It was a dark, overcast day, with clouds hovering low.
The room was kept at the most appropriate temperatures and humidity, as usual. Yoko sat on a couch in an untidy manner, killing time with a silly game. But she would not talk to me.
I’m bored. I could not be bored more . . .
So begins a short story titled “A day when a computer writes fiction.” The story, in which boredom prods a computer into writing a novel, was created with the help of artificial intelligence.
Since 2013, a team of computer scientists led by Hitoshi Matsubara of Future University Hakodate in Hokkaido has been trying to get AI-based computers to write fiction.
Last year, the team entered two works, including the one above, as candidates for the Hoshi Prize, a novel-writing contest held in honor of Shinichi Hoshi (1926-1997), known as the “God of short fiction” in Japan.
In March, the team announced that it had been notified that one of the two stories had passed the first round of screening (the organizers did not disclose which). To win, the team would have to clear three more screenings by the panel of human judges.
Matsubara’s project raises a question that more and more AI scientists around the world are seeking answers to: Can computers truly be creative?
In fact, creativity seems to be the latest turf war in the ongoing human-versus-computer battle, after Google DeepMind, an AI program armed with “deep learning” technologies, outsmarted the world’s best human go player with a stunning 4-1 victory in March over master Lee Sedol.
Google, which acquired U.K.-based DeepMind in 2014, is no doubt the forerunner in the race to test the limits of “computer creativity.” Through its Magenta project, Google recently released its first batch of machine-made art and music.
Matsubara, who is also chairman of the Japanese Society for Artificial Intelligence, said the Hakodate team’s AI program is not yet at a level where it can craft readable plots and engaging characters by itself.
In AI-generated fiction, 80 percent of the creative process is actually handled by humans, he said.
But Matsubara foresees a future in which AI creators, be it novelists or painters, will outperform mediocre and average human counterparts.
“I thought what separated AI from humans was that AI had not yet acquired sensibility, intuition, and artistic sense,” Matsubara said. “And yet for decades, sci-fi novels have featured computer characters writing fiction. I thought that, if we worked really hard, we might be able to turn such a scenario into reality.”
Having a machine generate an interesting story from scratch has proved exceptionally difficult, however. To create a story just like Hoshi would have, the team made the computer read and analyze 1,000 of his short works. That made possible the “formalistic” copying of Hoshi’s writing style, including average sentence length and frequency of kanji used versus hiragana. But to enable a program to put together a convincing work of fiction that emulates Hoshi’s style represents a completely different challenge.
“Hoshi’s pieces can be broken down into several dozen patterns, especially the ochi (shocking or humorous endings) part. We are still analyzing story lines. The AI has created numerous stories for us, but they have all been crappy. So we concluded that, at this point, we are better at coming up with the plot,” Matsubara said.
In the piece about a bored computer writing fiction, the computer program added various descriptions, such as the weather, the name and sex of the lead character, and whether the character was meticulous or lazy, by choosing from options prepared by humans.
The AI program was also taught to choose adjectives that won’t contradict its earlier choices; if the character was untidy or out of shape, for example, descriptions used later in the story would back up such traits, he said.
The researchers did not proof the AI manuscript and the judges were not notified who the authors were.
“The fact that our submission passed the first screening means that, as Japanese prose, it reached a level considered not incomprehensible; it read like a story at least,” Matsubara said. “But since we have three more layers of screening to pass, we have a long road ahead.”
Yet Matsubara is not sure AI can someday achieve the creative heights of James Joyce or Vincent van Gogh.
“AI can learn the traits of creators from past data, so if they get access to a lot of data, they can produce works resembling the original creators. Whether AI can spawn geniuses like Picasso or Nobel laureates, it’s hard to say,” he said.
Geniuses they may not be, but if machine-made literary creations became marketable, they would eventually steal people’s jobs and careers.
The success of Microsoft’s AI “schoolgirl” character Rinna, which charmed Line users with its quirky and witty responses, has shown it is possible for people to grow attached to and spend hours communicating with computer programs.
The proliferation of AI-generated content has also raised many legal issues. Should all of AI-generated creations be copyright-protected? Who should get credit for them? Would it drive humans out of the market?
The Cabinet Office is discussing how to define the rights of AI-generated creations ranging from music and writing to visual arts. In the government’s annual intellectual property policy plan adopted in May, it mentioned AI content for the first time, advocating the need to create a legal framework for it.
By law, AI-developed content cannot be copyrighted, with creations worthy of protection defined as “works that creatively express thoughts or emotions,” a clause interpreted as only applying to human efforts.
Lawyer and copyright expert Kensaku Fukui, who served on the Cabinet Office panel that came up with a report on intellectual property rights in April, says AI-made creations should warrant some protection but not something as rigid as a copyright. If AI works were given the same rights as human works, they would quickly overwhelm the latter in volume, leading to the dominance of content by huge internet platforms like Google and Facebook, he warned.
“Google has all the key assets,” he said.
“It occupies the longest time users spend on the internet, as it has a market share of 90 percent or so as a search engine. Google now has enormous access to ‘big data,’ or data of what people like, with which it is capable of letting AI generate ‘marketable’ content.
“Soon, AI will start churning out billions of pieces of creative content. Since computers never get tired, they can produce an infinite number of works. If we gave copyrights to all of them, it would bar people from creating similar works, potentially threatening the human exercise of creativity, and hand these platforms an automatic monopoly on content,” Fukui said.
The power of AI might scare some people, but a professor at Tokyo University of the Arts envisions a future where AI can coexist with real artists.
Last month, the university organized a live concert in Tokyo, inviting four members of the Scharoun Ensemble, the famed chamber music unit of the Berlin Philharmonic Orchestra, and letting them perform with an AI-assisted piano.
Dubbed Disklavier, the automated piano developed by Yamaha Corp. is equipped with a computer program that allows it to play just like the late Russian pianist Sviatoslav Richter (1915-1997), who is known for his superb technique.
As the ensemble performed Franz Schubert’s Trout, a signature Richter piece, the AI-powered piano synced with the string players, picking up the ebb and flow of the music through a camera and microphone. In return, a gesturing, cartoon-like shadow of the imaginary AI pianist linked to the computer was projected on stage to give the human players a clue on how to adjust their rhythm and tempo.
“AI will continue to grow and prosper, regardless of whether we like it or not,” said Isao Matsushita, deputy president of the university and a composer himself, who organized the project in cooperation with Yamaha. “Unlike in shogi or chess matches, which come down to winning or losing, we wanted to present a way to collaborate with AI, by reviving the great music of Richter.”
From an engineering perspective, the biggest challenge was to make the piano sound like Richter and yet play in perfect sync with the human performers, said Motoichi Tamura, the Yamaha manager who led the project. “Musicians can feel the difference of 1/100th of a second. So it was challenging just to make them play together,” he said.
During a rehearsal in Germany in April, members of the Berlin Phil were irritated when the piano didn’t play as they wanted. But as they continued to practice, they grew attached to Diskavier, noting how the AI-powered instrument was “stubborn,” said Akira Maezawa, another project member at Yamaha.
“Artists have always been at the forefront of science, harnessing the power of science and technology to improve their instruments and music,” Matsushita said.
“We can use AI in education, by having students practice with an AI piano that adjusts to them, for example. Of course, there’s fear that we might be rendered useless if AI artists overtake us in the future. But we must find ways that such a scenario can be avoided.”
Takafumi Nakanishi, senior research fellow at the Center for Global Communications at International University of Japan, said that for the time being at least, AI will be a handy tool to help expand human creativity.
He referred to the example of Dramatica, a screenwriting program heavily used in Hollywood. The software makes suggestions on story structure and plausible characters every step of the way as the users write the script, utilizing vast data from past hits.
“Dramatica exemplifies the idea of humans supplementing their creativity with AI,” he said. “Dramatica itself is not creative. In the near future, using AI to boost knowledge will become a mainstream style of work for people.”