Anthropic details Claude text watermarking methods
TL;DR. Anthropic revealed its plan to watermark text generated by Claude by subtly altering inconsequential word choices to comply with the EU AI Act. - The technique modifies model outputs in a detectable way, similar to Google DeepMind's SynthID-Text approach. - Anthropic asserts these changes do not alter the core meaning of the generated sentences. - Other AI model makers are expected to adopt similar text watermarking strategies for content authentication.
- Anthropic plans to watermark Claude's text output.
- Watermarking involves influencing 'inconsequential' word choices by the model.
- The method aims to comply with the EU AI Act and detect AI-generated content.
- Similar techniques are anticipated from other AI model developers.
Sources
- Anthropic says text watermarking scheme relies on inconsequential words — theregister.com
- engadget.com — engadget.com
- techcrunch.com — techcrunch.com
- gizmodo.com — gizmodo.com
- daringfireball.net — daringfireball.net