Safety
Watermarking (AI)
Embedding hidden, machine-readable signals in AI-generated text, images, or audio to prove content provenance — for example Google's SynthID.
Watermarking is the technique of embedding hidden, machine-readable signals — imperceptible to the human eye or ear — into content produced by artificial intelligence. The goal is to mark provenance: to make it possible to later determine whether a piece of text, an image, audio, or video was AI-generated, even after ordinary edits.
The method depends on the content type. For text, the system subtly adjusts the probabilities of which tokens are chosen during generation, creating a statistical pattern that a detector can recognize. For images from diffusion models, as well as audio and video, the watermark is embedded into the signal itself so it survives cropping, compression, or added noise. Google’s SynthID is one example of this approach across all four media types.
Through 2025-2026, watermarking is a key tool against deepfakes and disinformation and a support for regulatory transparency — Article 50 of the EU AI Act calls for machine-readable marking of AI content. Watermarks are not indestructible, though: sufficiently aggressive editing or paraphrasing can weaken them.