Julia C. Keller, "Death penalty’s justice might not be so blind . Recent work in artificial intelligence shows that prisoners may be executed for arbitrary reasons" (2005)

Is lady justice as blind as she seems?
> (Photo: Kevin ConnorsMorguefile)

Justice may not be so blind, according to recent work in artificial intelligence that raises theological questions by showing that prisoners may be executed for arbitrary reasons.

The U.S. Conference of Catholic Bishops’ announcement of a new campaign against the death penalty and the U.S. Supreme Court’s decision to abolish capital punishment for minors have prompted the reevaluation of the death penalty system. 

An article in the February issue of the journal Artificial Intelligence and Applications highlights serious flaws in a system that determines life or death.

Stamos Karamouzis and Dee Wood Harper of Loyola University in New Orleans developed AI that can accurately predict — with more than 90 percent accuracy — whether prisoners on death row will be executed. The AI, called the Artificial Neural Network, makes predictions without any knowledge about the nature of the prisoners’ capital crimes.

The network, which doesn’t actually have any biological neurons, despite its name, is essentially a sophisticated data spreadsheet running on Karamouzis’ computer. Though it might not be much to look at, the network can “learn” and make predictions based on this information.

“The difference comparing artificial neural networks to traditional programming is that you do not program the machine with what to do step by step,” said Karamouzis, a computer scientist.

Karamouzis trained the artificial neural network to study more than 1,000 cases of prisoners who were on death row between 1973 and 2000. Half of the cases represented offenders who had been executed.

The cases contained only basic demographic information such as birth date, sex, race and the year and state in which a person was sentenced to death row. These variables represented 17 different “neurons,” or computer processors connected like cells in a human brain, emulating human intelligence, said Karamouzis.

“It learns pretty much the way a human being learns by examining all the relationships between all these variables,” said Harper, a professor in Loyola’s department of criminal justice. “And we tested it by giving it an exam, so to speak.”

They tested the neural network with 300 additional capital offense cases, and found that the AI program accurately predicted whether these additional prisoners would be executed 92 percent of the time.

“It has nothing to do with the heinousness of the crime or the aggravating or the mitigating circumstances of the crime,” said Harper.

“It tells me that overall the process is very arbitrary. If we can predict this, then clearly the decision to execute or not execute doesn’t flow from the crime itself,” he said.

Harper did not set out to question the death penalty. “I started casting around to find something to test this neural network,” he said. He settled on capital punishment in part “because of the nationally growing debate regarding the fairness of the [death penalty] system in light of recent revelations about DNA testing.”

Harper said he doesn’t support the death penalty, but he doesn’t address his opinions from a theological perspective. “I’m not a particularly religious person,” he said. “You’ll probably never see me petitioning at Angola penitentiary up here or participating in a candlelight vigil.”

He’s said he’s more concerned about the judicial system’s impartiality. However, the current technology can’t be used to argue in court during a post-conviction appeal. “The science isn’t there yet,” said Harper. For example, a lawyer couldn’t “plug this person into this little black box, and show the jury or the judge in an appeal situation, and say this man isn’t being executed because he committed a heinous crime; he’s being executed because he’s male and he was born in 1955.”

The justice system’s lack of neutrality is raising concern outside Harper’s lab. “People ought to be disturbed by the findings,” said Susan McGraugh, an assistant clinical professor of law at Saint Louis University in Missouri — another state, like Louisiana, which sanctions the death penalty.

McGraugh, who is also a public defender, said that the technology is “interesting if it’s used the right way.”

She said she was worried that people might misuse the technology. “I am concerned that someone might use this kind of research to say, ‘Well then, next, let’s predict future dangerousness of criminals. If we can teach a computer who is executed and who’s not, then let’s figure out who’s going to be dangerous and use that in the capital arena.’”

Saint Louis University, a Jesuit institution, represents a part of a growing subset of Catholic institutions that are concerned about the fairness of the death penalty system.

The U.S. Conference of Catholic Bishops’ opposition to capital punishment, first stated in 1974, was reinforced in a major 1980 policy paper and other pronouncements since then.

“We cannot teach that killing is wrong by killing. We cannot defend life by taking life,” said Cardinal Theodore McCarrick, the Washington archbishop, speaking at a March conference on the Catholic Church’s new campaign against the death penalty.

McGraugh said that no matter what action the 36 states that currently permit capital punishment take, it must be fair from an ethical and legal standpoint.

“We’re concerned with social justice. We’re not pro-criminal defendant or anti-victim,” said McGraugh. “But we want to see that the results are just. That if the death penalty is going to be applied, it is going to be applied fairly across the board.”

Harper and Karamouzis are working to reduce the number of neurons the network needs to still be as accurate as a network with 17 variables.

“That’s where the utility is,” said Karamouzis, “knowing the exact variables that predict the outcome.”

Both Karamouzis and Harper said they believed that a variable like race would be much less influential to the network than the neurons for the state of sentencing and the year of sentencing for the capital offense.

“One of the biggest factors is how long you’ve been on death row,” said Harper, adding that some prisoners have sat in isolation off the green mile for 14 to 16 years.

“You get to 14 years and you’re still there? You’re in trouble, especially if you’re in Texas, Virginia or Florida,” he said.

“I think what state you’re in has a huge amount of influence on what happens,” said McGraugh. “Some states are just more likely to seek death. There’s a hugely disproportionate amount of inmates on death row in Texas.”

The ethical and legal problem the network shows, said McGraugh, is “if you’re more likely to get executed if you’re from Texas than Missouri, then that violates equal protection under the law.”

Though Harper said he wasn’t sure when the data on the revised AI network would be released, he said he was encouraged by these initial steps.

“If you can develop some hard, empirical facts to show that this isn’t fair, then you’re really winning the battle. We’re not wringing our hands and going on. We’re saying, ‘Here are the facts. This is the way it works,’” said Harper. “That goes a lot longer way to making a difference.”

Julia C. Keller is science editor at Science & Theology News.