How Ensembling AI and Public Managers Improves Decision-Making

TY - UNPB

T1 - How Ensembling AI and Public Managers Improves Decision-Making

AU - Keppeler, Florian

AU - Borchert, Jana

AU - Pedersen, Mogens Jin

AU - Nielsen, Vibeke Lehmann

PY - 2025/4/18

Y1 - 2025/4/18

N2 - Artificial Intelligence applications (AI) transform public sector decision-making. However, most research conceptualizes AI as a form of specialized decision support tool. In contrast, this study presents a different form of human-AI collaboration, the concept of human-AI ensembles, where public managers and AI tackle the same decision tasks, rather than specializing in certain subtasks. This is particularly relevant for many public sector decisions, where neither human nor AI predictions have a clear advantage over the other. We illustrate this within the context of public hiring, focusing on two key areas: (a) the potential of ensembling human and AI to reduce biases and (b) the willingness of public managers to implement ensembling. Study 1 uses data from the assessment of profiles of real-life job candidates (n = 695) at the intersection of gender and ethnicity by public managers compared to AI. The exploratory OLS regression results illustrate how ensembled decision-making may alleviate ethnic biases. The OLS regression results of study 2, a pre-registered survey experiment, show that public managers (n = 538 with 4 observations each) put equal weight on AI advice and human advice, and, when reminded of the unlawfulness of hiring discrimination, may even prioritize AI over human advice.

AB - Artificial Intelligence applications (AI) transform public sector decision-making. However, most research conceptualizes AI as a form of specialized decision support tool. In contrast, this study presents a different form of human-AI collaboration, the concept of human-AI ensembles, where public managers and AI tackle the same decision tasks, rather than specializing in certain subtasks. This is particularly relevant for many public sector decisions, where neither human nor AI predictions have a clear advantage over the other. We illustrate this within the context of public hiring, focusing on two key areas: (a) the potential of ensembling human and AI to reduce biases and (b) the willingness of public managers to implement ensembling. Study 1 uses data from the assessment of profiles of real-life job candidates (n = 695) at the intersection of gender and ethnicity by public managers compared to AI. The exploratory OLS regression results illustrate how ensembled decision-making may alleviate ethnic biases. The OLS regression results of study 2, a pre-registered survey experiment, show that public managers (n = 538 with 4 observations each) put equal weight on AI advice and human advice, and, when reminded of the unlawfulness of hiring discrimination, may even prioritize AI over human advice.

KW - public sector

KW - Human-AI collaboration

KW - decision-making

KW - experiment

KW - bias

U2 - 10.31219/osf.io/2yf6r_v3

DO - 10.31219/osf.io/2yf6r_v3

M3 - Preprint

T3 - Journal of Public Administration Research and Theory

BT - How Ensembling AI and Public Managers Improves Decision-Making

ER -