When should people trust an AI's recommendation?
A pre-registered, repeated-decision experiment studying appropriate reliance on AI advice: how AI accuracy, confidence displays, explanations, feedback and task difficulty shape whether people follow AI recommendations when they should — and override them when they shouldn't.
What this is
Two literatures disagree about how people respond to algorithmic advice: one documents automation bias — following AI even when it's wrong — and another documents algorithm aversion — under-using AI even when it's more accurate than the person. Both can be true at once, in different conditions. This project builds the experiment, data pipeline and analysis needed to say which conditions produce which failure, rather than asserting one story.
How it works
Participants complete short, repeated binary decisions — comparing dot counts, forecasting a noisy series — first alone, then after seeing a recommendation from an AI system whose accuracy and confidence display are experimentally controlled. The design and every outcome definition are fixed in the methodology before any human data are collected.
Evidence discipline
No human-subject data exist yet. Every chart on this site carries a label saying whether it is built from synthetic data (an archetype-based participant simulator, used only to validate the experiment and the statistics), pilot data, or real pre-registered data. Synthetic results are never presented as findings about people — see the data page for the provenance system that enforces this.