Google has expanded its AI watermarking technology, SynthID, to include AI-generated text and video. This announcement was made during the Google I/O 2024 developer conference, where DeepMind CEO Demis Hassabis introduced the enhanced system. Initially designed for AI-generated images and audio, SynthID now extends its capabilities to additional media formats, improving the identification of AI-created content.
SynthID’s New Capabilities
SynthID, which was first introduced in August of last year, initially embedded invisible watermarks in AI-generated images. These watermarks are undetectable to the human eye but can be identified by specialized systems. Unlike other watermarking protocols such as C2PA, which use cryptographic metadata, SynthID employs a unique imprinting method. The latest update allows SynthID to mark AI-generated video and text, expanding its scope and utility.
How SynthID Works
SynthID for text is designed to complement widely-available AI text generation models and can be deployed at scale. It adjusts the probability score of tokens generated by a large language model, creating a pattern that serves as the watermark. This method is less effective on responses to factual prompts due to fewer opportunities to adjust the token distribution without affecting factual accuracy. For video, SynthID builds upon its image and audio watermarking method to include all frames in generated videos. This innovative approach embeds an imperceptible watermark without impacting the quality, accuracy, creativity, or speed of the text or video generation process.
Importance of AI Watermarking
The need for watermarking AI-generated content has increased as the technology becomes more prevalent. AI tools have been misused for malicious activities, including spreading political misinformation, fabricating statements, and creating nonconsensual explicit content. The enhanced SynthID tool aims to mitigate these risks by providing a reliable method to identify AI-generated material, thereby supporting efforts to maintain content integrity and authenticity.
Microsoft already watermarks images generated by Bing Image Creator. The addition of the watermark is intended to help users identify AI-generated content. Meta introduced mandatory AI watermarks across its platforms. OpenAI also added watermarking – a method of adding metadata to digital content to establish its origin– to bolster security measures against deepfakes produced by their AI models.
Government and Industry Safeguards
SynthID is part of a broader initiative to develop safeguards against the misuse of AI technology. The Biden administration has directed federal agencies to establish guidelines for AI use, emphasizing the importance of such protective measures. Google’s efforts with SynthID align with these directives, aiming to create a safer digital environment as AI continues to evolve.
Starting today, all videos generated by Veo on VideoFX will be watermarked by SynthID. Later this summer, Google plans to open-source SynthID for text watermarking, allowing developers to build with this technology and incorporate it into their models. Additionally, Google will publish a detailed research paper on its text watermarking technology and update its Responsible Generative AI Toolkit to include SynthID.
Last Updated on November 7, 2024 8:23 pm CET
