Algorithm aversion is the tendency to abandon an algorithm after seeing it make a mistake, even when it demonstrably outperforms the human alternative. It is the mirror image of automation bias, it was identified by Berkeley Dietvorst and colleagues in 2015, and the two are almost never discussed together despite describing opposite failures of the same relationship.
Definition#
Algorithm aversion: the disproportionate loss of confidence in an algorithmic forecaster after observing it err, relative to the loss of confidence in a human forecaster making the same error, resulting in the rejection of a system that performs better.
The finding#
Dietvorst, Simmons and Massey ran a series of experiments in which participants chose between their own forecasts and an algorithm's. Those who saw the algorithm perform, including its mistakes, abandoned it more readily than those who never saw it work, even when they had also seen that it outperformed them.
The asymmetry is the point. People forgive human error far more readily than machine error. A colleague who gets something wrong is having an off day. A system that gets something wrong is broken. The standard applied is different. It is applied to the system that was performing better.
Later work by the same group found that giving people even slight ability to modify the algorithm's output substantially restored their willingness to use it, which suggests the aversion is partly about control rather than about accuracy.
The apparent contradiction#
Logg and colleagues found in 2019 that people frequently weight algorithmic advice more heavily than human advice, with domain experts the notable exception. So which is it?
The reconciliation is about what has been seen. Before observing failure, people often over-trust algorithms, which is automation bias. After observing failure, they under-trust them, which is aversion. Expertise moves the starting point but not the shape of the response.
Which means an organisation can be suffering both failures simultaneously, in different teams, with the same system. That is what you would expect from a population differing in exposure and expertise, and a strong argument against uniform policies. No paradox is involved.
Most of this work uses forecasting tasks#
Most of this work uses forecasting tasks with clear, quickly revealed outcomes. Much professional work has neither: the outcome arrives late, ambiguously, or not at all, and people rarely learn whether their override was correct. Whether aversion behaves the same way without that feedback is untested.
There is also a definitional problem worth naming. Rejecting an algorithm that is right on average but wrong in a specific case may be entirely correct if you hold context the system does not. Not all aversion is a bias, and the literature is better at identifying the pattern than at telling you when it is a mistake.
Two failures of the same calibration#
The reason these two concepts belong on the same page is that both are failures of calibration rather than of trust. The useful capability is knowing, for your own domain, where the system is reliable and where it is not, and holding that map steadily enough that a single visible error does not overturn it and a run of successes does not inflate it.
That is the jagged frontier stated as a psychological problem. It is also why a stated override rule matters: written in advance, it protects against both failures, because it does not move when you have just been impressed or just been embarrassed.
Related SuperSkills research#
On the opposite failure, automation bias and automation complacency. On the rule that guards both, when should I override AI. On the boundary, the jagged frontier. On decision design, human and AI decision making.
Key sources
- Dietvorst, B. J., Simmons, J. P. and Massey, C. (2015). Algorithm aversion: people erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1).
- Logg, J. M. et al. (2019). Algorithm appreciation. Organizational Behavior and Human Decision Processes, 151.
- Parasuraman, R. and Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors, 39(2).
About this definition#
Algorithm aversion is Dietvorst, Simmons and Massey's term and is not a SuperSkills coinage. Rahim Hirji is the author of SuperSkills (Kogan Page, 2026).
About this research#
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.
Cite this
Hirji, R. (2026). What is algorithm aversion? The SuperSkills Intelligence Company. Last reviewed 26 August 2026. thesuperskills.com/research/what-is-algorithm-aversion