Human-Machine Collaboration: Performance and Responsibility Attribution
COM3 Level 2
SR21, COM3 02-60
Abstract:
The recently emergent paradigm of human-machine collaboration has garnered great interest from practitioners and researchers. Compared with the human-machine substitution paradigm, human-machine collaboration is expected to enhance performance while also addressing concerns about responsibility attribution. However, these assumptions remain largely unexamined, which hence motivates this thesis.
The first essay, “Distraction as a Unique Human Advantage: Comparing the Performance of Human, Machine, and Hybrid Decision-Making in Wicked Problems,” explores the decision-making performance of human, machine, and hybrid agents in wicked problems—situations where objectives lack universal definitions. Using team formation as a context, I underscore that distraction may serve as a unique human advantage. While machines excel in “inside-box” alignment as they can form teams that align with predefined goal functions, specifically skill complementarity and capability similarity, humans excel in “outside-box” alignment as their attention is often diverted to alternative goal functions, such as familiarity and prosociality. The analysis of Multiplayer Online Battle Arena (MOBA) game data reveals that such distraction hinders the performance of human decision-making when complexity is low, but enhances performance when complexity is high.
Beyond performance, ethical considerations are also crucial for human-machine collaboration. The second essay, “Alleviate the Tension in Dual Identity of AI Agents: Responsibility Attribution and Consumer Engagement in (Human-) AI Services” examines how perceived agency and accountability influence consumer engagement by affecting responsibility attribution and concern. AI agents present a unique tension in responsibility attribution due to their dual identity: while AI agents’ perceived agency prompts users to attribute responsibility to them similar to a social actor, their inability to hold legal responsibility exposes a tension between perceived agency (actor identity) and accountability (tool identity). In a stress assessment service scenario featuring AI, human, and large language model (LLM) doctors, I find that this tension is unique to AI agents. Moreover, this concern intensifies when AI agents display higher perceived agency. Additionally, I identify two strategies to mitigate this tension: (1) incorporating a human service agent to create a human-AI joint service and (2) explicitly stating that the service provider assumes full responsibility for the joint service.
In summary, this thesis aims to contribute to the human-machine collaboration literatures by identifying how humans can complement machines. In practice, this thesis reveals not only the immediate impact of machine implementation but also their broader effects on final performance and consumer engagement.