Algorithmic Decision-Making and Managerial Bias: An Empirical Investigation

Abstract

Algorithms are introduced into managerial decisions such as hiring, promotion and credit partly to
reduce the biases of human decision-makers, yet algorithms are also accused of perpetuating those
biases. Both claims can be true, and which one holds depends on how the algorithm is built and how
it is used. This paper sets out an investigation design that distinguishes five decision regimes and
applies it in a simulation of 6,000 decisions in which a group attribute is unrelated to true outcome
potential but human decision-makers are biased against one group. Human decisions select members
of the disadvantaged group at 57% of the rate of the advantaged group. An algorithm trained on those
human decisions inherits part of the bias (selection ratio 0.92) despite higher accuracy; an algorithm
trained on actual outcomes reaches parity (1.01) with the highest accuracy (82%); removing a group
proxy from its inputs changes little (1.02); and allowing managers to override 30% of the outcometrained
algorithm’s decisions reintroduces bias (0.86). The results locate bias in the labels an
algorithm learns from and in the human discretion that surrounds it, rather than in algorithmic
decision-making as such, and they show that accuracy and fairness were aligned in this setting. The
paper proposes an audit protocol built on the five regimes and discusses implications for the design
of human–algorithm decision processes.