How Anthropic is invisibility watermarking Claude’s text to meet EU regulations – Innovation Village

In a major step toward AI transparency, Anthropic has unveiled a sophisticated new watermarking system for text generated by its Claude AI models. The move is primarily designed to ensure full compliance with the European Union’s sweeping new AI regulations. However, unlike traditional watermarks that stamp a visible logo on an image, Anthropic’s text watermark is entirely imperceptible. It does not alter the formatting, it does not rely on hidden Unicode characters, and it is completely indistinguishable to the human eye.

Instead, the company has engineered a way to embed a hidden cryptographic pattern directly into the mathematical decision-making process of the AI itself.

The Mechanics: A Cryptographic “Key” to Word Selection

To understand how this works, it helps to know how large language models (LLMs) operate. When Claude writes a sentence, it doesn’t think in whole phrases; it calculates the probability of what the next most logical word (or “token”) should be, generating a list of viable candidates. Usually, the model uses an arbitrary random number generator to pick one of the top choices from this list, ensuring the text flows naturally and creatively.

Anthropic’s watermarking intercepts this final step. Instead of using a random number generator, Claude uses a proprietary cryptographic “key” to dictate which word to select.

Anthropic illustrated this by using the digits of pi as a theoretical key (3.14159…). Following this…