How AI text watermarking works

AI text watermarking works by subtly influencing which words a language model chooses when multiple options would be equally valid, creating an invisible pattern that only reveals itself to someone with the secret key. The watermark doesn't exist in the characters themselves but in the statistical choices between synonymous or interchangeable words—the key tilts the odds slightly toward certain "green" words while allowing "red" words to occasionally appear, keeping the text natural-sounding. Google and Anthropic have implemented this technique in their AI models since 2024-2026, making watermarks that survive copying and pasting while remaining undetectable to casual readers.