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

Google’s SynthID text watermarking technology, aimed at making AI-generated text easier to identify, is now open-source, accessible through the Google Responsible Generative AI Toolkit. This allows AI developers to incorporate watermarking tools into their systems, facilitating responsible AI development.

Pushmeet Kohli of Google DeepMind emphasizes that watermarking is crucial for verifying the source of text, especially as AI is increasingly used for spreading misinformation and creating harmful content.

Regulatory moves are underway with California considering mandatory AI watermarking, mirroring an existing requirement in China, highlighting the global push for transparency in AI-generated content.

SynthID subtly modifies the probability of textual output by an AI in a way that is detectable by software but imperceptible to humans. It involves tweaking the likelihood scores assigned to each piece of generated text.

Despite challenges in managing short or significantly altered texts, Google has integrated SynthID into its Gemini chatbot without impairing text quality, accuracy, or creative integrity. The system is designed to be effective even on text as brief as three sentences.

While not foolproof, Google considers SynthID a critical step toward creating reliable AI identification tools, aiding users in making informed decisions about interacting with AI-generated content. #

--

Yoooo, this is a quick note on a link that made me go, WTF? Find all past links here.