A program in Virginia designed to algorithmically assign risk scores to those convicted of crimes and sentence them according to those scores does not work. Instead, the robot judge—which I fully understand is technically neither a robot nor a judge—allows for harsher sentences on young or black defendants and lighter sentences for sex offenders.
The algorithm was adopted in Virginia in 2002 in order to lower prison populations by predicting the chance of recidivism so that judges could focus on rehabilitation for those with less likelihood of re-offending. Defendants were automatically scored based on offense type, age, prior convictions, employment, and marital status. Soon, 28 other states followed suit. However, a recent paper authored by Megan Stevenson from George Mason University and Jennifer Doleac of Texas A&M found that not only does the algorithm score unfairly, but much of the time judges, who are given no risk assessment training, are either ignoring the scores or using them selectively.
One of the biggest issues with the system is how lightly it sentences sex offenders. Despite the fact that the algorithm was designed to approve more jail time for sex offenses, the effect of the cumulative risk score seems to be that the algorithm suggests 24 percent shorter sentences. Sex offenders were also sentenced to prison time five percent less when judges based decisions on the points system.
And though the algorithm could not officially use race as a scoring factor, black defendants were four percent more likely to be sentenced to prison time after the algorithm was implemented and their sentences were 17 percent longer when compared to those of their white counterparts.
“This is partially explained by the fact that black defendants have higher risk scores,” the study says, “and partially because black defendants are sentenced more harshly than white defendants with the same risk score.”
Offenders younger than 30 were also scored unfairly, with the automatic addition of 13 points to their risk assessment score, according to the study. This inflated score means that the programs recommend a longer sentence for younger defendants, which seems contradictory to the stated goal of the program, which was focusing on rehabilitation. Young people were four percent more likely to be incarcerated by judges who considered their risk score and received 12 percent longer sentences than peers.
Part of the reason for flaws in the system, according to some experts, is the fact that judges receive little training in risk assessment and are not required to consider them, meaning that they can ignore the assessment when it’s convenient while using it when they need justification for their sentencing:
“You can have a lot of scenarios where the AI algorithms end up being a rationalization for what the judge wants to do,” University of Maryland professor Frank Pasquale told the Washington Post. Across Virginia, judges who reported confusion around risk assessment scores unsurprisingly also reported using them inconsistently, with one judge even comparing the algorithm to consulting a psychic.