What is Deep Mind Synthetic ID? How It Works?
SynthID is a tool that embeds an invisible watermark into AI-generated text and videos. This watermark helps identify whether content was created by AI, without affecting the quality or speed of the content. Today, we’re expanding SynthID to watermark text in the Gemini app and videos in Veo, our advanced generative video model.


What is SynthID?
SynthID is a tool that embeds an invisible watermark into AI-generated text and videos. This watermark helps identify whether content was created by AI, without affecting the quality or speed of the content. DeepMind expanding SynthID to watermark text in the Gemini app and videos in Veo.
How Does Text Watermarking Work?
When a large language model (LLM) generates text, it predicts and selects the most likely next word or phrase, known as tokens. SynthID subtly adjusts the probability of these tokens to create a unique pattern — a watermark — in the text.
Example:
Imagine asking an AI to explain quantum mechanics to a child. The AI selects words based on probability scores. SynthID tweaks these scores slightly to embed a hidden watermark in the text.
Benefits:
Effective in Long Texts: Works best for longer responses like essays or scripts.
Resilient to Minor Changes: The watermark remains detectable even if the text is slightly modified.
Limitations:
Less Effective for Factual Prompts: Short, factual answers like “What is the capital of France?” offer fewer chances to adjust tokens.
Struggles with Major Rewrites: If the text is heavily rewritten or translated, detecting the watermark becomes difficult.
How Does Video Watermarking Work?
Videos are made up of individual frames. SynthID applies an invisible watermark to every frame, making it undetectable to the human eye but identifiable by detection tools. This is particularly important as video generation technology evolves, helping to prevent the spread of AI-generated misinformation.
Example:
If you use Veo model to create a video on VideoFX, SynthID will watermark every frame. This ensures viewers can verify if a video was AI-generated.
Future of SynthID:
SynthID across various AI platforms and content types to reduce the misuse of AI-generated content. Google Also Plan to publish a detailed research paper and open-source the text watermarking technology, allowing developers to use and integrate it into their models.