ทันเอไอ

Decoding AI Text Watermarking: How it Distinguishes Artificial Intelligence Output

Understanding Anthropic's AI text watermarking system, which works through token selection statistics rather than hiding data.

Amid the widespread adoption of artificial intelligence, distinguishing AI-generated text has become crucial, especially as developers like Anthropic, the creator of Claude, implement machine-readable watermarks to comply with the EU AI Act and enhance transparency.

Technical analysis and open-source data reveal that AI text watermarking operates entirely differently from embedding metadata or hiding special characters like invisible Unicode in files.

The core principle relies on "statistical bias" at the token selection level. While the model generates text, a secret key influences it to choose certain synonymous words instead of purely selecting the most naturally probable ones. This acts like embedding a statistical fingerprint into the content.

However, this system has both strengths and limitations. The watermark persists even through copy-pasting, but detection accuracy depends on text length; very short texts are difficult to detect. Furthermore, the statistical signal can be disrupted and fade if the text undergoes heavy editing or paraphrasing.