From Competition to Complementarity: Foundations and Evidence for Effective Human-AI Collaboration

The rapid development of artificial intelligence (AI) and its growing ability to autonomously perform a wide range of tasks raises questions about the future role of humans in various domains. In this thesis, we argue that despite the potential ability of AI to automate tasks, there is complementarity between humans and AI that can be harnessed to achieve complementary team performance (CTP), i.e., a level of performance that neither can achieve individually. Therefore, we propose that in the presence of complementarity, humans and AI should work together, a situation commonly referred to as human-AI collaboration. However, existing empirical work shows that the realization of CTP seems to be inconsistent and elusive. Some studies confirm its achievement, while others contradict this finding, essentially weakening human-AI collaboration as an alternative to automation. Therefore, this thesis investigates the challenges of achieving CTP and aims to provide guidance on how to obtain it.

Since the current overview of CTP is rather anecdotal, we begin the thesis by conducting a structured literature review and a meta-analysis to provide a comprehensive and statistical perspective on the current state of empirical studies in the field of human-AI collaboration. ... mehrOur results show that human-AI collaboration can outperform human individuals but often does not exceed the AI’s individual performance. These findings underscore the need for a deeper understanding of the factors that influence CTP.

To address the identified shortcomings, we derive a theoretical foundation of the impact factors of CTP, which we refer to as the human-AI complementarity concept.
Our concept, anchored in existing justificatory knowledge from the information systems field, provides a formalization, highlights possible sources of complementarity potential, and introduces a classification of mechanisms to fuse human and AI decisions during collaboration. Based on this concept, the remainder of the thesis focuses on how existing complementarity potential can be harnessed.

We first explore the process of harnessing complementarity potential in AI-assisted decision-making, the most common form of human-AI collaboration. In this setting, a human receives advice from an AI agent and then decides whether to accept or adjust it. To achieve CTP in this form of collaboration, it is critical that humans rely appropriately on the advice provided by AI. This area of research is still in its infancy and lacks a solid foundation and established measures. To address this research gap, we develop a new measure and a research model. In addition, we conduct two empirical studies to validate and test both the measure and the model. We then shift our focus beyond AI-assisted decision-making to show how complementarity can be harnessed in other forms of human-AI collaboration. Finally, we address the challenge of preserving complementarity over time by conceptualizing a new form of human-AI collaboration-intelligent decision assistance.

This thesis enriches the research on human-AI collaboration by conducting a comprehensive analysis of the current state of empirical studies, introducing new theoretical concepts, developing a measure and research models, and addressing critical challenges. From a practical standpoint, we enable organizations to maximize AI’s utility and yield higher return on investment by offering clear guidance to consistently achieve CTP. Finally, we contribute to the ubiquitous discourse on the future of work, which predominantly focuses on either full automation or pure human work, by providing foundations and evidence for effective human-AI collaboration.