Can All Humans Benefit from AI Assistance? Algorithm Aversion and Overconfidence
To what extent can AI assistance improve human decision comparing to human or AI alone? It is a crucial question that has profound policy implications. We experimentally study this issue in a setting of loan approval decision based on historical cash loan data. We find that human decision with AI assistance improves upon human decision alone, but still cannot beat AI alone. However, subjects demonstrate algorithm aversion in the form of AI information underweighting, which is significantly related to overconfidence. AI assistance narrows the performance gap among subjects; surprisingly, due to the underweighting bias some originally good-performers suffer from it.
