Business
Announcing our new watermarking method for AI-generated text and videos, and how we're integrating SynthID into key Google products
Generative AI tools – and the big language model technologies that underpin them – have captured the public imagination. Whether helping with work tasks or enhancing creativity, these tools are quickly becoming part of the products used by millions of people in their daily lives.
These technologies can be extremely beneficial, but as they become more popular, the risk of people causing accidental or intentional harm, such as spreading misinformation and phishing, if AI-generated content is not correctly identified. That is why last year we launched SynthIDour new digital toolkit for watermarking AI-generated content.
Today we are expanding SynthID Capabilities to the AI-generated text watermark in the Gemini app and web experienceand video in Veoour highest performing generative video model.
SynthID for Text is designed to complement the most widely available AI text generation models and for large-scale deployment, while SynthID for Video builds on our image and sound watermark method to include all images in the generated videos. This innovative method integrates an imperceptible watermark without impacting the quality, precision, creativity or speed of the text or video generation process.
SynthID is not a silver bullet for identifying AI-generated content, but is an important building block for developing more reliable AI identification tools and can help millions of people make informed decisions on how they interact with AI-generated content. Later this summer, we plan to make SynthID open source for text watermarking, so developers can use this technology and incorporate it into their models.
How Text Watermark Works
Large language models generate sequences of text when given a prompt such as “Explain quantum mechanics to me like I'm five” or “What's your favorite fruit?” “. LLMs predict which token will likely follow another, one token at a time.
Tokens are the basic elements that a generative model uses to process information. In this case, it can be a single character, a word or part of a sentence. Each possible token is assigned a score, which corresponds to the percentage chance that it is the correct one. Tokens with higher scores are more likely to be used. LLMs repeat these steps to construct a coherent answer.
SynthID is designed to embed imperceptible watermarks directly into the text generation process. It does this by introducing additional information into the distribution of tokens at the point of generation by modulating the probability of token generation, all without compromising the quality, accuracy, creativity or speed of text generation.
SynthID adjusts the probability score of tokens generated by a large language model.
The final pattern of scores for the model's word choices combined with the adjusted probability scores is considered the watermark. This pattern of scores is compared to the expected pattern of scores for watermarked and non-watermarked text, helping SynthID detect whether an AI tool generated the text or whether it may have come from other sources.
A piece of text generated by Gemini with the watermark highlighted in blue.
The advantages and limitations of this technique
SynthID for text watermarking works best when a language model generates longer responses, and in a variety of ways, such as when asked to generate an essay, a theater script, or variations of an email .
It works well even under some transformations, such as cropping chunks of text, changing a few words, and light paraphrasing. However, its confidence scores can be significantly reduced when AI-generated text is completely rewritten or translated into another language.
The SynthID text watermark is less effective on responses to factual prompts because there is less opportunity to adjust token distribution without affecting factual accuracy. This includes prompts such as “What is the capital of France?” ” or requests for which little or no variation is expected, such as “recite a poem by William Wordsworth.”
Many AI detection tools currently available use algorithms to label and sort data, called classifiers. These classifiers often only perform well on particular tasks, making them less flexible. When the same classifier is applied across different types of platforms and content, its performance is not always reliable or consistent. This can lead to mislabeled text, which can cause problems, for example when text can be misidentified as AI-generated.
SynthID works effectively on its own, but it can also be combined with other AI detection approaches to provide better coverage of content types and platforms. Although this technique is not designed to directly prevent motivated adversaries such as cyberattackers or hackers from causing damage, it can make it harder for AI-generated content to be used for malicious purposes.
How Video Watermark Works
At this year's I/O, we announced Veo, our highest performing generative video model. Although video generation technologies are not as widely available as image generation technologies, they are evolving rapidly and it will become increasingly important to help people know whether a video is generated by AI or not .
Videos are made up of individual images or still images. We therefore developed a watermarking technique inspired by our SynthID for image tool. This technique embeds a watermark directly into the pixels of each video frame, making it imperceptible to the human eye, but detectable for identification.
Empowering people to know when they are interacting with AI-generated media can play an important role in preventing the spread of misinformation. From today, all videos generated by Veo on VideoFX will be watermarked by SynthID.
SynthID for video watermark marks each frame of a generated video
Bringing SynthID to the broader AI ecosystem
SynthID's text watermarking technology is designed to be compatible with most AI text generation models and to adapt to different content types and platforms. To help prevent widespread misuse of AI-generated content, we are working to integrate this technology into the broader AI ecosystem.
This summer, we plan to publish more about our text watermarking technology in a detailed research paper, and we will release the SynthID text watermark as open source through our updated version. Responsible Generative AI Toolkitwhich provides essential guidance and tools for building safer AI applications, so developers can use this technology and integrate it into their models.