Facets of Human–AI Collaboration: The Importance of Collaboration Type, Invocation Type, and AI Output Quality

Recent advancements in artificial intelligence (AI) have made human–AI collaboration increasingly relevant in organizations across various industries. AI-based agents are able to act autonomously, learn from experience, and often outperform humans when solving problems. Moreover, they can generate content indistinguishable from that of human experts and interact in a conversational manner. The advanced capabilities of AI-based agents make human–AI collaboration a promising way for organizations to increase their productivity and performance. Due to this potential, human–AI collaboration has already established itself in several industries and proven to be effective. In software development, AI-based agents can support developers in writing computer code. In financial services, AI-based agents are utilized to make market predictions, and in healthcare, AI-based agents can enhance diagnosis by analyzing patient data. Notably, human–AI collaborations can take multiple forms that differ across several facets, including how responsibilities are distributed between humans and AI-based agents, how the collaboration is invoked, or the quality of the AI-generated output. However, while human–AI collaboration can increase productivity, improve decision-making, or enhance overall performance, previous research has demonstrated that collaborating with AI-based agents can have significant effects on humans in various forms. It is therefore important to understand how the different facets of human–AI collaborations influence the humans involved, to fully realize the potential of such collaboration. Consequently, this thesis aims to investigate how various facets of human–AI collaboration, specifically the type of collaboration, the type of invocation to collaborate, and the quality of output of the AI-based agent within the collaboration (i.e., quality of AI output), can affect humans involved in or affected by the outcomes of such collaborations. To this end, this thesis aims in its first strand to investigate how the type of human–AI collaboration can affect the perceptions of humans involved in or affected by the outcomes of such collaborations. To do so, the first study of this thesis (Article A) investigates how the two types of collaboration, namely delegation- and ensemble-based human–AI collaboration, can affect employees’ perceptions of fairness and the trustworthiness of the human collaborator when employees are affected by decisions made in such collaborations. The findings of Article A demonstrate that, in the context of our study, employees perceive a decision-making process as less fair when the decision is made in a delegation-based manager–AI collaboration rather than solely by the human manager, leading to a lower perceived trustworthiness of the manager. Notably, the results revealed that this effect does not occur when the decision is made in an ensemble-based manager–AI collaboration instead. Building on these findings, the second study of this thesis (Article B) explicitly aims to examine how ensemble-based human–AI collaborations can affect the perceptions of the human collaborators. In particular, Article B investigates how ensembling a hiring decision made by a human resource manager with that of an AI-based agent to form the final decision, as in an ensemble-based human–AI collaboration, can affect the managers’ process satisfaction. The findings of Article B show that ensembling a human manager’s decision with an AI-based agent’s decision can negatively influence the manager’s satisfaction with the decision-making process. More importantly, the results revealed that a loss of competence-based self-esteem is a crucial explanatory factor for this effect. The second strand of this thesis aims to examine how the type of invocation to collaborate in a human–AI collaboration can affect the human collaborators’ perceptions and intentions. Therefore, the third and fourth studies of this thesis investigate how the two types of invocation to collaborate, AI-invoked versus human-invoked, can affect the human collaborator of a human–AI collaboration in this regard. That said, Article C reveals that, in work settings, AI-based agents offering to collaborate proactively (i.e., AI-invoked collaboration) rather than reactively, in response to a human request (i.e., human-invoked collaboration), can result in lower satisfaction with the AI-based agent. Furthermore, the article identified a higher loss of competence-based self-esteem as a mediator for this effect. However, the results indicate that these effects do not occur in humans with low levels of knowledge about AI (i.e., AI knowledge) but rather increase with rising AI knowledge. The fourth study of this thesis (Article D) continues, indicating that the invocation of a delegation-based human–AI collaboration by an AI-based agent rather than by the human collaborator can threaten their self-view, negatively affecting their intention to collaborate. In particular, the findings reveal that when an AI-based agent invokes a collaboration by offering to take over a task on its own initiative rather than upon the human’s request, it can lead to higher self-threat for the human collaborator and, in turn, reduce their willingness to accept the offer. Furthermore, the results indicate that a high level of perceived human control after delegation can mitigate these effects. The last strand of this thesis aims to examine how the quality of output provided by AI-based agents in human–AI collaboration can affect the human collaborators’ knowledge. In this vein, the fifth study (Article E) analyzes the context of software development with particular focus on how the quality of code produced by AI-based agents during human–AI collaboration affects developers’ knowledge of how to employ programming syntax and underlying conceptual principles to write computer code that addresses programming tasks (i.e., developers’ procedural knowledge). The findings of this study indicate that developers receiving assistance with higher (vs. lower) code quality from AI-based agents exhibit relatively lower procedural knowledge because they experience lower cognitive load (i.e., the amount of mental resources used in working memory to process information). Overall, the five articles of this thesis advance research on human–AI collaboration by conceptualizing and empirically examining the type of collaboration, the type of invocation, and the quality of AI output as three essential facets that can shape the perceptions, intentions, and knowledge of humans involved in or affected by the collaboration. In doing so, this thesis emphasizes the multidimensional nature of human–AI collaboration and shows how these facets can influence the humans involved, highlighting the need for future research to move beyond general or unidimensional perspectives to better understand and design human–AI collaborations.