A team of scientists at Google DeepMind has developed a tool that adds watermarks to text generated by large language models (LLMs), enhancing the ability to identify and track AI-created content.
LLMs are widely used in applications like chatbots and writing assistance, but identifying the source of AI-generated text remains a challenge, raising concerns about information reliability.
While watermarking is common in images and videos, applying it to text is complex, as any alteration can change meaning and quality. The newly introduced SynthID-Text employs a novel sampling algorithm to subtly bias word choice, embedding a signature recognizable by associated detection software.
In a study published in the journal Nature, researchers Sumanth Dathathri and Pushmeet Kohli reported that SynthID-Text showed improved effectiveness compared to existing methods. Importantly, it requires minimal additional computational power, facilitating its implementation.
The ability to identify synthetic text can help mitigate accidental or deliberate misuse. The authors emphasize that SynthID-Text maintains text quality while allowing for high detection accuracy, presenting a technically robust solution for identifying AI-generated text.
Experts highlight the need for such technologies as current systems for detecting AI-generated documents have low accuracy rates. However, widespread adoption faces challenges, particularly as watermarks can be vulnerable to modifications that reduce their detectability.