Google DeepMind has made its AI watermarking technology SynthID open-source. This allows users of models other than Google’s to watermark and detect their content this way. The tech giant says it will encourage the conscious use of (Gen)AI.
For some time now, Google Deepmind has been very active in providing invisible or audible watermarks to text, audio, and images generated by GenAI for human perception. According to the company, this should ensure that the use of (Gen)AI is done in a more conscious, safer, and ethical way.
SynthID technology
With the toolkit SynthID, which is still in beta, developers can subtly watermark their AI-generated content. Human senses cannot detect this watermark, but it can be found via a specialized detector within SynthID. Thus, in addition to creating AI watermarks, SynthID can also detect them.
In summary, SynthID adds the watermark by using additional parameters in the AI-generated text, audio, or image. Developers insert a certain watermark configuration that subtly adds it during the creation process of the text, audio, or image.
Google DeepMind recently published more information about this technology in a paper in Nature.
More conscious use of GenAI
By making this technology open-source and no longer limiting it to its own Gemini LLMs, the tech giant wants to help other developers do the same. Thus, different companies can also integrate this watermarking technology within the output of their own (GenAI) LLMs.
Of course, there’s a little quid quo pro involved. Google Deepmind wants to benefit from the experiences of these developers and other community stakeholders and improve the tool through further feedback.
No cure-all
Google Deepmind specialists explicitly state that GenAI developers should not regard SynthID as a panacea. The tool does have some limitations here and there, they indicate. In their view, however, it is an important step toward more reliable AI identification tools. The ultimate intention is for GenAI users to be more aware of generated content.
SynthID is now open-source via an update to the Responsible Generative AI Toolkit. The tool is also available in Vertex AI for the Google image-LLMs Imagen 3 and Imagen 2. For GenAI audio, the technology is in the Google Lyria LLM. The tool for watermarking text is available through HuggingFace and GitHub.
Also read: Google shows off new set of AI tools for creating advanced videos and images