Anthropic, the developer of one of the leading language models Claude, has officially revealed details about the implementation of an invisible text watermarking system. This technology is designed to help users and regulators distinguish content generated by artificial intelligence from text written by humans. The decision is a response to tightening regulations from the European Union regarding AI, as well as a global demand for digital content transparency.
SynthID Technology: A Hidden Mathematical Pattern
The new system is based on the SynthID-Text algorithm, originally developed by the Google DeepMind team. The watermarking mechanism does not require adding special characters, hidden tokens, or altering the visual formatting of the text. Instead, the Claude model uses a special random number generation key to choose between synonymous word options. For example, when describing the weather, the model might choose the word "cloudy" or "gloomy" not only based on context but also in accordance with a mathematical pattern that forms a unique signature.
Data Security and Privacy
Anthropic developers emphasize that the implemented technology does not affect the speed of response generation and does not increase the cost of using the service. An important aspect is privacy: the hidden pattern does not contain user personal data, information about organizations, or the history of specific dialogues. This allows content to be marked without violating user privacy or creating vulnerabilities for information leaks.
Limitations and Scope of Application
The effectiveness of the watermark depends on the nature of the text. In situations requiring strict accuracy, such as factual data or programming code, the algorithm is hardly applied, as the model has no freedom to choose synonyms. Exceptions may only be arbitrary variable names or comments in the code. Furthermore, minor editing of the text by a human will not completely remove the watermark, however, a complete rewrite with the replacement of every word destroys the marking.
Verification Tools and C2PA Standards
To verify the presence of a watermark, the company plans to launch a special API. This tool will determine the probability of Claude's involvement in creating the text, but will not be able to identify the specific author or confirm that the text was written entirely by a human. For multimedia content, including images in .png, .jpg, and .svg formats, cryptographic metadata of the C2PA standard will be used, ensuring file integrity without changing their content.
