Anthropic is introducing hidden watermarking for text created through its Claude AI models. If users make no edits, the system could identify Claude’s involvement. However, the watermark will not add visible letters or symbols. Instead, Claude will follow a specific pattern while selecting words. A detection system can then identify that pattern within the text. According to Anthropic, the system aims to detect Claude’s role in creating content. When several words fit equally well, Claude will alter the source of randomness behind its selection. This process can create a distinct pattern across longer text. Anthropic adapted a version of Google DeepMind’s 2024 SynthID-Text technology for this system. Internal testing found no impact on content, creativity, or readability.
The Watermark Cannot Prove Every Text’s Full Origin
The watermark will indicate the likelihood of Claude’s involvement in a text. However, it cannot prove the complete origin of that content. It also will not reveal the user’s identity, organisation, or conversation details. Furthermore, it cannot identify text created by another AI model. Shorter content may produce less reliable detection results. Factual writing, proofreading, and coding may also weaken the watermark. These formats allow fewer changes in word selection. Anthropic says light editing may leave the watermark intact. However, a complete rewrite can remove the watermark. Claude translations will also retain the watermark. Traditional AI detectors examine writing styles and patterns. In contrast, this system targets a watermark linked to one specific model. Anthropic is also developing a Detection API for users and third parties.
Claude Will Also Add C2PA Credentials to Supported Files
Anthropic plans to use C2PA Content Credentials for supported files created or processed through Claude. These files include PNG, JPG, and SVG formats. Cryptographically signed metadata will indicate Claude’s involvement in a file. However, the metadata will not reveal private user information. C2PA operates as an open standard for content credentials. Camera makers and image-editing software already use the standard.
EU AI Rules Are Driving Global Deployment
Anthropic plans to deploy the watermarking system worldwide. The company currently lacks a reliable way to restrict it by region. The European Union’s AI Act also influenced this decision. Anthropic signed the EU’s AI content transparency code in July 2026. Around 190 other signatories joined the initiative alongside several major AI companies. Meanwhile, Anthropic is working to add watermarks to older models. The company plans those additions during the coming months.
