Home / Technology / AI Text Watermarked: EU Rules Drive Anthropic's Claude
AI Text Watermarked: EU Rules Drive Anthropic's Claude
15 Aug
Summary
- Anthropic uses invisible text patterns for AI watermarking.
- Watermarking ensures compliance with EU AI transparency laws.
- Method applies to text and image generation by Claude.
Anthropic has detailed its strategy for watermarking text generated by its AI, Claude, in adherence to the European Union's latest AI transparency directives. The company's approach embeds a subtle, undetectable pattern within the text, discernible only with a specific key.
This technique influences word selection during text generation, adapting a method pioneered by Google DeepMind. Anthropic assures that this watermarking does not compromise the quality or speed of Claude's output, nor does it incur additional costs.
The watermarking applies to both text and images produced by Claude. For text, it indicates if Claude was involved in its creation, though it cannot distinguish between original generation and editing. Translations and lightly edited texts may also bear the watermark.
While code generation might have less watermarking due to its requirement for exactness, Anthropic aims to watermark all output from models released after August 2. Older models will receive this capability in the coming months, with watermarking applied globally at launch.
