Algorithms predicting parole outcomes

The US has the highest incarceration rate in the world, which results in overcrowded prisons and all the additional violence that implies.

Funnelling felons back onto the street through granting parole is thus a critical safety mechanism and management tool – but assessing which inmates will likely not reoffend when granted liberty is a difficult and troubling task.

For some years now, the people responsible for calculating the chances of someone reoffending have been assisted in their decision-making by computational frameworks known as risk-assessment instruments (RAIs).

The validity of these algorithms was thrown into question in 2018 after a major study tested their predictive power against that of untrained humans. The machines and the people were given brief information on 400 inmates, including sex, age, current charge and prior convictions, and asked to make a determination.

Both cohorts made the correct call in 65% of cases, which was pretty perceptive on the part of the untrained humans, but rather ordinary for the algorithms, given what was at stake.

Now a new study, led by Sharad Goel, a computational social scientist at Stanford University, US, has repeated and extended the earlier research, and finds in favour of the software.

In the first phase of the research, Goel and colleagues replicated the previous work, and came up with similar results. They then repeated the exercise with several additional variables in play – a situation, they suggest, that much better resembles real-world conditions.

With the extra information, the algorithms performed much better, correctly predicting recidivism in 90% of cases. The humans got it right only 60% of the time.

“Risk assessment has long been a part of decision-making in the criminal justice system,” says co-author Jennifer Skeem.

“Although recent debate has raised important questions about algorithm-based tools, our research shows that in contexts resembling real criminal justice settings, risk assessments are often more accurate than human judgment in predicting recidivism.

That’s consistent with a long line of research comparing humans to statistical tools.”

In their paper, published in the journal Science Advances, the researchers say the more accurate RAI results will be helpful in the management of the over-burdened US penal system.

The algorithm will be useful not only in helping to decide which inmates can be safely released into the community but will also assist in allocating prisoners to low or high security facilities.