Anthropic has announced that future Claude models will generate watermarked text. This feature is designed to determine the probability that Claude contributed to writing a given text. This implementation aligns with similar efforts by other major AI providers to comply with the EU AI Act.
Claude's watermarking method operates by influencing low-stakes word choices during text generation. Large language models like Claude build text one word at a time, selecting the most probable next word. When multiple sensible options exist, the model typically uses a random number to decide. Watermarking replaces this arbitrary random number with a key-derived random source, leaving a detectable pattern in the output.
Key Points
- Future Claude models will generate watermarked text to indicate the likelihood of Claude's involvement in text creation.
- This change is being implemented by Anthropic and other major AI providers to comply with the EU AI Act.
- The watermarking method uses low-stakes word choices, which occur frequently in generated text, to embed a pattern.
- The pattern is undetectable to human readers but can be identified by those possessing the encoding key.
- Watermarking does not impact the quality, creativity, or readability of Claude's text, according to internal testing.
- The technique used is a version of the SynthID-Text approach, published by Google DeepMind in a 2024 Nature paper.
- Watermark detection is less effective on small text samples and factual passages where word choices are constrained.
Context
According to Anthropic, the watermarking process does not force Claude to select words it would not naturally consider. The model's choices remain random, but the source of that randomness is altered to create a verifiable sequence. This allows for a probabilistic assessment of whether Claude generated the text after the fact.
Why It Matters
This initiative indicates a shift in how AI-generated content will be auditable, particularly for deployers and users who must comply with emerging regulations like the EU AI Act. The ability to determine the likelihood of AI involvement in text generation could influence content verification processes and accountability frameworks.
What To Do
- Note that this watermarking applies to future Claude models.
- Watch for further guidance from Anthropic regarding the availability and specifics of the watermarking detection key.
- Consider how the probabilistic nature of watermark detection might integrate into content verification workflows.
- Review the SynthID-Text paper from Google DeepMind for technical details on the underlying method.