Watermarking

Watermarking is a class of techniques where a language model embeds a statistical signal in its output that a downstream detector can check for. The signal is invisible to readers but detectable by code that knows what pattern to look for.

In practice, watermarking has not become widespread. Major model providers haven't committed to it at the production scale, and watermarks are fragile — light editing, translation, or paraphrasing breaks them quickly. Detection that relies on watermarking is best treated as one signal among many, never the primary one.

AI Detector API does not rely on watermarks. We detect based on statistical and learned features of the text itself, which means we work on output from any model, watermarked or not.