LOS ANGELES/TOKYO -- "Don't go beyond the cross section," the cab driver advised me before speeding off.
Just a few blocks past the upmarket cafes and high-end boutiques of downtown Los Angeles lies the city's less glamorous side. The stench hits you as soon as you step out of the car. People lie on the sidewalk under swarms of flies.
Homelessness is a major social problem in the United States; the number of people sleeping rough total 47,000 in Los Angeles alone.
To help fight HIV infection among LA's homeless, researchers at the University of Southern California have started a project seeking to analyze their behavior through artificial intelligence.
The idea was to use AI to shed light on how homeless people interact with each other. Specifically, the researchers wanted to identify the most influential people within the city's ranks of homeless, and seek their help in persuading others to undergo HIV examinations.
It soon became clear, however, that this was going to be no easy task. The personal relationship information that social workers had was insufficient for the computer system to work properly.
USC assistant professor Eric Rice, who works on the project, decided that to collect more reliable information about the personal networks among homeless people, he and his team would have to make repeated personal visits. But whenever a person the AI system had identified as "influential" left the community, the researchers would have to start the process from scratch.
The experience suggests that AI can be a very fussy research assistant.
Simply feeding huge amounts of data does not necessarily make the technology useful for real-life applications, and it is down to the users to find the best way of putting it to use.
Last October, Midori Fukushima of major Japanese supermarket operator Aeon was struggling with a project to develop an internal "AI call center" for the company's employees. The plan was to make the system take calls from staff and answer inquiries automatically.
When the project started, she recalls, the inquiry: "I can't open the attendance record screen," was answered with totally irrelevant response: "Regarding email log-in."
After nearly half a year, the system can barely return 140 correct responses.
Tokyo Institute of Technology professor Takao Terano said there are still "many businesses that are not yet ready to take advantage of AI in terms of preparing data or securing skilled staff."
At the Henn-na Hotel near the city of Nagasaki, where robots are supposed to carry out almost all the tasks usually handled by the staff, a process of trial and error continues. But Hideo Sawada, president of operator Huis Ten Bosch, is determined to make the concept work.
"Eventually, we want to have just one person operate the hotel," he said.
Staff numbers have already been reduced to six, just one-fifth the figure when the hotel opened. In March, the company opened the second Henn-na Hotel in Chiba, just west of the capital. Business may be ticking along, but insiders say there are still many issues to solve.
When the hotel introduced a conversation robot, for example, guests complained that it kept interrupting when they were chatting among themselves. Ironically, the robot's speech recognition system was too sensitive, making it a distinctly insensitive conversation partner.
Nikkei staff writers Hiromi Sato, Masayuki Nakagawa and Nikkei senior staff writer Natsuko Segawa contributed to this article.