Anthropic Details AI Text Watermarking for Claude Models

Image credit: Search Engine Journal

Anthropic on Tuesday disclosed technical specifics of its artificial intelligence text watermarking method for Claude models, a system designed to identify AI-generated content.

The technology, a version of Google DeepMind‘s SynthID-Text developed in 2024, embeds an imperceptible signal into generated text by subtly altering the randomness of word selection during the AI’s output process.

Anthropic stated that the watermark does not rely on Unicode characters, metadata, or hidden characters, nor does it impact the quality or content of the generated text, making it undetectable to human readers without a specific key.

Scott Aaronson, a researcher involved in the project, indicated that while the watermark is robust against minor edits, extensive paraphrasing or a complete rewrite of the text could defeat the detection mechanism.

The company confirmed that future Claude models will produce watermarked text to comply with emerging regulations, such as the European Union’s AI Act, which mandates transparency for AI-generated content.

This initiative aims to address concerns about the authenticity of information and the potential misuse of large language models for generating deceptive content.

Source: Search Engine Journal