eople often prefer a human judgment to a formula even when the formula has the better record. That preference has a name, algorithm aversion, and a long prehistory in clinical psychology.

the old finding

In 1954 Paul Meehl compared clinicians' predictions with simple statistical rules in the studies then available, and found the rules as good or better in nearly every comparison (Meehl, 1954) [paraphrase, not a quotation]. Dawes (1979) added the uncomfortable twist: even linear models with crude equal weights tended to beat expert judgment, because the experts were inconsistent [paraphrase, not a quotation]. Grove and colleagues (2000) pooled the literature in a meta-analysis and found mechanical prediction modestly more accurate on average, and clearly worse in only a small minority of comparisons [paraphrase, not a quotation]. This is the clinical versus statistical debate.

the aversion

Dietvorst, Simmons & Massey (2015) let people watch a forecasting model and a human err, then choose whom to rely on. Having seen the model make mistakes, people turned away from it more than from the human, even when the model was better overall [paraphrase, not a quotation]. In a follow-up, the same authors (2018) found that letting people adjust the model's forecast, even by a little, made them much more willing to use it [paraphrase, not a quotation].

The aversion depends on the task. Castelo, Bos & Lehmann (2019) found people trusted algorithms less for tasks that seem subjective, and that making a task seem more objective raised reliance [paraphrase, not a quotation]. And it is not universal: Logg, Minson & Moore (2019) found many lay participants gave more weight to advice labelled as coming from an algorithm, a pattern they called algorithm appreciation, while experts discounted it [paraphrase, not a quotation].

Two errors, one dial. Aversion is under-reliance; automation bias is over-reliance. Both are failures of calibration of trust, and the same person can show either, depending on the task and on what they saw last.

live debates

Aversion or appreciation? Dietvorst et al. (2015) found people abandoned an algorithm after seeing it err; Logg et al. (2019) found people weighted algorithmic advice more than human advice. Two systematic reviews try to reconcile them. Burton, Stein & Jensen (2020) organise the causes under expectations, decision autonomy, incentives, cognitive compatibility and divergent rationalities; Mahmud et al. (2022) sort influencing factors into algorithm, individual, task and higher-level ones [paraphrase, not a quotation]. Domain matters: Longoni et al. (2019) traced resistance to medical AI to uniqueness neglect, the worry that a machine cannot account for one's particular case, and Castelo et al. (2019) found aversion stronger for tasks seen as subjective. Prahl & Van Swol (2017) found advice labelled as automated was discounted more steeply than human advice after a bad recommendation.

methodological critiques

What is being measured. "Aversion" has meant different things: choosing a human over an algorithm, weighting algorithmic advice less, or abandoning it after errors. Most studies use vignettes or judge–advisor tasks with modest stakes, and "aversion" is often scored against a benchmark in which the algorithm is known to be better, a fact participants do not have. Under-use is only an error relative to a performance gap the participant could have known about. The modifiable-algorithm result (Dietvorst, Simmons & Massey, 2018) suggests the effect is partly about control, not accuracy.

still open

Whether reliance can be made to track actual accuracy case by case, and not just on average; and whether the right fix is a better machine, a better explanation, or a better role for the person, which is the subject of complementarity.

in the tour: the forking corridor has a placard called "Meehl: clinical and statistical" ···

Reading list

1. Meehl (1954). Clinical versus Statistical Prediction: A Theoretical Analysis and a Review of the Evidence. Minneapolis: University of Minnesota Press.
2. Dawes (1979). The robust beauty of improper linear models in decision making. American Psychologist. doi:10.1037/0003-066x.34.7.571
3. Grove et al. (2000). Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment. doi:10.1037/1040-3590.12.1.19
4. Dietvorst, Simmons & Massey (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General. doi:10.1037/xge0000033
5. Dietvorst, Simmons & Massey (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science. doi:10.1287/mnsc.2016.2643
6. Castelo, Bos & Lehmann (2019). Task-Dependent Algorithm Aversion. Journal of Marketing Research. doi:10.1177/0022243719851788
7. Logg, Minson & Moore (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes. doi:10.1016/j.obhdp.2018.12.005
8. Burton, Stein & Jensen (2020). A systematic review of algorithm aversion in augmented decision making. Journal of Behavioral Decision Making. doi:10.1002/bdm.2155
9. Mahmud et al. (2022). What influences algorithmic decision-making? A systematic literature review on algorithm aversion. Technological Forecasting and Social Change. doi:10.1016/j.techfore.2021.121390
10. Longoni, Bonezzi & Morewedge (2019). Resistance to Medical Artificial Intelligence. Journal of Consumer Research. doi:10.1093/jcr/ucz013
11. Prahl & Van Swol (2017). Understanding algorithm aversion: When is advice from automation discounted?. Journal of Forecasting. doi:10.1002/for.2464

see also

	automation bias — the opposite error
	complementarity — when the team beats either alone
	Paul Meehl — the clinician who counted
	annotated: human–AI interaction — the reading shelf