Anthropic Reveals Mechanism Behind Invisible Watermarks in Claude
Photo: ITmedia
Quick answer
Anthropic has embedded statistical watermarks in AI-generated texts from its model Claude using a secret key, ensuring compliance with EU AI Act regulations.
Anthropic, the developer of the AI model Claude, has revealed technical details about its invisible watermarking mechanism for AI-generated text. The technology uses a secret key to create statistical patterns that do not compromise response quality, processing speed, or service cost.
According to Anthropic, the primary purpose of this innovation is to comply with the EU AI Act, a European regulation governing AI usage. However, the watermarking system is applied globally to all users of the model. The company notes that the technology has limitations: it is ineffective for short phrases, code, or other specialized content formats.
Watermarks remain invisible to users and do not affect text perception, but they enable origin identification when necessary. Complete rewrites of the text may eliminate these markers, as documented by the company.
Common questions
- What are watermarks in AI-generated texts?
- They are hidden statistical patterns embedded in AI-generated text to identify its origin. These watermarks confirm AI creation without affecting content quality or readability.
- Why does Anthropic watermark Claude's outputs?
- The primary goal is to meet EU AI Act compliance requirements and enhance transparency in AI usage. The technology is deployed globally but has limitations with certain content types.
- Can watermarks be removed from text?
- Yes, watermarks may disappear if the text is heavily rewritten or modified. They also perform poorly with short phrases or code snippets.
Dzen feed: /feed/dzen.xml · RSS: /feed.xml