Watermark AI-generated text and video with SynthID

We’re announcing our groundbreaking watermarking method for AI-generated text and video, and how we’re bringing SynthID to key Google products

Generative AI tools and the technologies behind them based on enormous language models have captured the public imagination. From helping with professional tasks to enhancing creativity, these tools are quickly becoming part of the products millions of people apply every day.

These technologies can be extremely beneficial, but as their popularity increases, so does the risk of people causing accidental or intentional harm, such as spreading misinformation or phishing, if AI-generated content is not properly identified. That’s why last year we launched SynthID, our groundbreaking digital toolkit for watermarking AI-generated content.

Today we are expanding SynthID’s capabilities to watermark AI-generated text in files Gemini app and web experienceand video in Veo, our most competent generative video model.

SynthID for Text is designed to complement most commonly available AI text generation models and can be deployed at scale, while SynthID for Video relies on our image and audio watermarking method to include all frames in generated videos. This groundbreaking method embeds an unnoticeable watermark without affecting the quality, accuracy, creativity or speed of the text or video generation process.

SynthID is not a silver bullet for identifying AI-generated content, but it is an crucial building block for developing more resilient AI identification tools and has the potential to support millions of people make informed decisions about how they interact with AI-generated content. Later this summer, we plan to open-source SynthID for text watermarking so that developers can build with the technology and incorporate it into their models.

How text watermarking works

Huge language models generate sequences of text when prompted such as “Explain quantum mechanics to me as if I were five years old” or “What is your favorite fruit?” LLMs predict which token is most likely to follow another, one token at a time.

Tokens are the building blocks that the generative model uses to process information. In this case it can be a single character, word or part of a phrase. Each possible token is assigned a score, which is the percentage chance that it will be correct. Tokens with higher scores are used more often. LLMs repeat these steps to build a consistent response.

SynthID is designed to embed hidden watermarks directly into the text generation process. It does this by introducing additional information into the token distribution at the time of generation, modulating the probability of generating tokens – all without compromising the quality, accuracy, creativity and speed of text generation.

The final pattern of scores for both words selected by the model combined with the adjusted probability scores is considered as a watermark. This pattern of results is compared to the expected pattern of results for watermarked and unwatermarked text, which helps SynthID detect whether the AI ​​tool generated the text or whether it may have come from other sources.

Benefits and limitations of this technique

SynthID for text watermarking works best when the language model generates longer responses and in a variety of ways—for example, when prompted to generate an essay, a play script, or variations of an email.

It even handles some transformations well, such as trimming parts of text, modifying a few words, and subtle paraphrasing. However, its confidence level can be significantly reduced if the AI-generated text is carefully transcribed or translated into another language.

SynthID text watermarking is less effective in responding to fact-based 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 queries where little or no change is expected, such as “recite a poem by William Wordsworth.”

Many AI detection tools currently available apply algorithms to label and sort data, called classifiers. These classifiers are often only good at specific tasks, which makes them less elastic. When the same classifier is used across different types of platforms and content, its performance is not always reliable and consistent. This can lead to text being mislabeled, which can cause problems, for example where text is incorrectly identified as AI-generated.

SynthID works effectively on its own, but can also be combined with other AI detection methods to provide better coverage for different types of content and platforms. While this technique is not intended to directly stop motivated adversaries such as cybercriminals or hackers from causing damage, may make it more difficult to use AI-generated content for malicious purposes.

How video watermarking works

At this year’s I/O, we introduced Veo, our most powerful generative video model. Although video generation technologies are not as widely available as image generation technologies, they are developing rapidly and it will become increasingly crucial to inform people whether a video is generated by artificial intelligence or not.

Movies consist of single frames or still images. That’s why we developed a watermarking technique inspired by our SynthID tool for images. This technique embeds a watermark directly into the pixels of each video frame, making it hidden to the human eye but detectable for identification purposes.

Ensuring people know when they are interacting with AI-generated media can play an crucial role in preventing the spread of misinformation. From today, all videos generated by Veo on VideoFX will be watermarked SynthID.

Introducing SynthID into the broader AI ecosystem

SynthID’s text watermarking technology is designed to be compatible with most AI text generation models and to scale across a variety of content types and platforms. To prevent widespread misuse of AI-generated content, we are working to introduce this technology into the broader AI ecosystem.

We plan to publish more about our text watermarking technology in a detailed research paper this summer, and we will also open-source SynthID text watermarking via our updated A responsible set of tools for generating artificial intelligencewhich provides guidance and vital tools for building safer AI applications, so developers can build with the technology and incorporate it into their models.