Anthropic details how Claude's invisible text watermarks will function
Anthropic has explained the mechanics behind a new system that will embed invisible watermarks in Claude's generated text to comply with EU AI transparency rules. The approach is based on Google's open-source SynthID-Text technology.
Anthropic has clarified how it plans to apply invisible watermarks to text generated by its Claude chatbot. On Friday, the company said its marking system is "a version" of the SynthID-Text approach, an open-source watermarking technology developed by Google DeepMind that produces detectable patterns based on word-choice probabilities.
The feature, together with C2PA support for images processed by Claude, is being rolled out so Anthropic can meet its obligations under the European Union's AI Act. The law requires that synthetic audio, image, video, and text content carry machine-readable marks allowing it to be identified as artificially generated. Anthropic says the watermarking will not raise costs for users and will have no practical effect on the quality or content of Claude's responses.
How the system works
Anthropic illustrates the concept with an example sentence: "The weather today was cold and..." The next word is highly unlikely to be something unrelated, but several similar options — such as "overcast" or "grey" — would fit almost equally well without changing the sentence's meaning. In such low-stakes situations, a random number typically decides which word is chosen. The watermarking system relies on exactly these kinds of choices, which recur many times throughout a piece of text, to embed a pattern that is invisible to readers but detectable by anyone holding the relevant key. The source of randomness changes under this system — instead of an arbitrary random-number generator, the choice of the next word is determined by the key combined with the preceding words.
Anthropic notes that the EU's transparency requirements will also affect other major AI developers, meaning Claude won't be the only model to introduce text watermarks. Google's Gemini chatbot has used the SynthID Text solution since 2024, while OpenAI has not yet detailed any text-watermarking plans for ChatGPT in its AI Act compliance roadmap, even though it too will be subject to the law.
