Two new scientific papers outlining different approaches to artificial intelligence (AI) have both concluded that reading books is the best way to learn human behavior. “Using Stories to Teach Human Values to Artificial Agents,” from a team at the School of Interactive Computing at Georgia Institute of Technology, describes “using stories to generate a
value-aligned reward signal for reinforcement learning agents.” Meanwhile, a Stanford University team mining Wattpad stories has developed “Augur: Mining Human Behaviors from Fiction to Power Interactive Systems.”
The Georgia team focus chiefly on “preliminary work” developing theoretical and formal structures so that “crowdsourced narrative examples can be used to train an agent to act in a human-like fashion.” The Stanford team claim something a little more practical. They state that they have developed Augur, “a broad knowledge base of human behavior by analyzing more than one billion words of modern fiction,” which could power applications “from smart homes that prepare coffee when we wake, to phones that know not to interrupt us during important conversations.” To create Augur, they indexed “more than one billion words of fiction writing from 600,000 stories written by more than 500,000 writers on the Wattpad writing community,” using data mining DSL and activity models they developed. They then tested the resulting activity models and predictions, often via social media, with a claimed success rate of up to 94 percent in certain applications.
You wonder what might happen if Google or Amazon turned such agents loose on the huge volumes of text they have digitized. Better predictive algorithms for Amazon’s marketing machine? Better ad targeting for Google? Skynet? The Matrix? The Singularity? If you want to know what the future might hold, there’s probably a story about it somewhere. Facebook, meanwhile, has reportedly set its AI to learn human behavior by reading children’s books.
Meanwhile, both approaches pay fundamental tribute to the power of stories and storytelling. The Georgia team concur that “an artificial intelligence [or presumably a human one] that can read and understand stories can learn the values tacitly held by the culture from which the stories originate.” The Stanford team, meanwhile, maintain that “fictional human lives provide surprisingly accurate accounts of real human activities.” So if computer scientists believe that stories provide the best guides to human behavior and morals even for computers, then perhaps politicians, businesspeople, other scientists, and the public at large, ought to turn to, and respect, them more too. Because if things go on like this, the novel you read today may be powering your toaster – or deciding your court case – tomorrow.
I suppose that the accuracy of this is tied to the degree to which our stories are “true to life.” Do most authors write about what is or do they write about what they hope or fear will be. If so, these algorithms will need to implement some sort of interpolation strategy to get at the truth of human behavior.
One of the major things that makes the social sciences so difficult is the fact that the objects of our study are self-aware and often aware of being studied. So, unlike an amoeba under the gaze of a biologist, a person may behave differently when being studied. Computer science needs to learn from the social sciences before setting out on investigations such as this.