Daniela Hernandez reports in the Wall Street Journal:
In most cases, humans are still very much a part of the design process. When left to its own devices, AI software can create robotic or nonsensical things—or may miss the mark humans want to hit. They lack deep knowledge of the human creator’s intentions and motivations, as well as the cultural and emotional context to make (it) pop on the first try.
Artificial intelligence has been flexing its creative muscles recently, making images, music, logos and other designs.
In most cases, though, humans are still very much a part of the design process. When left to its own devices, AI software can create robotic or nonsensical things—or may miss the mark humans want to hit.
Researchers at Sony Corp.’s Computer Science Laboratory in Paris recently used AI to produce music in the styles of the Beatles and Johann Sebastian Bach. For instance, machines generated the melody and harmony of a vaguely Beatle-esque song, entitled “Daddy’s Car.” Lyrics, however—“Down on the ground/ Lines leads me to Nowhere Man/ Please mother drive me again in your car”—are the work of humans.
The melodies are good proofs-of-concepts, says CSL lab director François Pachet, but getting machines to create commercial hits autonomously is still years away. “Currently, there’s no automated system that’s even able to distinguish between a good song and a very, very good song,” he said.
Mike Cook, a computational-creativity researcher at Falmouth University in the U.K., is developing software that automatically generates videogame story plots based on a prompt Dr. Cook gives it. The more abstract the prompt is, the harder it is for the program, he said. Recently, the software, which is called Angelina, created a game about a bridesmaid trying to escape from evil women based on the prompt “You Only Got One,” according to Dr. Cook. Explaining its rationale for its music selections and story lines, the program wrote:
“This is a game about a bridesmaid. The objective is to reach the exit. Along the way, you must avoid the Daughters as they kill you. I use some sound effects ...like the sound of a wedding. People feel bored by bridesmaids sometimes, so I chose a aggressive piece of music...Let me know what you think.”Such results point to the limits of AI, experts say. “It’s very hard to change the outputs of these systems. They don’t do what you want them to do,” said Rebecca Fiebrink, a machine-learning researcher at Goldsmiths, University of London. They lack deep knowledge of the human creator’s intentions and motivations, as well as the cultural and emotional context to make art pop on the first try, she and other researchers said.
Alphabet Inc.’s Google in June introduced a project dubbed Magenta to explore how machine learning could be harnessed to create compelling art and music, according to Doug Eck, a Google machine-learning researcher and Magenta chief. Since then, Google engineers working on Magenta have started using the open-source system to craft short stories and simple jokes, he said.
Google declined to make its machine’s literary musings available for public consumption.