A paper published in the latest issue of Nature magazine claims that Google's deep thinking team has developed a tool that can add watermarks to text generated by large language models, improving the ability to identify and trace synthesized content. The Large Language Model (LLM) is a widely used artificial intelligence (AI) tool that can generate text for chatbots, article writing, and other purposes. However, it is difficult for people to identify and trace the source of AI generated text, which calls into question the reliability of the information. Watermarks are considered to solve this problem, but the strict requirements for quality and computational efficiency in production systems hinder their large-scale application. The deep thinking team has developed a system called SynthID Text, which adds watermarks to AI generated text using a new sampling algorithm. This tool cleverly shifts the vocabulary selection of LLM through sampling algorithms and inserts a signature that can be recognized by relevant detection software. This can be achieved through a "twisted" path (which can improve watermark quality but slightly affect output quality), or through another "non twisted" path that can preserve text quality. Deep thinking has evaluated the detectability of these watermarks on multiple public models, and the results show that SynthID Text has better detectability than other current methods. They also evaluated the quality of these texts using Gemini LLM's nearly 20 million online conversations and ultimately concluded that the non distorted watermark form would not reduce the quality of the text. In addition, the impact of using SynthID Text on the computing power required for LLM operation is negligible, thereby reducing application barriers. This study demonstrates that adding watermarks to generative text is feasible and can enhance the accountability and transparency of AI usage. (New Society)