Anthropic explains how Claude's new AI watermarking will work
What's the story
Anthropic has published a blog post addressing basic questions about how it plans to watermark text generated by its Claude chatbot. The company explained how the watermarking will work, whether it can be removed through editing, and what the system means for code generated by Claude. Claude users have been debating the move since Anthropic announced earlier this week that it would introduce watermarking to comply with the EU AI Act's Transparency Code, the regulation to identify AI-generated content.
User reactions
How the watermarking system works
The watermarking system works by creating a pattern in Claude's responses during "low-stakes choices," like picking between "overcast" and "gray."
This pattern is invisible to the reader but detectable to anyone with a key that encodes it.
Anthropic clarified that the watermarking process doesn't compromise the quality of Claude's output, making watermarked and unwatermarked responses indistinguishable to readers.
Technical details
Plans to adopt Google DeepMind's SynthID-Text method in 2024
Anthropic plans to use the SynthID-Text method proposed by Google DeepMind in 2024. The company also intends to launch a watermark detection API.
It's important to note that this watermarking system is different from AI detection methods offered by companies like Pangram, which look for "tells" in writing patterns to reveal AI usage.
User control
Will editing remove the watermark?
When it comes to editing, Anthropic says light editing may not completely remove the watermark, but a complete rewrite where every word is replaced will.
The company also clarified that if text has only been lightly edited, "nearly all the words" will have been written by the human author and "there's very little (if anything) for the watermark to attach to."
Code constraints
What about coding?
In coding, the watermark will be less prominent as the model has to generate working code without much leeway for word selection.
However, Anthropic said it can be used in areas where there's an arbitrary choice of words or terms within the code, such as comments.
When editing human-written text, if Claude only makes minor grammar or punctuation changes, there may not be enough Claude-generated words for reliable watermark detection.
Rollout plan
No extra costs for watermarking process
Anthropic has said that the watermarking process will not incur any extra costs as it doesn't add additional tokens.
The company is also working to implement watermarking on models launched before August 2, 2026, with a rollout expected over the coming months.
This transition period under EU rules shows Anthropic's commitment to transparency and compliance in its AI technology.