Google has made the SynthID Text technology available to developers, allowing them to add “watermarks” to texts generated by artificial intelligence and to detect them. Now, this tool can be downloaded from the Hugging Face platform or accessed via Google’s updated Responsible GenAI Toolkit.
SynthID Text will help developers identify AI-generated content without compromising its quality
SynthID Text works by adjusting the probabilities for individual words or characters (tokens) generated by models, without affecting the accuracy or speed of the AI. As a result, a pattern of scores is formed, which acts as a kind of digital watermark. This method remains effective even after editing or partial modification of the text.
Google acknowledges that the technology has its limitations, especially when it comes to short texts or precise answers to queries. For example, with questions like “What is the capital of France?” the possibilities for adjustment are limited, as factual accuracy must be preserved.
SynthID is already integrated into Gemini models and is widely used in the company’s new products. Its mass adoption remains relevant, particularly in the context of new legislative requirements for labeling AI-generated content, which have already appeared in China and California.
