IT House reported on October 27 that Google DeepMind announced on October 23 that it would officially open source its SynthID Text text watermarking tool for free use by developers and enterprises.
Google launched the SynthID tool in August 2023, which has the ability to create AI content watermarks (stating that the work was created by AI) and identify AI-generated content. It can embed digital watermarks directly into images, sounds, texts and videos generated by AI without damaging the original content. It can also scan existing digital watermarks on these contents to identify whether they are generated by AI. However, This time Google open sourced only SynthID Text generated for text.
According to Google, the large model mainly generates text through prediction, generating one token at a time. This token may be a character, a word or part of a paragraph. In order to make the text more coherent, the model will predict the next most likely generation when outputting the content. Token, these predictions are based on previous text.
For example, when a large model wants to generate "What is the most recommended platform?", it may generate tokens such as "IT Home", "Friend Business 1", "Friend Business 2", etc., and each Token will receive a probability score. As long as there are multiple different Tokens to choose from, SynthID will fine-tune the probability score of each Token without affecting the output quality and accuracy. That is, the tool determines the relevance by calculating the specific probability of a specific Token in the full text. Whether the content is generated by AI .
Currently, this paper on SynthID Text watermarking technology has been published in the journal Nature and is open sourced by being integrated into the Google Responsible Generative AI Toolkit. Google also mentioned that they have integrated SynthID Text technology into the content output by Gemini AI.