Cognitive Challenges in Human–Artificial Intelligence Collaboration: Investigating the Path Toward Productive Delegation

A consensus is beginning to emerge that the next phase of artificial intelligence (AI) induction in business organizations will require humans to work with AI in a variety of work arrangements. This article explores the issues related to human capabilities to work with AI. A key to working in many work arrangements is the ability to delegate work to entities that can do them most efficiently. Modern AI can do a remarkable job of efficient delegation to humans because it knows what it knows well and what it does not. Humans, on the other hand, are poor judges of their metaknowledge and are not good at delegating knowledge work to AI—this might prove to be a big stumbling block to create work environments where humans and AI work together. Humans have often created machines to serve them. The sentiment is perhaps exemplified by Oscar Wilde’s statement that “civilization requires slaves…. Human slavery is wrong, insecure and demoralizing. On mechanical slavery, on the slavery of the machine, the future of the world depends.” However, the time has come when humans might switch roles with machines. Our study highlights capabilities that humans need to effectively work with AI and still be in control rather than just being directed. We study how humans make decisions when they collaborate with an artificial intelligence (AI) in a setting where humans and the AI perform classification tasks. Our experimental results suggest that humans and AI who work together can outperform the AI that outperforms humans when it works on its own. However, the combined performance improves only when the AI delegates work to humans but not when humans delegate work to the AI. The AI’s delegation performance improved even when it delegated to low-performing subjects; by contrast, humans did not delegate well and did not benefit from delegation to the AI. This bad delegation performance cannot be explained with some kind of algorithm aversion. On the contrary, subjects acted rationally in an internally consistent manner by trying to follow a proven delegation strategy and appeared to appreciate the AI support. However, human performance suffered as a result of a lack of metaknowledge—that is, humans were not able to assess their own capabilities correctly, which in turn led to poor delegation decisions. Lacking metaknowledge, in contrast to reluctance to use AI, is an unconscious trait. It fundamentally limits how well human decision makers can collaborate with AI and other algorith