Waterway marking text and video with synthesizer

By announcing our innovative water marking method for the text and video generated by AI and the way we introduce synthesis to key Google products

Generative AI tools – and large technologies of the language model behind them – stimulated public imagination. From help in work to increase creativity, these tools quickly become part of products used by millions of people in their daily lives.

These technologies can be extremely beneficial, but because they are becoming more and more popular to use, an increase in the risk of people causing accidental or intentional damage, such as the spread of disinformation and phishing, if the content generated by AI is not properly identified. That is why last year we introduced Synthid, our innovative digital set of marking tools generated by AI.

Today we are expanding the possibilities of Synthid to the water marking of the text generated by AI in Gemini's application and internet experienceand video in VEO, our most talented video generative model.

Synthid for Text has been designed to supplement the most common models of generating AI text and to be implemented on a scale, while Synthid for video is based on our image and sound imaging method to contain all frames in generated movies. This innovative method embeds an imperceptible watermark without affecting the quality, accuracy, creativity or speed of the process generation or video generation.

Synthid is not a silver ball to identify the content generated by AI, but is an important element of construction to develop more reliable AI identification tools and can help millions of people in making informed decisions regarding interaction with the content generated by AI. Later this summer, we are planning Synthid Open Source to mark the watering of the text so that developers can build with this technology and include it in their models.

How the text text works

Large language models generate text sequences when he received a prompt, such as “explain quantum mechanics to me as if I had five” or “What is your favorite fruit?”. LLM predicts which token most likely follows another, one token at once.

Toxes are component blocks whose generative model uses to process information. In this case, they can be one sign, word or part of the phrase. Each possible token is assigned a result that is a chance that it will be right. Tokens with higher results are more often used. LLM repeat these steps to build a coherent answer.

Synthid has been designed to set unnoticed watermarks directly into the process of generating text. He does this by introducing additional information in the distribution of the token at the generation point by modulating the likelihood of generating tokens – all without risking quality, accuracy, creativity or speed of text generation.

The final model of the results for both the model's words in combination with the corrected probability results is considered a water sign. This pattern of results is compared with the expected design pattern for the text in Apartin and unwanted, helping the synthesizer detect whether the AI ​​tool generated the text or can come from other sources.

The benefits and limitations of this technique

Synthid for the labeling of water text works best when the language model generates longer answers and in various ways – for example, when it is asked to generate an essay, theater script or varieties of E -Mail messages.

It works well even under some transformation, such as pruning text, modifying a few words and mild paraphrasing. However, his trust results can be significantly reduced when the text generated by AI is thoroughly prescribed or translated into another language.

Synthid Water -Water -Water -Water -Water Resave to reactions to actual prompts, because there are less possibilities to adjust the token distribution without affecting the actual accuracy. This includes hints like “What is the capital of France?” or inquiries in which small or no variants are expected, such as “recite the poem by William Wordsworth”.

Many currently available tools for detecting artificial intelligence use algorithms for labeling and sorting data known as classifiers. These classifiers often work well only in individual tasks, which makes them less flexible. When the same classifier is used on different types of platforms and content, its performance is not always reliable or consistent. This can lead to the erroneous significance of the text, which can cause problems, for example, in which the text can be incorrectly identified as generated by AI.

Synthid works effectively independently, but it can also be combined with other approaches to detect artificial intelligence to ensure better range of content and platform types. Although this technique is not built to directly stop motivated opponents such as cyber attacks or hackers, they cause damage, it It may make it difficult to use the content generated by AI for malicious purposes.

How the video watermark works

At this year's We/O we have announced VEO, our most talented video generative model. Although video generation technologies are not as widely available as image generation technologies, they are developing quickly and it will be more and more important to help people know if the video is generated by AI or not.

The films consist of individual frames or photos. So we have developed a water marking technique inspired by our synthesis of the image tools. This technique embeds the watermark directly into the pixels of each video frame, which makes it imperceptible to the human eye, but detected to identify.

Enabling people to know about interaction with the media generated by AI can play an important role in preventing the spread of disinformation. From today, all films generated by veo on Videofx will be marked by Synthid.

Transferring the synthesis to the wider AI ecosystem

Synthid text texts technology is designed to be compatible with most AI text generation models and for scaling various types of content and platforms. To prevent the widespread improper use of content generated by AI, we are working on introducing this technology to the wider AI ecosystem.

This summer we plan to publish more about our water marking technology in a detailed research document, and we synthid SMSs about water marking through our updated Responsible AI tool setwhich provides tips and necessary tools for creating safer AI applications, so that developers can build with this technology and include it in their models.