The participants were playing MBA admissions officers. For each real applicant they saw GMAT scores, interview and essay ratings and work experience, and forecast how well that person had gone on to do. A statistical model, built from hundreds of past students, made the same forecasts.
Some of them were shown, for 15 applicants in a row, their own forecast, the model’s forecast and the true result.
Then they made a choice with money on it. For the next 10 forecasts, they would earn $1 for each one that landed close to the truth, and they had to decide whose forecasts would count, theirs or the model’s.
Forgiving people, not formulas
Seeing both sides didn’t help the model. The study, published in 2015 by Berkeley Dietvorst, Joseph Simmons and Cade Massey of the Wharton School, ran five experiments, including a second task, ranking US states by airline passengers, and one study in which people chose between the model and another participant. Combining them, 610 of the 741 people who saw both the model’s and a human’s results watched the model beat the human forecaster. That group was still among the least likely to choose the model.
On average the model beat people in every study, by a wide margin on the airline task, and in every study choosing the model paid a higher average bonus. But people who had seen the model’s forecasts and their misses were less likely to bet on it than people who hadn’t seen it work at all. Seeing their own misses, which were bigger, didn’t significantly shake their faith in themselves.
The authors called it algorithm aversion. People, they wrote, “more quickly lose confidence in algorithmic than human forecasters after seeing them make the same mistake.”
That cuts against a long record, which the Wharton authors summarize. In 1954 the psychologist Paul Meehl reviewed 20 forecasting studies, in areas from academic performance to parole violations, and showed that the formulas outperformed human judges. A 2000 analysis of 136 studies found algorithms beat human forecasters by about 10% on average.
Not always, not everyone
The aversion isn’t universal. In 2018 Jennifer Logg, Julia Minson and Don Moore found that ordinary people often lean more on advice when told it comes from an algorithm than from a person, and called that algorithm appreciation.
But it faded in two situations that matter at work. It weakened when people were choosing between an algorithm’s estimate and their own, rather than another person’s. And it weakened with expertise: when 70 US national security professionals, used to forecasting world events, got the same advice as a group of online participants, they relied on the algorithm less, “which hurt their accuracy.”
These findings come from different experiments, and those experiments didn’t test how they combine. But each points to a moment worth watching: right after a tool makes a visible mistake, when people are choosing between its answer and their own, and when the people involved are experienced at the job.
A little control
Three years after their first paper, the Wharton team tested a remedy, with people forecasting students’ math test scores. In one study, taken together, 70% of the participants allowed to adjust the model’s forecasts by up to 10, 5 or 2 percentiles chose to use it, compared with 47% of those who had to take its numbers as they were. How much adjustment was allowed barely mattered. And because more of them used the model, the groups allowed to adjust made smaller errors overall than the group that wasn’t. Those were statistical forecasts with a number to nudge. Whether open-ended AI output is as easy to adjust, and whether people’s edits make it better, has to be tested on its own.
The December miss
Imagine a sales director whose team gets a demand-forecasting tool. For ten months it beats the planners’ spreadsheets. In December it badly underestimates a holiday rush, the warehouse runs short, and by January the planners are back to their spreadsheets. Nobody compares the two over the year. Everyone remembers December.
The habits that help are simple. Keep score over time, on the same cases, for the tool and for the people, so one vivid miss is weighed against the whole record. Let people adjust the tool’s output within set limits, and log both the original and the adjusted number, so you learn whose changes help. And when the tool does fail, ask whether the miss shows a real blind spot, like a kind of case it never handles, or just the ordinary error that every forecaster, human or machine, will make.
The same choice now comes up with AI assistants every day. The trust calibration page covers the opposite failure, trusting a tool past what it has earned. Algorithm aversion is the other edge: 610 of those 741 people watched the model win, and their group was still among the least likely to bet on it.