Watermark AI-generated text and video with SynthID

Announcing new watermarking techniques for AI-generated text and video and how to bring SynthID to key Google services

Generative AI tools and the large-scale language modeling technology behind them have captured the imagination. From assisting with work to enhancing creativity, these tools are quickly becoming part of the products millions of people use in their daily lives.

These technologies are highly beneficial, but as their use becomes increasingly widespread, AI-generated content risks causing accidental or intentional harm, such as spreading misinformation or phishing, if it is not properly identified. will increase. That’s why last year we launched SynthID, a new digital toolkit for watermarking AI-generated content.

We are currently extending SynthID’s capabilities to watermark AI-generated text in Gemini apps and web experiences and embed video with Veo, the most capable generative video model.

SynthID for Text is designed to complement the most widely used AI text generation models and be deployed at scale. SynthID for video, on the other hand, is built on image and audio watermarking methods to include every frame in the generated video. This innovative method embeds invisible watermarks without affecting the quality, accuracy, creativity or speed of the text or video generation process.

While SynthID is not a silver bullet for identifying AI-generated content, it is a critical building block for developing more reliable AI-identification tools and will improve how millions of people interact with AI-generated content. help you make informed decisions about what to do. Later this summer, we plan to open source SynthID for text watermarks. This allows developers to build with this technology and incorporate it into their models.

How text watermarks work

A large-scale language model can generate a sequence of text when given a prompt such as “Please explain quantum mechanics to me like I was five years old” or “What is your favorite fruit?” Generate. LLM predicts which token is most likely to follow another token, one token at a time.

Tokens are constructs that generative models use to process information. In this case, it can be a single character, a word, or part of a phrase. Each possible token is assigned a score. The score is the percentage probability that it is the correct token. Tokens with higher scores are more likely to be used. LLM repeats these steps to build a consistent response.

SynthID is designed to embed invisible watermarks directly into the text generation process. This is achieved by introducing additional information into the distribution of tokens at the point of generation by adjusting the probability that a token will be generated. All this is done without compromising quality, accuracy, creativity, or speed of text generation.

The final pattern of scores for both word choices of the model, 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, allowing SynthID to determine whether the text was generated by an AI tool or if it came from another source. It helps to detect whether

Advantages and limitations of this technique

SynthID for text watermarks works best when a language model is asked to generate longer responses in different ways, for example when it is asked to generate variations of an essay, a play script, or an email.

Works well even with some transformations, such as cutting out parts of the text, changing a few words, or mildly paraphrasing. However, if the AI-generated text is completely rewritten or translated into another language, the trust score can drop significantly.

SynthID’s text watermarks are less effective for responding to factual prompts because there is less opportunity to adjust token distribution without affecting factual accuracy. This includes prompts such as “Where is the capital of France?” Or queries where little or no variation is expected, such as “Read a poem by William Wordsworth.”

Many of the AI ​​detection tools available today use algorithms to label and sort data called classifiers. These classifiers often work well only for specific tasks and are therefore less flexible. When the same classifier is applied to different types of platforms and content, its performance is not always reliable or consistent. This can lead to issues such as text being mislabeled and text being incorrectly recognized as being generated by AI.

While SynthID works effectively on its own, it can also be combined with other AI detection approaches to provide better coverage across content types and platforms. Although this technology was not built to directly prevent motivated adversaries such as cyber attackers and hackers from doing harm, it does prevent motivated adversaries such as cyber attackers and hackers from using AI-generated content for malicious purposes. can be difficult.

How video watermarking works

At I/O this year, we announced Veo, our most capable generative video model. Although video generation technology is not as widespread as image generation technology, it is rapidly evolving, and it will become increasingly important for people to be able to recognize whether a video has been generated by AI.

Videos are made up of individual frames or still images. So we developed a watermarking technology inspired by the image tool SynthID. This technology embeds a watermark directly into the pixels of every video frame, making it imperceptible to the human eye but detectable for identification.

Providing people with information when consuming AI-generated media can play an important role in preventing the spread of misinformation. Starting today, all videos generated by Veo on VideoFX will be watermarked by SynthID.

Introducing SynthID to the broader AI ecosystem

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

We will be publishing more details about our text watermarking technology in a detailed research paper this summer. We also plan to open source SynthID text watermarks through an updated Responsible Generative AI Toolkit, which provides guidance and essential tools for creating more secure AI applications. , so developers can use this technology to build and incorporate it into their models.