🔗 How AI text watermarking works: a visual guide

A watermark in plain text may seem impossible, as text lacks pixels and metadata. However, Google has implemented text watermarking from Gemini, starting in 2024, with new Claude models marking text invisibly and resiliently as of August 2026. These marks reside in the choices between words, not within the characters themselves.

The process involves small word choices made by a model that rolls weighted dice among acceptable options. Each choice can secretly lean toward certain words using a key, which makes counting colored choices easier for detection. The text can still read normally, but marked sections show an imbalance when re-evaluated with the correct key. Even editing alters the marks, as they depend on unbroken runs of original text.

Detection remains private and needs the provider’s key, contrasting with unreliable AI style detectors. This method can indicate whether a text has been processed without confirming authorship. Overall, it highlights the ongoing battle in maintaining and detecting AI-generated information.