Google is open-sourcing their SynthID text watermark which will make it easier to identify AI-generated text. Developers interested in testing out this new feature can access it through the Google Responsible Generative AI Toolkit and Hugging Face.
Pushmeet Kohli, Google DeepMind’s VP of research has stated that this tool will be helpful for AI developers to identify if text outputs were generated from their LLMs. It helps developers to create a more responsible AI. OpenAI has also spent years researching methods related to watermarking. However, due to certain commercial and technical issues, OpenAI has delayed the release of its watermarking technology. So this means Google is not the only tech company working on this kind of watermarking technology.
SynthID was announced in 2023 and its sole purpose was to detect AI-generated output. This is done through the addition of a watermark that cannot be seen on audio, videos, images and text. Google has stated that the SynthID text watermark makes the text less probable as it cannot be detected by humans. Instead, it can be detected by software instead.
SynthID Text Watermark: More on this new open-source tool
This technology has already been integrated into Google’s Gemini chatbot. The tech giant has stated that even with this integration, the quality, creativity, speed and accuracy have not been compromised, an issue prevalent in most watermarking systems. Furthermore, Google states that the SynthID text watermark can apply to texts that are paraphrased, cropped or even if it is only three sentences long.
However, the technology does not come with no setbacks. It has trouble with shorter texts or translated and rewritten texts. SynthID will not be able to identify texts that have been translated from other languages. For instance, if a person translates text from English to Spanish with AI, then SynthID will be unable to identify the text.
Suggested:
Google Flights Update Lets Users Get The Cheapest Flights
Smart Replies: Google’s New Feature Now Rolling Out
Image Credits: Screenshot Taken From Google DeepMind
