Examining Generative AI: Impact on Society / Generative AI Displays Misinformation at City Official Website; Towns Limit AI Services as Mistakes Creep in

The Yomiuri Shimbun

6:00 JST, May 15, 2024

The use of generative AI has been rapidly expanding in various areas of society. While people place high expectations on generative AI to improve convenience, confusion caused by the negative impact of generative AI is spreading in people’s daily lives. This is the second installment of a series which explores issues and potential countermeasures in the fields of education, government, business, medicine and sport.

***

Last November, Eiji Okazaki, 51, a local government official, sighs as he looks at the computer screen in his city office in Mitoyo, Kagawa Prefecture, a city on the Seto Inland Sea home to thriving manufacturing and agricultural industries.

He is the assistant manager of the city’s environmental sanitation division.

His eyes are fixed on text generated by ChatGPT, an interactive artificial intelligence service that the city government is running as an experiment to demonstrate the service’s capabilities.

In an answer to a question about how to dispose of garbage, the chatbot made a mistake.

Asked about where to dispose of air conditioners and televisions, ChatGPT responded, “Contact the environmental sanitation division of the Kanonji city office.”

The answer should have been to contact the Mitoyo city government or to take the items to a recycling company in neighboring Kanonji. ChatGPT, however, mentioned a nonexistent Kanonji city government department.

He also found that ChatGPT falsely labeled plastic natto packets as “flammable waste.” When asked how to dispose of a cell phone battery, the instructions included a method that could have caused the battery to catch fire. These instructions only deepened Okazaki’s concerns.

In June last year, Mitoyo city government began experimenting with an inquiry system for garbage disposal in cooperation with the University of Tokyo laboratory specializing in generative AI.

The city, which has a population of more than 60,000, has recently seen an increase in residents who speak foreign languages, including technical intern trainees.

The inquiry system is expected to answer citizens’ questions about garbage disposal 24 hours a day in different languages. Hopes were high for the system to reduce the workload.

However, “wrong answers” became a major stumbling block. The city set a goal of 99% or higher accuracy. After training the AI on more than 4,000 questions, the rate of correct answers improved from the initial 63% to 94%, but that was as far as it could improve.

“It is difficult to further increase the percentage of correct answers with the current technology,” said a person at the laboratory in charge of the system.

In December of last year, the city decided not to fully implement the system because it feared the system would increase its employees’ workload, as incorrect information could lead to noncompliance with waste disposal regulations or an increase in inquiries.

“It’s the role of the administration to deliver accurate information. We decided it was too early to implement the system for citizens,” said Takeshi Arawaki, 49, who was then head of the city’s digital promotion office and is now director of the regional strategy section.

The city is now considering using AI for other tasks in cooperation with the University of Tokyo laboratory.

Generative AI is expected to improve work efficiency due to its ability to produce human-like sentences. However, one of the risks of using AI is that it sometimes gives incorrect answers.

Tobetsu, Hokkaido, a town with a population of about 15,200, installed ChatGPT on all 200 of its employees’ computers in October last year.

Improving work efficiency is a pressing concern for the town, as many local government staff hold multiple jobs.

The town has used ChatGPT to generate ideas and create meeting materials, and the amount of work required to summarize meetings minutes has been cut in half.

While aiming to introduce the system in services for its citizens, the town learned that Mitoyo had “given up” on ChatGPT.

“Making good use of generative AI is a way for the town to survive. I was shocked to learn that we wouldn’t be able to introduce the system for resident services if the percentage of correct answers isn’t high enough,” said Hirohisa Usui, 41, chief manager of the town’s digital promotion division. “We have to set our sights even higher.”

The town said it has been promoting the use of AI since spring by appointing “missionaries” who are familiar with digital technology at each department, but services for citizens have yet to materialize.

Akihira Yoshimoto, director of the Association for Promotion of Public Local Information and Communication, believes that “local governments are expected to face further staff shortages and heavier workloads due to the declining population, falling birthrate and aging society. The expectation for AI is more apparent in rural areas.”

Meanwhile, problems such as misinformation persist.

“Since some mistakes can be fatal with today’s technology, generative AI needs to be used in a limited way to assist the staff,” he said.