Anthropic shares more details about how Claude’s new watermarks will work

Core Impact Points

• Anthropic's watermarking feature uses the SynthID-Text approach developed by Google DeepMind in 2024.
• The feature is designed to be undetectable to readers but can be detected by those with the necessary key.
• Watermarking does not impact the quality of Claude's output, and users can still expect the same level of performance.
• The feature is distinct from AI detection approaches offered by companies like Pangram, which look for 'tells' in the writing to reveal AI usage.
• Anthropic plans to release a watermark detection API to help developers and researchers understand the technology better.
• The company emphasizes that watermarking is not a conspiracy against users but rather a necessary step to comply with regulatory requirements.