Anthropic Adds Invisible Watermarks to Claude-Generated Content

Anthropic will be adding invisible watermarks to text generated by Claude AI. The move is part of the company’s efforts to comply with EU rules that are aimed at enhancing transparency in AI-generated content. Anthropic signed the EU AI Act’s Article 50 (2) Code of Practice on Transparency of AI-Generated Content.

The transparency obligations under Article 50 apply from August 2. However, there is a limited transition period for certain generative AI systems placed on the market before this date. The EU Commission extended the transition period for the marking and detection obligation of such systems to December 2, 2026.

How Does the Claude AI Watermark Work?

Anthropic will apply two techniques to make content generated and processed by Claude identifiable. The first technique involves embedding its AI content watermark on any text generated by Claude. The AI firm said the watermark will remain when users copy and paste the content, and could even survive edits.

“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response,” Anthropic noted in a statement published on its website.

Text originally written by humans then edited or processed by Claude will also carry the Claude AI watermark. The company said watermarks will be applied at model level and will be present on every text irrespective of the Claude product used.

For images and other supported files, Anthropic will apply the signed provenance metadata technique. This means information about content provenance will be attached to files to show they have been generated or processed by Claude.

Can Claude AI-Generated Text be Detected?

Anthropic did not immediately provide a way for users and third-parties to detect AI-generated content. However, the company is working on detection mechanisms that will allow them to check whether a file or a piece of text carries the Claude AI mark. The AI firm said details of the detection mechanisms will be published later.

But some experts have argued that AI detection technology have not been perfected yet. Text watermarking works by slightly influencing which words an AI model selects. This creates a pattern that can be detected. Experts claim it might be challenging for watermarks to survive “intense paraphrasing”. They also cautioned that making detection tools available to the public might enable users to find ways to remove them.

How Are Claude Users Reacting to Claude AI Marks?

Anthropic’s move to watermark Claude-generated content has attracted mixed reactions from users. Some have termed it unethical, others have welcomed it, while others have raised questions about the detection process.

“I gave the instructions, context, decisions, and countless refinements, Claude was the tool. If Claude starts watermarking the code or anything else it generates, what exactly is it claiming credit for?,” One user commented on Reddit.

Some users said watermarks are not new in the industry. On detection tools, one user asked whether “the general public have access to tools that can "read" those watermarks”.

Anthropic’s decision comes as major AI firms move to align with the EU’s new transparency rules. In July, Google committed to adopt the bloc’s AI Code of Practice. The AI firm is rolling out Claude’s invisible watermark globally. The company said versions released before August 2 will initially not have the new marking capability. However, the company will introduce the feature during the transition period set out under the EU AI Act.