Google SynthID
Digital watermarks are immediately embedded into AI-generated text, music, video, and photos by SynthID, which watermarks and identifies AI-generated content.
The watermarking method used by SynthID is observable for identification but invisible to humans.
Promoting information trust requires the ability to recognize AI-generated material.
It is a collection of promising technical answers to this urgent AI safety issue, even though it is not a panacea for issues like disinformation or misattribution.
This toolkit is still being developed and is now available in beta. It is currently being included into an expanding number of products, enabling individuals and institutions to handle AI-generated material in an ethical manner.
How does SynthID works?
SynthID watermarks and recognises AI-generated information using a range of deep learning models and algorithms.
Using watermarks
Without affecting the original material, SynthID incorporates a digital watermark straight into AI-generated output.
Recognition
It helps people identify whether content, or a portion of it, was created using Google’s artificial intelligence capabilities by scanning text, audio, video, and photos for digital watermarks.
SynthID for text produced by AI
One of the biggest challenges facing AI researchers has been coming up with a reliable way to watermark text produced by AI without sacrificing its accuracy, creativity, or quality.
One token at a time, an LLM produces text. One character, word, or phrase can be represented by these tokens. The model forecasts the next most likely token to generate in order to produce a string of cohesive text. The preceding words and the likelihood scores given to each possible token serve as the basis for these forecasts.
Take the statement, “My favorite tropical fruits are __,” for instance. The tokens “mango,” “lychee,” “papaya,” or “durian” may be used by the LLM to begin finishing the sentence; each token is assigned a likelihood score. When a variety of tokens are available, SynthID can modify each projected token’s probability score as long as doing so doesn’t impair the output’s quality, accuracy, or originality.
Because this procedure is repeated throughout the created text, a page may have hundreds of modified probability ratings, while a single sentence may contain ten or more. The watermark is the final pattern of scores for both the model’s word choices and the adjusted probability scores. As few as three sentences can benefit from this method. Additionally, SynthID becomes more accurate and robust as the text gets longer.
SynthID for music produced by AI
Google’s most sophisticated AI music creation model to date, Lyria, was the first to use SynthID. A SynthID watermark is integrated right into the waveform of every AI-generated audio file that its Lyria model publishes.
The audio wave, which is a one-dimensional representation of sound, is transformed into a spectrogram using SynthID. This two-dimensional graphic illustrates how a sound’s frequency spectrum changes over time.
The digital watermark is appended to the spectrogram after it has been calculated. The spectrogram is then transformed back into the waveform. In order to preserve the listening experience, SynthID uses audio attributes during this conversion process to make sure the watermark is inaudible to the human ear.
The watermark can withstand a wide range of typical changes, like adding noise, compressing MP3, and changing the track’s pace. In order to ascertain whether or not Lyria may have created some of the audio, it can additionally scan the music for the watermark at various locations.
SynthID for AI-produced visuals and audio
SynthID incorporates an invisible digital watermark into each frame of an AI-generated film or into the pixels of an AI-generated image.
Where can I find it?
Customers of Vertex AI can use this technique with its text-to-image models, Imagen 3 and Imagen 2, which produce excellent images in a broad range of artistic styles. Additionally, ImageFX watermarks the image outputs using SynthID technology. It also included SynthID into Veo, itsmost powerful video creation model to date, which is accessible to a limited number of VideoFX producers.
In order to identify digital watermarking, it may also scan a single image or the individual frames of a movie. The About this image feature in Chrome or Search allows users to determine whether an image or a portion of an image was created using Google’s AI capabilities.
Note: The model utilized on YouTube, Imagen on Vertex AI, may differ from the model used to generate the artificial images and audio samples on this page.
