Large language models (LLMs) have become essential in generating high-quality synthetic text for various applications, leading to the need for effective ways to distinguish this content from human-written text. The SynthID-Text watermarking scheme offers a solution by embedding watermarks that allow synthetic text to be easily identified while preserving text quality. Unlike previous methods, SynthID-Text operates efficiently with minimal latency, integrating seamlessly into large-scale production systems. This approach tackles the challenge of detection without compromising on quality or requiring complex, resource-intensive operations. SynthID-Text employs an innovative Tournament sampling algorithm that modifies the token selection process during text generation. It embeds a statistical signature within the text, allowing for precise detection without needing to access the underlying language model. The watermarking technique is compatible with speculative sampling, commonly used to enhance the speed of text generation, preserving the efficiency of production systems. It supports different configurations for enhancing detectability or maintaining text quality, making it adaptable to various scenarios. The implementation of SynthID-Text in real-world applications, such as the Gemini response system, demonstrates its utility and scalability. During a live experiment involving around 20 million responses, the watermarked texts maintained indistinguishable quality from non-watermarked versions, as per user feedback. This validates SynthID-Text as a practical approach for watermarking AI-generated content at scale, offering a reliable means of content management and attribution in the evolving information ecosystem.