For scientists who dream of a future in which computers churn out great American novels, interactive fiction makes the perfect starting point. Choose Your Own Adventure books already require a little math to get them on the page, and researchers at the Georgia Institute of Technology have been developing a program that covers part of the creative side as well.
Named after the Arabic storyteller of legend, “Scheherazade-IF” (interactive fiction) has learned to "write" by reading human-produced stories and studying their narrative structures. Scientists fed the program hundreds of stories on two subjects: bank robberies and date movies. Scheherazade can’t comprehend stories like a human reader would, but it can scan them for significant events and memorize their sequences. For example, after processing so many date movie stories, the bot eventually learned that buying popcorn was a typical event that occurred before the movie while a kiss goodnight (hopefully) happened after.
To test the program’s writing ability, researchers asked three test groups to judge Choose Your Own Adventure stories abut bank robberies and date movies that had been produced by Scheherazade-IF, as well as stories written by a random generator, and a program that had been manually plotted by a human expert. Readers were asked to report any logistical errors and rank stories in terms of overall entertainment value and coherence.
Scheherazade’s bank robbery story performed just as well as the manually-generated one, but its date movie piece scored 17 percent lower. For both stories, the AI was ranked as more eloquent than the random generator, and Scheherazade's and the human-programmed stories scored similarly on coherence, enjoyment, player involvement, and story recognition.
While Scheherazade has made progress in terms of mapping story structures, it still relies on human input for all its sentences. It looks like authors' jobs are still safe from the imminent AI uprising—for the time being at least.