Anthropic has clarified its plans to implement invisible watermarks on text generated by Claude to comply with the European Union’s AI transparency mandates. The company announced that its text marking system is a version of the "SynthID-Text" approach, an open-source technology developed by Google DeepMind.
The Mechanics of SynthID-Text
The watermarking system functions by creating detectable patterns based on wording probabilities. This feature, along with C2PA support for images processed by Claude, is designed to meet obligations under the EU AI Act. The law requires synthetic audio, images, video, and text to include machine-readable marks that identify them as artificially generated.
Anthropic explained the process using a practical example: in a sentence like "The weather today was cold and...", the model might choose between likely words such as "overcast" or "grey." Usually, this choice is settled by a random number generator. Under the new system, the source of randomness is replaced by a specific "key" and the preceding words to influence the word choice.
Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses.
This pattern remains undetectable to a human reader but can be identified by anyone possessing the encoding key.
User Impact and Industry Standards
Anthropic emphasized that these invisible marks will not make Claude more expensive for users, nor will they have any practical impact on the quality or content of the model's outputs.
The move aligns Anthropic with other major AI developers subject to the EU AI Act. Google’s Gemini chatbot has utilized the SynthID-Text solution since 2024. While OpenAI has yet to detail its specific text watermarking plans for ChatGPT, it remains legally obligated to follow the same transparency requirements.
