The participants were told to play an MBA admissions officer. For each real applicant, they saw the kind of information an application holds, such as GMAT scores, interview and essay ratings and work experience, and had to forecast how well that person had actually done in the program, measured by grades, how classmates rated them, the prestige of their first employer and their promotions and raises two years after graduating.
There was also a statistical model, built from hundreds of past students using the same information, making the same forecasts.
In the first stage, participants worked through 15 applicants. Depending on the group they were randomly put in, they saw feedback on their own forecasts, on the model’s, on both, or on neither. Then came the decision that mattered. They would make 10 more forecasts and earn $1 for each one within 5 percentiles of the truth, and they had to choose whose forecasts would count: their own or the model’s.
The model was better. Across five studies by the Wharton researchers Berkeley Dietvorst, Joseph Simmons and Cade Massey, published in 2015, people’s forecasts of MBA students had 15% to 29% more error than the model’s. In a second task, ranking US states by airline passengers, people’s error was 90% to 97% larger. In every condition of every study, participants would have earned more by choosing the model.
But watching the model work changed how people chose. Those who had seen its forecasts and their mistakes were less likely to bet on it than those who hadn’t. Watching their own forecasts miss did not significantly reduce their willingness to bet on themselves, though their misses were bigger.
The group that saw both sides side by side was the sharpest test. Across the studies, 610 of the 741 people in it saw the model beat the human forecaster in that first stage; a minority saw the human win. They were still among the least likely to choose the model. In the first study, confidence in the model’s forecasts, on a scale of 1 to 5, averaged 3.04 in the control group, which saw no forecast feedback at all, and 2.49 among those who had seen only the model’s forecasts.
The researchers named the pattern algorithm aversion. Their explanation was that people “more quickly lose confidence in algorithmic than human forecasters after seeing them make the same mistake.”
Three years later, the same team found one fix, in a task forecasting students’ scores on a standardized math test. In one study, people who could adjust the model’s forecasts by up to 10, 5 or 2 percentiles chose to use it 70% of the time, taken together, against 47% for those who had to take its numbers as they were. The preference barely changed with how much adjustment was allowed. And because more of them used the model, the groups allowed to adjust made smaller errors overall than the group that wasn’t.