New Delhi: Anthropic has introduced machine-readable watermarks to the content created or manipulated by its Claude AI models. The marks won’t show up
to users but will stay in the text after it’s been copied and pasted, the company claims.
This is a step Anthropic is taking to comply with Article 50 of the European Union’s AI Act, which mandates transparency.
The system will enable users and third parties to detect content created or manipulated by Claude, the company states.Invisible watermark for AI-generated text
Anthropic says that it will include an invisible watermark in generated text with its supported Claude models. The mark will not impact the meaning/content of the response and may be retained in part of the editing.
Claude will include signed (content provenance and authenticity) data for files like SVG, PNG and JPG based on the Coalition for Content Provenance and Authenticity (C2PA) standard. Claude’s metadata may reveal if a file was processed by Claude and if it has been modified.
Watermarks do not prove AI authorship
If a watermark is detected, it means that Claude might have processed the content, but it doesn’t necessarily mean that it was. Detecting a watermark will only tell you that Claude may have processed the content. It will NOT establish that Claude composed the total work.
I keep thinking Anthropic can’t possibly top their last bad decision but here we are https://t.co/xEuY21uzUt
— Alex (@The_Alex) August 10, 2026
For example, users may ask Claude to proofread, translate, summarise or modify existing content. Similarly, heavily edited or combined text may lose the watermark.
The company also states that a lack of watermarking is not a guarantee that the content was not created using AI.
New Claude models will carry the mark
The marking system will be used on Claude across all supported services, including worldwide on cloud platforms like AWS, Google Cloud and Microsoft Foundry, as well as on Claude, Claude Code, Claude Cowork and other supported services, Anthropic says.
The new Claude models will have content marking available at launch. Models prior to August 2, 2026, are still in the update process.
Google DeepMind’s SynthID is one such method that uses invisible signals to detect AI-generated content, and it’s very much the same approach Anthropic has taken.
