During content generation, SynthID subtly modifies the output — e.g. adjusting the token-selection distribution in text, embedding inaudible patterns in the audio signal, or invisible pixel-level changes in an image — encoding the watermark directly into the content (not into easily removed metadata). A separate model/detector analyzes the content and estimates the probability that it carries a SynthID watermark, remaining robust to common transformations.
The growing difficulty of distinguishing AI-generated content from authentic content raises the risk of disinformation; SynthID makes it possible to mark and identify generative content, supporting provenance transparency.
Debut of watermarking for Imagen-generated images.
Expansion to text and audio and the release of SynthID Text as open source.
SynthID watermarks all audio generated by the Gemini 3.8 Live voice models.
