Google makes AI text watermarking tool open source

Google DeepMind is making its AI watermarking tool for text generation available to external developers.

The open source launch of SynthID-Text, which aims to preserve text quality and enable high detection accuracy, comes as the increasingly sophisticated large language models (LLMs) behind platforms like ChatGPT make it more and more difficult to distinguish AI-generated text from human-written content.

The move comes after Google conducted a live experiment that assessed feedback from nearly 20 million Gemini6 responses to demonstrate the feasibility of watermarking in large-scale production systems.

The tech giant revealed details about its production-ready text watermarking scheme in a paper written by London-based scientists which was published in Nature this week.

The company’s wider SynthID technology suite, which it first announced at a conference in the US last year, uses deep learning models and algorithms for both watermarking and identifying AI-generated content across images, audio, text, or video.

“Given the widespread use of LLMs in education, software development and web content generation, identification and attribution of LLM text is critical to ensure safe and responsible use of the technology,” wrote the authors of the Nature report, which included DeepMind research scientist Sumanth Dathathri.

Google says that while watermarking can help to identify synthetic text, it has not yet widely been adopted due to “stringent quality, detectability and computational efficiency requirements”.

According to the company, valuations across multiple LLMs “empirically show” that SynthID-Text provides “improved detectability over comparable methods”, with standard benchmarks and human side-by-side ratings indicating no change in LLM capabilities.