Google open-sourced its watermarking tool for AI-generated text

Google’s SynthID substance watermarking technology, a instrumentality the institution created to marque AI-generated substance easier to identify, is present disposable open-source done the Google Responsible Generative AI Toolkit, the institution announced connected X.

SynthID, which was announced past August, helps marque AI-generated output detectable by adding an invisible watermark into images, audio, video, and substance arsenic they’re generated. Google says the substance mentation of SynthID works by making the substance output somewhat little probable successful a mode that is detectable by bundle but not humans:

An LLM generates substance 1 token astatine a time. These tokens tin correspond a azygous character, connection oregon portion of a phrase. To make a series of coherent text, the exemplary predicts the adjacent astir apt token to generate. These predictions are based connected the preceding words and the probability scores assigned to each imaginable token.

For example, with the operation “My favourite tropical fruits are __.” The LLM mightiness commencement completing the condemnation with the tokens “mango,” “lychee,” “papaya,” oregon “durian,” and each token is fixed a probability score. When there’s a scope of antithetic tokens to take from, SynthID tin set the probability people of each predicted token, successful cases wherever it won’t compromise the quality, accuracy and creativity of the output.

This process is repeated passim the generated text, truthful a azygous condemnation mightiness incorporate 10 oregon much adjusted probability scores, and a leafage could incorporate hundreds. The last signifier of scores for some the model’s connection choices combined with the adjusted probability scores are considered the watermark.