Text Watermarking for Non-Academics

Anthropic has unveiled how its Claude AI model incorporates text watermarking, a technique that brings statistical watermarking into the realm of practical applications. Unlike watermarks in images or files, which are easier to detect, a textual watermark embedded within the content of written text is less apparent. This is because the copied text tends to retain its provenance, making it challenging to discern the origin.

The article breaks down the mechanism behind statistical text watermarking, explaining that it relies on the inherent redundancy of natural language. As writers make choices among words, contractions, sentence structures, and grammatical constructions, these decisions create a unique statistical pattern that can be detected later. Even though individual choices seem inconsequential, when compiled over a substantial passage, they form a discernible pattern that can be identified by a detector.

This watermarking method is grounded in the principles of stylometry, a field that studies writing styles and uses them to estimate authorship. By analyzing features such as the frequency of function words, spelling preferences, sentence length distributions, and recurring grammatical constructions, analysts can compare texts and assess the likelihood of a shared author.

However, the effectiveness of this technique heavily depends on the volume of text analyzed; a small sample may not provide sufficient evidence, while a larger sample allows for clearer patterns to emerge.

It is important to note that while Anthropic's announcement brings this concept to immediate attention, the underlying technique of text watermarking is not exclusive to Claude or any specific AI model. Various vendors can employ different token-selection rules and detection methods, and the public descriptions of these technologies do not provide a complete implementation specification.

The focus of this article is to elucidate the general principles of text watermarking, rather than attributing a specific design to any particular AI product.