Google made a number of artificial intelligence (AI)-based announcements during its I/O 2024 keynote on Tuesday night. This includes new AI models, upgrades to existing core models, integration of AI features into Google products, and more. The tech giant has also focused on AI security and expanded the use of its original AI-generated content watermarking technology, called SynthID. This new toolkit will now embed watermarks for text generated by the Gemini app and web client and videos generated by Veo.
SynthID was first introduced by Google DeepMind in August 2023 as a beta project aimed at correctly tagging AI-generated content. The need for such a solution was felt due to the increase in cases in which these synthetically created media were shared as real. They have been used to spread disinformation and cybercrimes such as identity theft. 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 tools watermarked the audio as a waveform to make it invisible yet detectable.
Now Google is expanding the use of SynthID to generate text and video. It will now watermark text generated using the Gemini app and website. For this, the tools will target the generation process itself. Each text AI model uses tokens – which can be words, syllables or phrases – for training. The training process also involves understanding the flow of using these tokens, or the order the tokens should follow to generate the most coherent response.
SynthID introduces “additional information into the token distribution at the time of generation by modulating the probability of token generation.” In this way, it assigns a number to certain words in the block of generated text. When detecting whether artificial intelligence was used to generate the text, it checks the result against its adjusted probability scores to determine whether the source could be an artificial intelligence model. DeepMind noted in the post that this technique is useful when artificial intelligence generates long creative text because it’s easier for probabilistic models to verify how it was created. However, for shorter factual responses the detection may not be as accurate.
The company is also extending SynthID to Veo’s recently introduced AI-generated videos. Google said the technology will embed watermarks directly into the pixels of each video frame that will be invisible to the human eye but will appear when a detection system is used.
In the coming months, Google plans to open-source SynthID watermarking through its Responsible Generative AI tool. They will also publish a detailed research paper explaining text watermarking technology.
