Noise: A Flaw in Human Judgment

The authors’ research shows that while some groups think that noise-reduction strategies contribute to a fairer society in ensuring that fewer irrelevant factors inform decisions, others think that noise-reduction impedes justice. This is nowhere more evident than in the debate on the federal sentencing guidelines implemented in 1987, which aimed to reduce unwanted variation in the sentencing of similar cases. Judge Marvin Frankel and his disciples argued that a system in which sentences reflected the predilections of particular judges rather than the crimes committed was inequitable. In such a system, all types of noise—level, pattern and occasion—prevailed, as the outcome of a case and a defendant’s fate were determined by unrelated information. However, the guidelines that aimed to counter such noise proved unpopular with judges who felt that they were prevented from exercising their specialized skills. Stith and Cabanes further argued that true justice could not be served when the particulars of each case were not being considered. They felt that the sentencing guidelines did not allow for fair consideration of a defendant’s background, and such blanket approaches contributed to dehumanizing defendants and reducing their dignity. While the authors acknowledge Stith and Cabanes’ view, they point to disparities in sentencing, since the 2005 demotion of the guidelines from mandatory to advisory, to argue that a noisier system is more unfair than one in which every particular of a defendant’s background is considered. Whereas Stith and Cabanes believe in the nobility and distinction of human judgment, the authors argue that human judgment is deeply flawed and needs checks and balances to prevail.

While the authors contend that algorithms can often help reduce noise because they cannot be influenced by the same predilections as humans, they acknowledge that algorithms can sometimes promote biases that favor traditionally dominant groups. For example, when discriminating against people who live in a zip code associated with a particular ethnicity, the algorithm can reinforce the norms of a backwards, regressive society. While the authors agree that prioritizing noise-reduction over the opportunities of traditionally marginalized groups is wrong, they argue for adapting the algorithms to reflect shifting societal attitudes by making them more complex. They also point out that “as humans, we are keenly aware that we make mistakes, but that is a privilege we are not prepared to share. We expect machines to be perfect. If this expectation is violated, we discard them” (160). Instead of having an all-or-nothing attitude where technology is concerned, the authors advocate for the use algorithms as a complement to structured human judgment. Simply by presenting data in a new way, algorithms can suggest a broader range of candidates for a job position than a human, who may succumb to the distraction of different types of noise. Overall, the authors demonstrate that devices, whether technological or human, that slow down and complicate conclusion bias help humans resist making repeated erroneous judgments that disadvantage certain demographics. While they acknowledge that no system will ever reach perfect judgment, the process of questioning and modifying decision-making practices will promote a greater level of fairness over time.