Google made a large number of artificial intelligence (AI)-based announcements late Tuesday during its I/O 2024 keynote session. These include new AI models, upgrades to existing foundation models, integration of AI features into Google’s products, and more. The tech giant also focused on AI safety and expanded the usage of its native watermarking technology for AI-generated content, dubbed SynthID. This new toolkit will now be embedding watermarks for text generated by the Gemini app and web client, and videos generated by Veo.
SynthID was first unveiled by Google DeepMind in August 2023 as a beta project aimed at correctly labelling AI-generated content. The need for such a solution was felt due to the rise of instances where these synthetically created media were shared as real. These were used to spread misinformation and cybercrimes such as phishing. The tech giant first used this technology in November 2023, when it was used to watermark AI-generated audio created through its Lyria model. The toolkit added watermarks as a waveform to the audio to make it imperceptible yet detectable.
Now, Google is expanding the usage of SynthID to include text and video generation. It will now watermark the text generated using the Gemini app and the website. For this, the toolkit will target the generation process itself. Every text-based AI model uses tokens — which can be words, syllables, or phrases — to train. The training process also includes understanding the flow of using these tokens, or the order the tokens should follow to generate the most coherent response.
SynthID introduces “additional information in the token distribution at the point of generation by modulating the likelihood of tokens being generated.” This way it assigns a number to certain words in a block of generated text. When detecting whether AI was used to generate the text, it checks the score against its adjusted probability scores to determine whether the source could be an AI model. DeepMind highlighted in a post that this technique is useful when an AI generates long creative text as it is easier for probability models to check how it was created. However, for shorter factual responses, the detection may not be as accurate.
The company is also expanding SynthID to recently unveiled Veo’s AI-generated videos. Google said the technology will embed watermarks directly into the pixels of every video frame which will be imperceptible to the human eye but will show up when a detection system is used.
In the coming months, Google plans to open-source SynthID text watermarking through its Responsible Generative AI toolkit. It will also publish a detailed research paper explaining the text watermarking technology.
