ANTHROPIC TO ADD INVISIBLE WATERMARKS TO CLAUDE-GENERATED TEXT AND IMAGES

Anthropic announced this week that its Claude AI models will embed invisible watermarks into text and images to identify content generated by the system. The watermarks will be implemented across all Claude models launched on or after 2 August, including those accessed through the API, Claude Code and other variants. The change applies globally, not only within the European Union, though it responds to EU regulations requiring AI companies to mark artificially generated content.

The invisible markers will be embedded into letters, words and image data in ways that remain detectable by automated systems even if users copy and paste text across platforms or edit the content manually. Anthropic stated that the watermarks could include variations in word spacing, word sequencing and other signals imperceptible to human readers. The company acknowledged limitations: watermarks in images may be removed through conversion or screenshot, whilst text watermarks might persist even when humans contribute substantially to edited content. Anthropic said the marks should "provide a signal that content was processed by Claude" rather than constitute definitive proof of AI generation.

The watermarking initiative addresses concerns about low-quality AI-generated content circulating online. The EU's Code of Practice on Transparency of AI-generated Content requires companies providing AI systems to inform users when they interact with AI tools and to include watermarks on generated material. Some users have raised concerns that the text watermark feature could incorrectly flag human-edited content as AI-generated. Anthropic stated it would share technical details on detecting the watermarks in future communications.