Cognitive challenges in human-AI collaboration: Investigating the path towards productive delegation

We study how humans make decisions when they collaborate with an artificial intelligence (AI): each \ninstance of a classification task could be classified by themselves or by the AI. Experimental results suggest \nthat humans and AI who work together can outperform the superior AI when it works alone. However, \nthis only occurred when the AI delegated work to humans, not when humans delegated work to the AI. \nThe AI profited, even from working with low-performing subjects, but humans did not delegate well. This \nbad delegation performance cannot be explained with algorithm aversion. On the contrary, subjects tried to \nfollow a provided delegation strategy diligently and appeared to appreciate the AI support. However, human \nresults suffered due to a lack of metaknowledge. They were not able to assess their own capabilities correctly, \nwhich in turn leads to poor delegation decisions. In contrast to reluctance to use AI, lacking metaknowledge \nis an unconscious tra