks to text and images generated with Claude

Anthropic will add machine-readable markers to text and images generated with Claude as part of the EU's artificial intelligence transparency rules.

12punto 11 August 2026

Anthropic is preparing to implement a new marking system to ensure that text and images generated by its artificial intelligence model, Claude, can be identified. According to the company's statement, Claude outputs will contain data that is imperceptible to the human eye but readable by machines and supported platforms.

The move is stated to be in compliance with the European Union's artificial intelligence transparency rules. According to information on Anthropic's support page, embedded watermarks will be added to text created by Claude, while files will carry digitally signed data on appropriate platforms.

Anthropic is switching to an invisible marking system to facilitate the detection of Claude content.

REGULATION TO BE APPLIED GLOBALLY

The new transparency obligations that entered into force on August 2 under the EU AI Act grant a four-month compliance period for existing AI products already on the market. Anthropic announced that newly released Claude models will mark content from day one, and that integration efforts are ongoing for existing models.

It was stated that the machine-readable markers will be applied globally across supported Claude models, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. For visual content, the C2PA standard, also used by companies such as Adobe, OpenAI, and Google, will be adopted.

On the text side, Anthropic will use a different watermarking method with limited details provided. The company states that this method will not degrade the meaning, quality, or readability of the text, and that the watermark is designed as an integral part of the content.

The goal is for text watermarks to be preserved even when content is copied and moved elsewhere.

According to Anthropic, this "invisible watermark" will be able to travel with the text even if the content is copied and pasted, and it will be able to maintain its presence after certain editing processes. Since the system will be implemented at the model level, it will be activated automatically regardless of which product or interface is used to access Claude.

The company also emphasized that these watermarks will be valid when Claude models are accessed via AWS, Google Cloud, or Microsoft Foundry. This is intended to make content originating from Claude easier to detect by both users and digital platforms.