In the fast-evolving world of AI, keeping track of AI-generated content has become more important than ever. Google DeepMind has taken a bold step by open-sourcing SynthID, an advanced tool designed to watermark AI-generated text. While SynthID has the potential to watermark various types of media like images, videos, and audio, its initial release focuses on text and is currently available only for businesses and developers. The company plans for a wider adoption of the tool to make AI-generated content easier to identify. Both individuals and enterprises can now get this tool via Google’s enhanced Responsible Generative AI Toolkit.
It was in a recent post on X that Google DeepMind revealed that the SynthID’s text watermarking tool is now open and free for developers and businesses. In addition to being accessible through the Responsible GenAI Toolkit, SynthID can also be downloaded from Google’s Hugging Face listing. AI-generated text is rapidly filling the online space. Earlier this year, a study from Amazon Web Services’ AI lab reported that up to 57.1% of all sentences translated into two or more languages on the internet may have been created using AI tools.
Though AI chatbots filling the internet with AI-generated text may seem like harmless spam, they pose a serious risk. In the wrong hands, AI can be used to spread misinformation and sway public opinion, influencing real-world events like elections and fueling propaganda against public figures. Detecting AI-generated text has been particularly challenging, primarily because watermarking individual words isn’t feasible. Even if it were possible, malicious users could easily rephrase the content. However, Google DeepMind’s SynthID introduces an innovative approach to watermarking AI-generated text. It employs machine learning to predict which words are likely to follow a given word in a sentence.
By analyzing the content generation styles of various AI models, SynthID can predict the word that should come next and substitute it with a synonym from its database. This watermarking tool embeds these words throughout the entire text. Later, when looking for AI-generated content, it checks the frequency of these specific words to verify its authenticity.
Importantly, for images and videos, SynthID embeds a watermark directly into the pixels of the frames, making it invisible to the naked eye while still detectable by the tool. In the case of audio, the audio waves are converted into a spectrogragh, where the watermark is added to the visual representation. However, these features are currently exclusive to Google and not available to the public.
- OpenAI Announces a Major Shift Towards For-Profit Operations - January 6, 2025
- xAI Secures $6 Billion in New Funding, Valuation Soars Past $40 Billion - January 2, 2025
- OpenAI Reveals o3 Reasoning AI Model, Official Launch Set for Early Next Year - December 27, 2024
