Algorithms have algorithm aversion

Abstract

Purpose This study investigates the phenomenon of algorithm aversion in AI systems. We examine whether AI agents exhibit preferences or aversions to algorithm-generated advice and explores how this aligns with or diverges from human behaviour in similar decision-making contexts. Design/methodology/approach The research replicates three seminal studies on algorithm aversion using GPT models as participants. Scenarios span forecasting tasks, advice-weighting experiments and healthcare recommendations. Findings GPT models exhibit algorithm aversion, preferring human advice over algorithmic inputs. However, the mechanisms driving this behaviour differ from humans. While humans resist algorithmic advice due to biases like uniqueness neglect, GPT models show a broader aversion rooted in performance comparisons. Model version and temperature also influence these preferences. Originality/value This study extends the understanding of algorithm aversion beyond human contexts to AI systems. By systematically comparing human and GPT behaviours, it highlights differences in how aversion manifests, providing insights for designing AI systems that integrate human and algorithmic inputs effectively.

Ähnliche Arbeiten