How AI Can Lead to Better Predictions
By Paula Klein
As social scientists and labor economists debate the impact of machine learning and AI on human tasks, data scientists are drilling down deeper: By dissecting the intricacies of human decision-making, they hope to find the best uses of AI technologies.
Avi Goldfarb, Chief Data Scientist of the Creative Destruction Lab (shown above), discussed his preliminary research, the “Impact of Artificial Intelligence: Prediction versus Judgment,” at a recent MIT IDE seminar.
A Professor of Marketing at the Rotman School of Management, University of Toronto, Goldfarb offered a “framework for thinking about AI” that describes why some tasks are well-suited for machines while others require human judgment. In this view, the value of human judgment skills will usually increase as machines take on more basic work.
Goldfarb’s work, with Ajay Agrawal and Joshua Gans, focuses on AI as a predictive technology. “The current technology for machine intelligence is, in its essence, a prediction technology, so the economic shift will center on a drop in the cost of prediction,” he wrote in a 2016 Harvard Business Review article co-authored with his Rotman colleagues. The team defines prediction as “the ability to take known information to generate new information.” As machine intelligence lowers the cost of prediction, they wrote, new applications will emerge.
Among the growing list of predictive AI applications are loan default determination, tracking consumer habits, medical diagnosis, and object identification/classification. Even when AI is used in autonomous vehicles, Goldfarb said that prediction and probability are at work. By linking the car’s sensor data to the driving decisions made by the human drivers (e.g. steering, braking, and accelerating), the vehicle’s AI learns to predict how humans would react while driving.
Goldfarb also discussed the changing role of human judgment as machine learning improves. “Anything that affects a decision other than prediction or data” is judgment, he said. Usually, this includes ethics, knowledge, task definition, emotional intelligence and other factors considered to be human functions. In the future, “the most valuable skills will be those that are complementary to prediction — in other words, those related to judgment,” Goldfarb wrote in a recent Sloan Management Review article.
Many questions still remain unanswered about human and machine decision-making. For instance, when to delegate decisions to machines, and when it is best to ignore human judgment and base decisions purely on data. Goldfarb’s team is continuing to study these topics. Watch the full seminar video here.