Strategy. Innovation. Brand.

Are Machines Better Than Judges?

No bail for you.

Police make about 10 million arrests every year in the United States. In many cases, a judge must then make a jail or bail decision. Should the person be jailed until the trial or can he or she be released on bail? The judge considers several factors and predicts how the person will behave. There are several relevant outcomes if the person is released:

  1. He or she will show up for trial and will not commit a crime in the interim.
  2. He or she will commit a crime prior to the trial.
  3. He or she will not show up for the trial.

A person in Category 1 should be released. People in Categories 2 and 3 should be jailed. Two possible error types exist:

Type 1 – a person who should be released is jailed.

Type 2 – a person who should be jailed is released.

Jail, bail, and criminal records are public information and researchers can massively aggregate them. Jon Kleinberg, a professor of computer science at Cornell, and his colleagues did exactly that and produced a National Bureau of Economic Research Working Paper earlier this year.

Kleinberg and his colleagues asked an intriguing question: Could a machine-learning algorithm, using the same information available to judges, reach different decisions than the human judges and reduce either Type 1 or Type 2 errors or both?

The simple answer: yes, a machine can do better.

Klein and his colleagues first studied 758,027 defendants arrested in New York City between 2008 and 2013. The researchers developed an algorithm and used it to decide which defendants should be jailed and which should be bailed. There are several different questions here:

  1. Would the algorithm make different decisions than the judges?
  2. Would the decisions provide societal benefits by either:
    1. Reducing crimes committed by people who were erroneously released?
    2. Reducing the number of people held in jails unnecessarily?

The answer to the first question is very clear: the algorithm produced decisions that varied in important ways from those that the judges actually made.

The algorithm also produced significant societal benefits. If we wanted to hold the crime rate the same, we need only have jailed 48.2% of the people who were actually jailed. In other words, 51.8% of those jailed could have been released without committing additional crimes. On the other hand, if we kept the number of people in jail the same – but changed the mix of who was jailed and who was bailed – the algorithm could reduce the number of crimes committed by those on bail by 75.8%.

The researchers replicated the study using nationwide data on 151,461 felons arrested between 1990 and 2009 in 40 urban counties scattered around the country. For this dataset, “… the algorithm could reduce crime by 18.8% holding the release rate constant, or holding the crime rate constant, the algorithm could jail 24.5% fewer people.”

Given the variables examined, the algorithm appears to make better decisions, with better societal outcomes. But what if the judges are acting on other variables as well? What if, for instance, the judges are considering racial information and aiming to reduce racial inequality? The algorithm would not be as attractive if it reduced crime but also exacerbated racial inequality. The researchers studied this possibility and found that the algorithm actually produces better racial equity. Most observers would consider this an additional societal benefit.

Similarly, the judges may have aimed to reduce specific types of crime – like murder or rape – while de-emphasizing less violent crime. Perhaps the algorithm reduces overall crime but increases violent crime. The researchers probed this question and, again, the results were negative. The algorithm did a better job of reducing all crimes, including very violent crimes.

What’s it all mean? For very structured predictions with clearly defined outcomes, an algorithm produced by machine learning can produce decisions that reduce both Type I and Type II errors as compared to decisions made by human judges.

Does this mean that machine algorithms are better than human judges? At this point, all we can say is that algorithms produce better results only when judges make predictions in very bounded circumstances. As the researchers point out, most decisions that judges make do not fit this description. For instance, judges regularly make sentencing decisions, which are far less clear-cut than bail decisions. To date, machine-learning algorithms are not sufficient to improve on these kinds of decisions.

 

(This article is based on NBER Working Paper 23180, “Human Decisions and Machine Predictions”, published in February 2017. The working paper is available here and here. It is copyrighted by its authors, Jon Kleinberg, Himabindu Lakkaraju, Jure Lesovec, Jens Ludwig, and Sendhil Mullainathan. The paper was also published, in somewhat modified form, as “Human Decisions and Machine Predictions” in The Quarterly Journal Of Economics on 26 August 2017. The paper is behind a pay wall but the abstract is available here).

Leave a Reply

Your email address will not be published. Required fields are marked *

My Social Media

YouTube Twitter Facebook LinkedIn

Newsletter Signup
Archives