Visual guide: How AI text watermarking works by hiding in choices

ArtificialOther · x · 2026-08-17

This article provides a visual guide to the inner workings of AI text watermarking.

Core Mechanism:

Not hidden in characters: Text lacks pixels for data storage. Instead, watermarks exist within the decision-making process of word selection.
How it works: When generating text, a model chooses from a list of probable words (weighted dice). A secret key subtly biases the weights towards specific options, embedding a statistical pattern without altering the text's meaning.
Industry Adoption: Google has watermarked text in the Gemini app since 2024, and Claude models have applied model-level watermarking since August 2026.

More from Safety