Will Anthropic’s new AI text watermarking method make Claude’s responses worse?

Anthropic has announced plans to watermark text generated by its Claude AI models, in response to increasing concerns that a model-wide implementation of watermarks could impact the quality of AI-generated responses. Any text longer than 200 tokens, which is approximately 150 words, will bear a machine-readable watermark that is invisible to human users and cannot be removed.

The watermark will also be present on AI-generated image files, attached through cryptographically signed provenance data. Anthropic claims its watermarking method meets several criteria, including being efficient, inexpensive, invisible to readers, and difficult to remove. The technique is based on Google DeepMind's SynthID-Text research paper and does not involve hiding non-printing Unicode characters in the text.

It is expected to comply with transparency obligations under the EU's AI Act, as well as the Code of Practice on Transparency of AI-Generated Content voluntarily signed by several major AI providers.