Thanks to the relentless work of dedicated engineers, artificial intelligence, or AI, becomes smarter by the day.
But while computers become better at replicating human tasks, reading comprehension, an area where machines have yet to catch up, is declining among young people, suggesting a chilling future in which AI may put people out of work.
That is why Noriko Arai, a mathematician at the National Institute of Informatics, decided in November to change the direction of her project from teaching an AI to pass the entrance exam for Japan’s most prestigious school — the University of Tokyo, better known as Todai — to focusing on improving the reading comprehension of future generations using AI technology.
“AI engineers have always said that humans don’t have to worry because only menial jobs will be taken over by machines. But what about the people who do such jobs?” Arai, the project’s director, said in a recent interview. “Does AI create new jobs? Sure. But the real question is, how many of those who may lose their jobs to AI will be able to land a new one that requires high creativity?”
Launched in 2011, the Todai Robot Project aimed to develop AI that could attain a high score on the National Center Test for University Admissions, a standardized university entrance examination, by fiscal 2016, and be smart enough to pass Todai’s entrance exam in fiscal 2021.
This year, the AI, dubbed Torobo-kun, scored high enough to outperform 80 percent of all high school students in a mock exam. This means it has an 80 percent chance or more of getting into about 80 percent of private universities in Japan, including some highly competitive schools that many students dream of attending.
Arai’s decision to suspend the quest, albeit temporarily, was due in part to the fact that her research team achieved its ultimate goal: to find out what AI currently can and cannot do, and to predict the pros and cons of a society in which AI is pervasive.
“When I started the project in 2011, I had a feeling that the AI would not be able to enter Todai,” Arai said. “But when I asked AI researchers, about 85 percent said it would be able to pass Todai’s entrance exam by fiscal 2021. The answer wasn’t very clear even among professionals.”
A report by private think tank Nomura Research Institute in December 2015 estimated that about 49 percent of Japan’s workforce would be replaced by advanced robots and AI within 10 to 20 years.
Now, after five years of research, Arai believes that’s not far off the mark.
AI is excellent at answering fact-based questions on topics such as mathematics, physics and history.
The Torobo-kun AI relies on a sea of factual information gleaned from textbooks, dictionaries, past exam questions and even Wikipedia. It interprets a question by identifying key words and phrases, and uses an algorithm to select the most likely answer from the vast information in its database.
But because it comes to a conclusion without the kind of deep understanding humans can have about the meaning of a question, the AI cannot provide a correct answer requiring common knowledge not outlined in the question.
For example, consider the following phrase: “Daytime in Europe in summer is shorter than in Japan because Europe is located at a higher latitude than Japan.” The AI is unable to use the information provided to deduct that nighttime in Europe in winter is longer than in Japan. It simply can’t think creatively to extrapolate beyond the bounds of the question.
Another reason that discouraged Arai from pursuing the Todai test was the vast time and resources still necessary for Torobo-kun to become smart enough to pass.
“When I learned that we need 15 billion conversation data patterns to further improve the AI’s English exam score, I thought, ‘That’s not something we should pour our money into,'” Arai said.
In the end, she paints a bleak picture of the future: A majority of young people probably will not be able to get a job that requires a high level of reading competency, the key skill area in which AI has been slow to catch up.
As part of the project, Arai conducted a multiple-choice reading skill test on 15,000 students at public and private junior high and high schools. Students were asked to read a passage and select a phrase that best describes the meaning.
The results strongly suggest that many students don’t understand what is written in their textbooks, Arai said. This means, she added, many students lack the ability to visualize an image from a written sentence, essentially to think for themselves, which is exactly the kind of skill they will need to get an edge in an AI-saturated society.
“The worst-case scenario is, from 2021, when the number of people aged 22 or below will drop off significantly, the world will be short of workers, but the number of unemployed will increase as well,” she said.
Whether people can survive in a society dominated by AI depends on “whether they are able to exploit machines to supplement their abilities, and not follow what machines tell you to do,” Arai said. “It is my mission to help children acquire the ability, so that the outcome of technological advancement won’t be ill-fated.”