Being able to adjust the algorithm made it more popular, but slightly less accurate
When predicting students’ math performance, human oversight increased algorithm choice by 7 percentage points and slightly increased error.
The key idea
People preferred the algorithm even though it was about as accurate as a human estimate. Being able to adjust it increased preference and reduced accuracy.
The study, in plain language
What happens when someone can modify an automated recommendation before using it? In an experiment, participants predicted secondary students’ performance on a math test and compared algorithmic recommendations with human estimates.
Many people preferred the algorithm’s recommendation even though its average accuracy was similar to that of humans. With the option to review and adjust it, preference was even greater, but average accuracy fell slightly.
People intervened less on recommendations with larger errors and, when they did, corrected less of those errors. Having the final say did not guarantee a better prediction in this task.
What they found
- Many people chose the algorithm’s prediction, although human predictions had similar average accuracy.
- With the option to review and adjust, preference was greater, but final predictions were slightly less accurate than when delegating.
- Participants intervened less on large errors and corrected less of them when they did intervene.
Something to think about
If you have the final say on an automated recommendation, ask yourself what specific evidence would justify changing it.
How they studied it
An online experiment with 292 participants who predicted secondary students’ math performance. One group delegated the decision, and another could review recommendations from an algorithm or people, which had similar average accuracy. This was a specific school-performance forecasting task, not a test of other kinds of decisions.
What the study cannot show
- A limited experimental task, young and mostly university-educated participants, recommendations carefully matched in quality, and an educational forecasting domain. The authors note that results could differ with expert supervisors, other available data, or other decisions.
- That human oversight always reduces accuracy, that every algorithm outperforms professionals, or that results transfer directly to employment, legal, medical, or credit decisions.
The original study
Putting a human in the loop: Increasing uptake, but decreasing accuracy of automated decision-making
Source under CC BY 4.0. English adaptation and explanation by PaperFlow, with AI assistance. The text has been reorganized for a general audience; it is not the complete paper and does not imply endorsement by its authors.
One study alone does not represent a field's consensus. Educational content does not replace professional advice. Report a possible error.
