Key facts
- Anthropic is implementing AI watermarking for its Claude models to comply with the EU AI Act.
- The watermarking embeds invisible signals into AI-generated text to identify its origin.
- Some users have expressed dissatisfaction with the new watermarking feature.
- Anthropic plans to offer an API for checking text for Claude's watermark.
- The watermarking technique is based on a 2024 Google DeepMind paper.
Anthropic has detailed its new AI watermarking feature for its Claude chatbot, aiming to comply with the EU AI Act's Transparency Code. The technology embeds imperceptible signals into AI-generated text, allowing for its origin to be identified. This move comes amid user concerns about potential exposure of AI use in professional and academic settings.
The company explained that the watermarking process involves making subtle word choices that create a detectable pattern, without affecting the quality or readability of the output. Anthropic plans to utilize the SynthID-Text approach developed by Google DeepMind and will offer an API for detecting these watermarks. They differentiate this method from AI detection tools that analyze writing patterns.
Anthropic clarified that while light editing might not fully remove the watermark, a complete rewrite would. The impact on watermarking for code generation is expected to be negligible due to the need for functional accuracy. The company also noted that other major AI developers are adopting similar watermarking practices to adhere to the EU's regulations.
