Can AI Nudge Us to Make Better Choices?
Executive Summary
The fields of behavioral science and machine learning provide some promising techniques for creating more “emotionally intelligent AI that organizations are putting to work to produce better outcomes. The key is to boost AI’s emotional by training algorithms to mimic the way people behave in constructive relationships. Thus new systems note changes in people’s patterns and nudge them to see if they want to make a correction; they encourage self-awareness by helping people compare their performance with others; they apply game theory to accept or challenge conclusions; and they help decision makers focus on the right right task at the right time.
The behavioral revolution in economics was triggered by a simple, haunting question: what if people don’t act rationally? This same question now vexes the technology field. In the online world, once expected to be a place of ready information and easy collaboration, lies and hate can spread faster than truth and kindness. Corporate systems, too, elicit irrational behavior. For example, when predicting sales, employees often hide bad deals and selectively report the good ones. AI stands at the crossroads of the behavioral question, with the potential to make matters worse or to elicit better outcomes from us. The key to better outcomes is to boost AI’s emotional quotient — its EQ. How? By training algorithms to mimic the way people behave in constructive relationships.
Whether or not we care to admit it, we build relationships with apps. And apps, like people, can elicit both positive and negative behaviors from us. When people with high EQ interact with us, they learn our patterns, empathize with our motivations, and carefully weigh their responses. They decide to ignore, challenge, or encourage us depending on how they anticipate we will react.
Insight Center
AI can be trained to do the same thing. Why? Because behaviors are more predictable than we like to think. The $70 billion weight-loss industry thrives because diet companies know that most people regain lost weight. The $40 billion casino industry profits from gamblers’ illogical hope of a comeback. Credit card companies know it is hard for people to break their spending habits.
While it’s still quite early, the fields of behavioral science and machine learning already provide some promising techniques for creating higher-EQ AI that organizations are putting to work to produce better outcomes. Those techniques include:
Noting pattern breaks and nudging. People who know you can easily tell when you are breaking a pattern and react accordingly. For example, a friend may notice that you suddenly changed your routine and ask you why. The Bank of America online bill paying system similarly notes pattern breaks to prevent user keying errors. The system remembers the pattern of payments you’ve made in the past and posts an alert if you substantially increase your payment to a vendor.
Encouraging self-awareness with benchmarks. Bluntly telling individuals they are performing poorly often backfires, provoking defensiveness rather than greater effort. A more diplomatic method simply allows people to see how they compare with others. For instance, a major technology firm used AI to generate more accurate sales forecasts than the sales team did. To induce the team to course-correct, the system provides each team member with personalized visualizations showing how their forecasts differ from the AI forecast. A simple nudge then inquires why this might be the case. The team member can provide a rational explanation, avoid providing feedback, or claim that the AI is incorrect. The AI learns about the substance and timing of the individual’s reaction, weighs it against the gap in the two forecasts, and can choose an appropriate second-order nudge.
Using game theory to accept or challenge conclusions. Imagine being on a team that must find errors in over 100,000 mutual fund transactions a day. A fund managing a trillion dollars in assets is tackling this daunting problem with AI. The first version of the AI scored potential errors (called “anomalies”) by risk and potential cost, then queued the riskiest anomalies first. The system then tracked the time the analyst spent on each anomaly. It was assumed that analysts would spend more time on the risker anomalies and less time on the “no-brainers.” In fact, some analysts were flying through the riskiest anomalies, reaching suspiciously fast conclusions.
In most massive screening systems, the rate of false positives is often extremely high. For example, secret teams from the Department of Homeland Security found that the TSA failed to stop 95% of inspectors’ attempts to smuggle weapons or explosive materials through screening. Mutual fund analysts scouring countless transactions, like TSA screeners dealing with thousands of passengers, their eyes glazing over, simply glide over anomalies.
The fund is tackling this dangerous, though highly predictable, behavior with an algorithm employed by chess playing programs. This modified version of sequential game theory first monitors whether the analyst concludes that an anomaly is a false positive or decides to spend more time on it. The AI, playing the role of a chess opponent, can decide to counter by accepting the analyst’s decision or challenging it.
Choosing the right time for insight and action. By any standard, Jeff Bezos is a master decision maker. In a recent interview with Bloomberg TV’s David Rubenstein, he described his framework for making decisions. When approached about a complex decision late in the afternoon he often replies, “That doesn’t sound like a 4 o’clock decision; that sounds like a 9 o’clock [in the morning] decision.”
My firm’s sales team A/B tested the right time of day to maximize responses to prospecting emails and found a dramatic difference in response rates between messages sent Tuesday morning and Friday afternoon. Many consumer messaging systems are tuned to maximize yield. The tuning algorithm can be enhanced to determine the type of decision to be made and the tendency of users to respond and make better choices. For example, decisions that need more thought could be presented at a time when the decision maker has more time to think — either through prediction or by the user’s scheduling.
Could higher-EQ AI help bring more civility to the internet? Social media companies might do well to consider a distinction Western business people soon learn when negotiating with their Japanese counterparts — “honne” (what one feels inside) versus “tatemae” (what one publicly expresses). A shared understanding of the distinction between what one feels and what one is expected to say leads to fewer miscalculations. An algorithm based on that distinction might conceivably be developed to address the predictable tendencies of people to say and do things under the influence of crowds (even if virtual ones) that they would otherwise hesitate to do. Someone preparing an inflammatory, misleading, or cruel post might be nudged to reconsider their language or to notice the mob-like tenor of a “trending” topic. The challenges of developing such emotionally charged, high-EQ AI are daunting, but instead of simply weeding out individual posts it might ultimately be more beneficial to change online behavior for the better.
Comments
Post a Comment