People abandon a more accurate algorithm after seeing it err, and use it far more readily when they are allowed to adjust its output, even slightly.
Why it matters when the plan changes
Evidence that formal methods predict better than unaided judgement has existed for decades, and many organisations still prefer the person. Berkeley Dietvorst, Joseph Simmons and Cade Massey showed why in 5 studies published in 2015: participants who watched an algorithm perform became less confident in it and less likely to choose it over an inferior human forecaster, even when they had seen it outperform the human.
The tension is that some scepticism is healthy, because a model can be wrong in ways its users are right to notice. The same authors found in 2018 that people were considerably more likely to use an imperfect algorithm when they could modify its forecasts, even when the modifications allowed were severely restricted, and that those who could adjust it performed better as a result. Control, even slight, changed how people used the model.
In practice
A hiring panel is given a structured scoring model for shortlisting. In its first month the model ranks highly a candidate who later withdraws, and the panel starts ignoring it. A redesign lets panellists move any candidate up or down one place with a written reason. Use of the model recovers, and the reasons become evidence about what it misses.
Evidence
In 5 studies, people who saw an algorithm err lost confidence in it faster than in a human forecaster making the same mistake.
Berkeley J. Dietvorst, Joseph P. Simmons and Cade Massey, Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err, Journal of Experimental Psychology: General (2015)People were considerably more likely to use an imperfect algorithm when they could modify its forecasts, even slightly.
Berkeley J. Dietvorst, Joseph P. Simmons and Cade Massey, Overcoming Algorithm Aversion, Management Science (2018)Algorithm aversion describes a tendency to reject advice or recommendations from an algorithm in favour of a human.
Algorithm aversion, Wikipedia (2026)
What it cannot tell you
Algorithm aversion describes behaviour in forecasting tasks, largely from experiments, and is not universal: in some settings people over-rely on automated advice instead. It explains resistance to a better model; it does not show that any particular model is better, which has to be established separately.
Questions
Berkeley Dietvorst, Joseph Simmons and Cade Massey, in a 2015 paper in the Journal of Experimental Psychology: General. Across 5 studies they showed people avoiding algorithmic forecasters after seeing them err, even when the algorithm had outperformed the human alternative they chose instead.
Because people forgive human error more readily than machine error. In the 2015 experiments, participants lost confidence in an algorithm more quickly than in a human forecaster after seeing both make the same mistake, so the algorithm was punished for errors a person would have been excused.
By giving people some control. In 2018 the same authors found participants were considerably more likely to use an imperfect algorithm when they could modify its forecasts, even when the modifications allowed were severely restricted, and those who could adjust it performed better as a result.
Yes. People can also defer too much to automated advice, especially under time pressure or when a system seems authoritative, a pattern usually called automation bias. The EU AI Act names automation bias among the risks that people overseeing high-risk systems must remain aware of.
That the way a model's output is presented matters as much as its accuracy. A structured score that experts can adjust within limits, with reasons recorded, is more likely to be used than one they must accept or reject whole, and the recorded reasons show where the model falls short.