Unveiling SynthID A Digital Watermark for AI Content

SynthID is a pioneering technology designed to embed digital watermarks into content created by artificial intelligence. This tool aims to enhance transparency and build trust by providing a reliable way to identify AI-generated media without degrading its quality.

Watermarking AI-Generated Images and Video

SynthID works by adding an invisible digital watermark directly into AI-generated images or video segments at the moment of their creation. This watermark is engineered to be robust and persistent, remaining detectable even after the content undergoes common modifications. These alterations include cropping, applying filters, changing frame rates, or using lossy compression techniques, ensuring the origin of the content can still be traced.

Securing AI-Generated Audio

For audio content, SynthID embeds a watermark into tracks produced by AI models like the music generator Lyria or the podcast feature in Notebook LM. This watermark is completely inaudible to the human ear, so it does not affect the listening experience. Furthermore, it is designed to withstand typical audio modifications such as the addition of noise, MP3 compression, or changes to the track's speed, preserving the integrity of the identification marker.

Identifying AI-Generated Text

SynthID's capabilities have also been extended to text generated by large language models, including those powering the Gemini app. The technology operates by subtly influencing the word selection process during text generation. Large language models assign a probability score to each potential next word or token. For instance, in the sentence "My favourite tropical fruits are mango and…", the word "bananas" would have a much higher probability than "airplanes". SynthID adjusts these probability scores slightly to embed a unique watermark pattern. This process is imperceptible to a human reader and has no impact on the overall quality or coherence of the written output.