Anthropic to Embed Watermarks in Claude Text and Images
The AI company will use C2PA metadata for images and a proprietary system for text to comply with EU transparency rules.
Anthropic plans to embed invisible watermarks in all text and images generated by its Claude AI models, the company announced in a new support page. The initiative responds to transparency requirements under the European Union's AI Act, which took effect August 2, 2026.
The watermarking will apply globally across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, as well as models accessed through AWS, Google Cloud, and Microsoft Foundry.
Two watermarking approaches
For images and other media files, Anthropic will use C2PA (Coalition for Content Provenance and Authenticity), a provenance metadata standard already adopted by Adobe, OpenAI, and Google. This standard attaches digitally signed information to files indicating their AI origin.
For text, Anthropic is developing what it describes as an "imperceptible watermark" woven directly into Claude's output. According to the company, this watermark doesn't affect the meaning, quality, or readability of generated text. The watermark persists when text is copied and pasted, and may survive some editing.
"Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere," Anthropic stated. The company did not name the specific watermarking technology it's using for text.
Compliance timeline and detection tools
The changes won't take effect immediately. The EU AI Act includes a four-month grace period for AI products that launched before August 2. Anthropic said new Claude models will include watermarking from release, while existing models will receive support as work progresses.
The company plans to enable users and third parties to detect these watermarks and provenance metadata, with technical documentation coming in the future. Several tools already exist for detecting C2PA metadata, including Google's Gemini chatbot, though it's unclear whether these will work with Claude-generated files.
Known limitations
Anthropic acknowledged the systems aren't foolproof. C2PA metadata can be stripped from files, sometimes accidentally when content is uploaded to online platforms. The company cautioned that content lacking detectable marks could still originate from generative AI models.
The text watermarking system's robustness remains unclear. Fanfiction communities have already developed informal methods to flag Claude-generated content on platforms like Archive of Our Own (AO3), but Anthropic's approach could provide more systematic detection across the internet.
Why it matters
This move represents a significant step toward making AI-generated content identifiable at scale, addressing growing concerns about synthetic media flooding online spaces. For businesses and platforms, machine-readable watermarks could enable automated filtering and labeling systems. For consumers who want to avoid AI-generated content, these technical standards could provide the infrastructure for meaningful choice—assuming the watermarks prove durable enough to survive real-world content distribution. The effectiveness of these systems will determine whether AI transparency becomes enforceable or merely aspirational.
The details were first reported by Jess Weatherbed at The Verge.
This is an original analysis by the Omega editorial team. Source reporting: AI Watch.
Want systems like this working for your business?
Book a Call
