What it is. An AI watermark is a hidden signal built into text, images, audio or video as it is generated, so that a detector can later tell that it came from a particular AI system. People cannot see or hear it.
An analogy. Banknotes carry a watermark you only notice when you hold them up to the light. An AI watermark is similar, except the “light” is a piece of software that knows what pattern to look for.
How it works. For images, audio and video, the signal is a tiny, carefully placed pattern in the pixels or sound wave. For text it is cleverer, because words cannot hide pixels. When a model chooses each next token, there are often several good options. A text watermark nudges those choices in a pattern linked to a secret key. Any single word looks normal, but over a few hundred words the pattern becomes statistically clear to a detector with the key. Google DeepMind described its text method, SynthID Text, in the journal Nature in October 2024.
Why it was in the news this week. Google DeepMind says all video and audio from Genie 4, its new world model, carries “SynthID, our imperceptible digital watermarking technology”. As AI-made worlds, voices and video get more realistic, watermarks are one way to label them. Our brief has the details.
One thing to remember. A watermark can only confirm content from a system that adds it. No watermark does not prove a person made something; another tool may simply not use one, and heavy editing can weaken the signal. A watermark also only helps if someone can run the detector, so who gets access to detection tools matters as much as the watermark itself. Watermarks help, but they are one clue, not a verdict. See also our voice AI explainer and the watermarking glossary entry.
Sources: Dathathri et al., “Scalable watermarking for identifying large language model outputs”, Nature, October 2024, doi.org; Google DeepMind SynthID Text reference code, github.com; Google DeepMind, “Genie 4: A new generation of world simulation”, 27 Oct 2026, deepmind.google.