Hitachi's "Watermarking" Tool Detects AI-Generated Writing

Hitachi Ltd. has developed the world's first artificial intelligence (AI) watermarking tool that can help distinguish between sentences created by generative AI and those written by humans, aiming to protect against abuses such as the spread of false information, Tatsuro Ando reported for Mainichi Japan.

Hitachi has developed a method that adds expected words to digital watermark word lists. I Photo: Gnsin Wikimedia Commons

Generative AI has led to challenges, such as schools being unable to discern whether reports were written by students themselves or outsourced to AI technology. When people write sentences, their word choice is influenced by various factors.

Hitachi has developed a method that adds expected words to digital watermark word lists.

When AI generates a sentence, it is designed to automatically select certain words from these lists.

Safety and other factors were enhanced by layering these lists. AI-generated sentences containing the telltale watermarks are more likely to include words that match those in the lists, making it possible to distinguish them.

Digital watermarking technologies that allow the generative source of information to be detected are advancing for mediums such as images and video.

If the use of this technology in written texts becomes widespread, it will reportedly be possible to determine if AI was involved, even in cases where the texts were copied repeatedly.

One of the developers, Koichi Nagatsuka, said, "As debate over the rules and regulations of generative AI continues, we'd like to first improve the accuracy of the watermarking technology so that it can be put to practical use."