Khái niệm cốt lõi
Complementarity between humans and AI can lead to superior team performance, but is often not realized in practice. This work establishes a conceptual foundation to understand and develop human-AI complementarity by introducing the notion of complementarity potential and its inherent and collaborative components. The authors demonstrate the value of this conceptualization through two empirical studies that explore information and capability asymmetries as sources of complementarity potential.
Tóm tắt
The paper establishes a conceptual foundation for understanding and developing complementarity in human-AI decision-making. It introduces the notion of complementarity potential, which comprises inherent and collaborative components.
The inherent complementarity potential represents improvements that could be contributed by including superior decisions from the overall less accurate team member. The collaborative complementarity potential captures decision-making synergies that only emerge through human-AI interaction.
The authors demonstrate the value of this conceptualization through two empirical studies:
- Information Asymmetry Study:
- Participants collaborate with an AI model to predict real estate prices.
- Humans receive additional contextual information (house photos) that the AI model does not have access to.
- The results show that the information asymmetry increases the inherent complementarity potential, allowing the human-AI team to achieve complementary team performance.
- Capability Asymmetry Study:
- Participants collaborate with AI models that have different capability levels compared to the human.
- The results demonstrate that capability asymmetry also increases the inherent complementarity potential, enabling the human-AI team to outperform the individual team members.
The studies illustrate that leveraging sources of complementarity potential, such as information and capability asymmetries, constitutes a viable pathway toward effective human-AI collaboration and superior team performance.
Thống kê
The mean absolute error (MAE) of the AI model on the hold-out set is $163,080.