Watermarking
Hiding a statistical signal in generated text so that it can later be identified as machine-written.
Also: AI text detection, SynthID
A text watermark nudges the model's word choices in a pattern invisible to readers but detectable by someone with the key. Researchers at the University of Maryland proposed a widely cited scheme in 2023, and Google DeepMind deployed SynthID for text in Gemini and released it openly in 2024. Watermarks can be weakened by paraphrasing or translation, and only work if the model's maker applies them. Detectors that try to spot machine text without a watermark have proved unreliable, and have wrongly flagged writing by non-native English speakers.