“`html
Anthropic has announced details regarding its implementation of invisible watermarks in Claude-generated text, designed to meet regulatory requirements established by the European Union. The company revealed that its watermarking system is based on SynthID-Text, an open-source technology originally created by Google DeepMind. This approach embeds detectable patterns within text by leveraging word choice probabilities, allowing the content to be identified as artificially generated.
The watermarking mechanism works by identifying moments in text generation where multiple word choices are equally valid from a reader’s perspective. Rather than using arbitrary random selection, the system applies a specialized key to determine which word the model selects in these low-stakes situations. This creates an invisible pattern throughout the generated content that can only be detected by those possessing the corresponding decryption key.
The introduction of text watermarks, combined with existing C2PA support for images processed by Claude, fulfills Anthropic’s obligations under the EU’s AI Act. This legislation requires that synthetic audio, video, images, and text include machine-readable markers enabling detection of artificially created or modified content. Anthropic has stated the watermarking will not increase costs for users or impact output quality. Other AI developers are similarly implementing such measures, with Google’s Gemini already supporting SynthID-Text since 2024.
“`
Originally reported by Theverge →
About the Author
