AI Learns to Predict Our Reactions by Reading Fiction

Artificial intelligence systems are becoming more and more advanced, but they still have one major flaw that sometimes causes them to behave unnaturally. The problem is that AI can’t always predict human reactions with enough accuracy, but it looks like a team of researchers from Stanford University have come up with a new way to “teach” AI how to be better at it. To be more specific, the researchers have given their Augur knowledge database access to an online writing community named Wattpad, which includes more than 600,000 stories. The point of this is to allow learning algorithms to predict how people would react to different stimuli and in different situations.

“Over many millions of words, these mundane patterns [of people’s reactions] are far more common than their dramatic counterparts. Characters in modern fiction turn on the lights after entering rooms; they react to compliments by blushing; they do not answer their phones when they are in meetings.”

Augur certainly sounds promising, as a system based on an Augur-powered wearable camera has managed to identify people and objects with great accuracy in 91 percent of cases. When it came to predicting the reactions of humans, the system’s rate of success was a bit lower at 71 percent. Using books to teach computers new things certainly sounds like a natural way to go about it, and this isn’t the first time that such an experiment has been conducted. Not too long ago, Facebook’s AI received access to a massive 1.96GB stack of children’s books.