Anthropic to implement invisible watermarking for Claude AI outputs

Anthropic has announced that its AI model, Claude, will implement invisible watermarking for both text and image outputs. This move follows the company’s commitment to transparency standards outlined in the EU AI Act. The text watermarking technique is expected to rely on specific statistical patterns in word choice and sequence that are imperceptible to humans but detectable by specialized systems. For images, Anthropic plans to use metadata embedding to track whether content has been altered.

To support this initiative, Anthropic will provide a watermarking detection API for developers and third-party services. This development aligns with broader industry trends, as platforms like Suno are implementing audio watermarking and Google’s Gemini utilizes DeepMind’s SynthID for image and audio marking.

In a related discussion on AI content, a16z crypto noted that the perceived ‘AI flavor’ in writing often stems from human writing habits—such as over-structuring and using vague, pseudo-profound language—which large language models tend to amplify. They suggest that rather than focusing solely on detection, writers should prioritize specificity and authenticity to distinguish their work from generic AI-generated content.