Google DeepMind has introduced SynthID, a tool for watermarking and detecting AI-generated content.
This technology embeds invisible watermarks into images, audio, text, and video created by generative AI.
With this feature, Google aims to mitigate misinformation and misattribution issues associated with generative AI technologies.
SynthID Text has been open-sourced to enable developers to implement watermarking in their text generation models.
The watermarking process functions as a logits processor, enhancing the generation pipeline of large language models while maintaining output quality.
Detection of these watermarks is probabilistic, utilizing a Bayesian detector that can be customized according to user preferences.
SynthID Text is designed to improve the identification of AI-generated content but is not entirely effective against motivated adversaries.
Despite its limitations, SynthID represents a significant advancement in promoting responsible AI usage and enhancing content verification across various platforms.
Be the first to comment