Testing it out on the Fujitsu Research Portal

Fujitsu Research Laboratories has made a website available to the public where anyone can experience Fujitsu's latest technologies with just a simple user registration.Fujitsu Research Portal Please refer to this article for details on user registration.

This time, I would like to introduce hallucination detection technology that identifies explanations that sound plausible but are incorrect from among them. When using conversational generative AI like ChatGPT, have you ever been told a plausible lie? For example, let's look at the following case.

First, referring to this example, I asked, “Please tell me what a Take-copter is.” Then, I received this answer:
“The Take-copter is a legendary creature in Japan, a helicopter-like existence made of bamboo. According to legend, the Take-copter has legs on its bamboo nodes, and it is said that it can fly through the air by rotating these nodes. Also, the Take-copter is extremely small, almost invisible to the human eye. However, the Take-copter is a fictional creature, and it has not been scientifically proven whether it actually exists.”

In this way, it created an explanation that sounds plausible but is incorrect (called a hallucination). We know this is a lie because everyone knows what a Take-copter is, but what will happen in the future as the use of conversational generative AI progresses and it starts being used in corporate operations, legal services, and research and development of advanced technologies? If you are asking because you don't know, and plausible lies are mixed in, you won't know whether you can trust the conversational generative AI in the first place, right?
So, let's use the Fujitsu Research Laboratories developed AI trust technology that can deal with hallucinations in conversational generative AI. If you click the shield mark on the bottom left of the speech bubble, the following screen will appear.

The number displayed is called the hallucination score, and the closer the hallucination score is to 100, the higher the possibility that the answer contains errors. In other words, it clearly denied that the AI's answer is a lie. This technology will continue to be refined, but in preparation for a future where conversational generative AI is widely used, Fujitsu Research Laboratories is advancing the development of such hallucination detection technology. ※The Take-copter is a secret gadget that appears in Doraemon.

For those who want to know more, please take a look at the press release and technical blog below! (If you would like to request an interview, please feel free to contact us here.)

Developed AI trust technology that can deal with hallucinations in conversational generative AI and adversarial attacks that deceive AI, and started providing it on “Fujitsu Kozuchi (code name) - Fujitsu AI Platform”

Introduction to the features and functions of Fujitsu Kozuchi conversational generative AI