Two-Layer Learning in Human–AI Collaboration in Uncertain Tasks

24.09.2026 12:00 – 13:00

MANAGEMENT BROWN BAG SEMINAR

ABSTRACT

Artificial intelligence increasingly supports decision making in uncertain environments, yet effective human–AI collaboration requires decision makers to learn not only about the task they face but also about the quality of the AI advising them. We conceptualize this challenge as a two-layer learning problem: outcomes simultaneously provide information about an uncertain environment and about the competence of an opaque AI advisor, making it difficult to disentangle the two sources of uncertainty. We examine this problem in an incentivized multi-armed bandit experiment in which participants repeatedly make decisions under uncertainty. We experimentally vary access to AI advice and whether AI quality remains stable or unexpectedly deteriorates while the underlying environment remains unchanged. Reliable AI substantially accelerates learning and improves performance, particularly among less-skilled decision makers. When AI quality deteriorates, however, participants continue relying on its recommendations and only gradually recalibrate their behavior, producing substantial performance losses. More-skilled decision makers adapt more effectively to declining AI quality. Importantly, similar aggregate levels of AI reliance can therefore reflect fundamentally different underlying learning processes. Our findings conceptualize human–AI collaboration as a dynamic inference problem and show that successful augmentation depends not only on AI capability, but also on humans’ ability to continuously learn about the quality of the AI with which they collaborate.

Lieu

Bâtiment: Uni Mail

Boulevard du Pont-d'Arve 40
1205 Geneva

Room M 3250, 3rd floor

entrée libre