SynthID

What is SynthID?

SynthID is a tool designed to watermark and identify AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text, or video. This technology helps users determine if content, or part of it, was created using Google's AI tools. SynthID's watermarking technique is imperceptible to humans but detectable for identification, addressing the challenge of watermarking AI-generated text without compromising quality and accuracy. By using deep learning models and algorithms, SynthID can scan various types of content for digital watermarks, ensuring authenticity and aiding in the identification of AI-generated material. This innovative toolkit is still in beta but shows promise in enhancing trust in information and addressing AI safety concerns.

Who created SynthID?

SynthID was created by DeepMind in collaboration with Hugging Face. The technology is integrated into the Gemini platform for developers to use in their models. The first deployment of SynthID was through the AI music generation model Lyria, where all AI-generated audio is embedded with a SynthID watermark. The watermarking technique is imperceptible to humans but detectable for identification, aiming to promote trust in information and address AI safety issues.

How to use SynthID?

To use SynthID for AI-generated text, follow these steps:

1. Token Generation: An LLM generates text one token at a time, with each token representing a character, word, or part of a phrase. The model predicts the next most likely token based on preceding words and assigned probability scores to generate a coherent sequence of text.

2. Token Adjustment: When there are multiple potential tokens to choose from for completing a phrase, SynthID adjusts the probability scores of predicted tokens to maintain quality, accuracy, and creativity in the output.

For AI-generated audio content:

1. Spectrogram Generation: Convert the audio into a spectrogram, a two-dimensional visualization showing how the frequency spectrum evolves over time.

2. Watermark Addition: The digital watermark is added to the spectrogram, ensuring it remains inaudible to human ears to preserve the listening experience.

3. Robustness: The watermark is designed to withstand common modifications like noise additions, MP3 compression, or speed changes in the audio track. SynthID can also scan the track to detect watermarks at various points to identify AI-generated content accurately.

For AI-generated images and videos:

1. Watermark Insertion: SynthID adds imperceptible digital watermarks directly into the pixels of AI-generated images or each frame of AI-generated videos.

By following these steps, users can effectively utilize SynthID to watermark and identify AI-generated content across different mediums such as text, audio, images, and videos.