Google made a number of artificial intelligence (AI)-based announcements during its I/O 2024 keynote session late Tuesday. This includes new AI models, upgrading existing foundational models, and integrating AI features into Google products. The tech giant has also focused on AI safety and expanded the use of its native watermarking technology for AI-generated content called SynthID. This new toolkit now includes watermarks for text generated in the Gemini app and web client, as well as video generated in Veo.
SynthID was first released by Google DeepMind in August 2023 as a beta project that aims to correctly label AI-generated content. As more and more synthetically created media is shared as real, the need for such a solution is felt. It has been used to spread misinformation and cybercrime, including phishing. The tech giant first used the technology for watermarking AI-generated audio generated through its Lyria model in November 2023. The toolkit added a watermark to the audio as a waveform, making it detectable even though it was otherwise undetectable.
Now Google is expanding its use of SynthID to include text and video generation. Text created using Gemini apps and websites now displays a watermark. To achieve this, the toolkit targets the creation process itself. All text-based AI models are trained using tokens such as words, syllables, and phrases. The training process also includes understanding the flow of using these tokens, or the order in which they must follow to produce the most consistent response.
SynthID introduces “additional information about token distribution at the time of creation by regulating the probability of token creation.” Assign numbers to specific words in blocks of text created in this way. When detecting whether AI was used to generate text, we compare the score with an adjusted probability score to determine whether the source may be an AI model. DeepMind highlighted in the post that this technology is useful when AI generates long ad text because it is easier to see how the probabilistic model was created. However, for shorter factual responses, detection may be inaccurate.
The company is also recently expanding SynthID to release AI-generated video from Veo. Google says the technology will insert a watermark directly into the pixels of every video frame, which will be undetectable to the human eye but will be visible using a detection system.
In the coming months, Google plans to open source SynthID text watermarking through its Responsible Generative AI toolkit. A detailed research paper explaining text watermarking technology will also be published.
