Algorithm Aversion

What it is

Dietvorst and colleagues, writing at the University of Pennsylvania, reported five studies in the Journal of Experimental Psychology: General in 2015. Participants who had seen a forecasting algorithm perform became less confident in it and less likely to tie their incentives to it rather than to an inferior human forecaster. That held even for participants who had just watched the algorithm outperform the human.

Noah Castelo, Maarten Bos and Donald Lehmann found in 2019 that the reluctance depends on the task. Across four online lab studies with more than 1,400 participants and two field studies with more than 56,000, people trusted algorithms less for subjective tasks than for objective ones. Reliance rose when a task was framed as benefiting from quantitative analysis and when the algorithm was made to seem more humanlike in its feelings.

Two other lines of work point the same way. Chiara Longoni, Andrea Bonezzi and Carey Morewedge reported nine studies in 2019 in which people were less willing to use healthcare from an AI provider, would pay less for it and were less sensitive to differences in its performance. They traced this to uniqueness neglect, the worry that a machine cannot take account of what is particular about a person. In 2023 Kobe Millet and colleagues found, in four experiments with 1,708 participants, that the same artwork was rated less creative and less awe-inspiring when labelled as made by AI.

In effect

For a consumer, one visible mistake by an automated tool can weigh more than a long record of accuracy. The reluctance is strongest where a judgement feels subjective or personal, such as medicine, art and taste.

What it does not say

It does not say people always prefer humans. Jennifer Logg and colleagues found in 2019 that lay people followed advice more closely when it was attributed to an algorithm, a pattern they called algorithm appreciation.

The original studies concerned forecasting under experimental incentives, and much of the later work used scenarios and online samples. The Castelo figures were read in a paraphrase on the publisher's page and a university summary, not in the full paper.

---