摘要
With the explosive growth of data and the rapid rise of artificial intelligence (AI) and automated working processes, humans (either as employees or consumers) inevitably fall into increasingly close collaboration with machines and large-scale information. On the one hand, due to various constraints such as machine-learning algorithms' inherent "black boxes'', human agents lack an efficient way to work with AI or even show resistance. On the other hand, such collaboration could be expensive for individual companies when taking into account the increasing costliness of labor, big data collection, and technical development. Hence, it is crucial as well as a matter of some urgency to explore how humans and machines behave in collaboration mode under different information conditions. Inspired by the information processing literature, which has suggested, by implication, respective pros and cons of humans and machines, we herein introduce the concept of information complexity (in the form of information volume) and identify its roles in determining collaboration performance. Specifically, we cooperate with a large Asian microloan company to conduct a two-stage field experiment in which we tune the treatments by level of information volume, the presence of collaboration, and the availability of machine transparency. We compare approved loans' default rates between human evaluators and machine-learning algorithms. We present at least three interesting findings herein. First, we show that humans, especially experienced ones, might resist alternative information sources but make decisions via a traditional process with small information volumes. Second, machines perform better on larger information scales and better than humans but would unintentionally incur gender biases. Third, in the human-machine collaboration mode, the presence of machine interpretations, when compared with their absence, could reduce humans’ potential resistance to machines’ recommendations. More importantly, the co-existence of large-scale information and machine interpretations can invoke humans’ systematic rethinking, which in turn, shrinks gender gaps and increases prediction accuracy simultaneously. Our experiments and empirical findings provide non-trivial implications that are both theoretical and practical, respectively.