go straight to jail.
On the other hand, the man denied everything. He had two previous criminal offenses—but they were misdemeanors, from many years ago. He had a job as a mechanic, which he would lose if he went to jail, and he had an ex-wife and a fifteen-year-old son whom he was supporting with that income. So Solomon had to think about that fifteen-year-old, relying on his father’s paycheck. He also surely knew that six-year-olds are not the most reliable of witnesses. So there was no way for Solomon to be sure whether this would all turn out to be a massive misunderstanding or part of some sinister pattern. In other words, the decision about whether to let the man in the guayabera go free—or to hold him in jail until trial—was impossibly difficult. And to help him make the right call, Solomon did what all of us would do in that situation: he looked the man right in the eyes and tried to get a sense of who he really was. So did that help? Or are judges subject to the same puzzle as Neville Chamberlain?
4.
The best answer we have to that question comes from a study conducted by a Harvard economist, three elite computer scientists, and a bail expert from the University of Chicago. The group—and for simplicity’s sake, I’ll refer to it by the economist’s name, Sendhil Mullainathan—decided to use New York City as their testing ground. They gathered up the records of 554,689 defendants brought before arraignment hearings in New York from 2008 to 2013—554,689 defendants in all. Of those, they found that the human judges of New York released just over 400,000.
Mullainathan then built an artificial intelligence system, fed it the same information the prosecutors had given judges in those arraignment cases (the defendant’s age and criminal record), and told the computer to go through those 554,689 cases and make its own list of 400,000 people to release. It was a bake-off: man versus machine. Who made the best decisions? Whose list committed the fewest crimes while out on bail and was most likely to show up for their trial date? The results weren’t even close. The people on the computer’s list were 25 percent less likely to commit a crime while awaiting trial than the 400,000 people released by the judges of New York City. 25 percent! In the bake-off, machine destroyed man.4
To give you just one sense of the mastery of Mullainathan’s machine, it flagged 1 percent of all the defendants as “high risk.” These are the people the computer thought should never be released prior to trial. According to the machine’s calculations, well over half of the people in that high-risk group would commit another crime if let out on bail. When the human judges looked at that same group of bad apples, though, they didn’t identify them as dangerous at all. They released 48.5 percent of them! “Many of the defendants flagged by the algorithm as high risk are treated by the judge as if they were low risk,” Team Mullainathan concluded in a particularly devastating passage. “Performing this exercise suggests that judges are not simply setting a high threshold for detention but are mis-ranking defendants.…The marginal defendants they select to detain are drawn from throughout the entire predicted risk distribution.” Translation: the bail decisions of judges are all over the place.
I think you’ll agree that this is baffling. When judges make their bail decisions, they have access to three sources of information. They have the defendant’s record—his age, previous offenses, what happened the last time he was granted bail, where he lives, where he works. They have the testimony of the district attorney and the defendant’s lawyer: whatever information is communicated in the courtroom. And they have the evidence of their own eyes. What is my feeling about this man before me?
Mullainathan’s computer, on the other hand, couldn’t see the defendant and it couldn’t hear anything that was said in the courtroom. All it had was the defendant’s age and rap sheet. It had a fraction of the information available to the judge—and it did a much better job at making bail decisions.
In my second book, Blink, I told the story of how orchestras made much smarter recruiting decisions once they had prospective hires audition behind a screen. Taking information away from the hiring committee made for better judgments. But that was because the information gleaned from watching someone play is largely irrelevant. If you’re judging whether someone is a