AI watermark
Updated June 10, 2026
A hidden, machine-checkable signature embedded in AI output at generation time — real technology, barely deployed for text.
Definition
An AI watermark marks content as machine-generated at the moment of creation. For text, the standard scheme biases the model's token choices using a secret key — slightly preferring words from a pseudorandom "green list" — so a checker holding the key can find statistically impossible green density. Readers notice nothing; the signature lives in the probabilities.
Deployment status
Google deployed SynthID-Text for Gemini output; OpenAI built a text watermarking system and has publicly held off shipping it. No universal cross-vendor scheme exists, so the detectors you actually encounter work statistically instead. Full story: does ChatGPT watermark its text?
Why text watermarks are fragile
Rewording washes the signal out — the green-word bias dies under paraphrase, a fragility vendors acknowledge. Watermarks also only cover the issuing model's output, leaving open-source models unmarked. Images and audio hold watermarks far better, which is where the technology has actually shipped widest.