Key facts
- Anthropic is implementing invisible watermarks in Claude AI's output.
- The watermarking method is adapted from Google DeepMind technology.
- This implementation is to comply with the EU AI Act.
- Anthropic acknowledges limitations with the watermarking system.
- Watermarks can potentially be removed through editing.
- Human text edited by Claude may be misidentified as AI-generated.
Anthropic is implementing an invisible watermarking system for its Claude AI model's output, a move designed to comply with the European Union's AI Act. The technology used for this watermarking is an adaptation of methods developed by Google DeepMind. The primary goal of this initiative is to enable the identification of content generated by AI systems.
Despite the implementation of this system, Anthropic has acknowledged certain limitations. The company states that the watermarks can potentially be removed through editing processes. Furthermore, there is a possibility that human-written text that has been edited by Claude could be misidentified as AI-generated content due to the watermark.
The EU AI Act mandates that AI systems must be transparent about their origins, and watermarking is one method to achieve this. The act aims to ensure that users are aware when they are interacting with AI or consuming AI-generated content. Anthropic's adoption of this technology reflects the growing regulatory pressure on AI developers to build safety and transparency features into their models.