Google DeepMind announced the open source of SynthID technology and the integration of it into the Gemini model. SynthID for Text is a technology that adds watermarks to text generated by AI, which is achieved by fine-tuning the next marking probability predicted by LLM without affecting text quality and creativity. This technology can reliably identify AI-generated text without changing the text content, providing a new solution to the challenge of AI-generated content. It is suitable for multiple languages and has been integrated into Gemini apps and websites for user convenience.
SynthID adds a hidden watermark to the generated text, which can be recognized even after editing. Google DeepMind claims its performance is better than existing text watermark systems and has opened the technology to facilitate use and improvements by developers and researchers. The open source release of this technology will help promote the reliability and transparency of AI-generated content and contribute to building a safer and more reliable AI ecosystem.
Recently, Google DeepMind is integrating its SynthID technology into its Gemini model and releasing it as an open source project. SynthID for Text uses complex processes to interfere with text generation of large language models (LLMs).
SynthID slightly adjusts the probability score of the next tag predicted by LLM, but only if it does not affect the quality, accuracy, or creativity of the output. This process repeats all generated text, forming a watermark.
According to Google DeepMind, this technology can be applied to at least three sentences. For longer text, the watermark becomes more robust and accurate. While this approach works in a variety of languages, it has some weaknesses when editing AI text.
Google DeepMind has integrated SynthID into Gemini applications and websites to add watermarks and recognition to generated text. The technology can also be used as an open source project on GitHub, Google Responsible Generative AI Toolkit, and Hugging Face.
Google DeepMind claims SynthID performs better than existing text watermark systems. Previously, Google DeepMind has launched SynthID for images, voice and music.
The move to Google DeepMind's open source SynthID technology marks an important step in combating the abuse of AI-generated content and reflects its commitment to responsible AI development. In the future, with the continuous improvement of technology and the expansion of application scope, SynthID will play a greater role in ensuring information security and maintaining the authenticity of content.