In sectors as diverse as health care, criminal justice, and finance, algorithms are increasingly used to help make complex decisions that are otherwise troubled by human biases. Imagine criminal justice decisions made without race as a factor or hiring decisions made without gender preference. The upside of AI is clear: human decisionmakers are far from perfect, and algorithms hold great promise for improving the quality of decisions. But disturbing examples of algorithmic bias have come to light. Our own work has shown, for example, that a widely-used algorithm recommended less health care to Black patients despite greater health needs. In this case, a deeply biased algorithm reached massive scale without anyone catching it—not the makers of the algorithm, not the purchasers, not those affected, and not regulators.
To stop algorithmic bias, we first have to define it (brookings.edu)