Probation algorithm may be discriminatory, often underestimates chance of reoffending
Dutch probation services are reportedly relying on a computer algorithm, OxRec, that has allegedly never accurately predicted the risk of reoffending and is used about 44,000 times a year to advise judges and prosecutors, according to the national Inspectorate of Justice and Security. The Inspectorate states that the system frequently underestimates the likelihood that offenders will commit new crimes, particularly among drug users and individuals with serious mental health conditions. There is also a risk that the computer-generated probation plans discriminate, the office claimed.
OxRec reportedly uses age, gender, substance use, income, and neighborhood to calculate risk. According to the Inspectorate, every probation recommendation applies OxRec, potentially impacting thousands of cases annually with inaccurate predictions. The Inspectorate recommended that the probation organizations correct the algorithm immediately or stop using it.
"It is important to clarify that the ‘flawed’ predictions identified by the Inspectorate were the result of errors in how the software was implemented within the local system. The original developers of OxRec - including myself - were not involved in this," said Seena Fazel in an email to NL Times.
Flawed predictions are partly caused by the algorithm using the wrong formulas for different groups—applying the model for prisoners to suspects and vice versa—and by relying on 16- to 24-year-old data from Swedish prisoners, which does not reflect the current Dutch offender population, the Inspectorate wrote.
"It is also not correct to state that the algorithm ‘relies' on older Swedish data," Fazel rebutted. "The model was revised and adapted for use in the Dutch prison and probation population, and the results were published in an open-access, peer-reviewed scientific journal."
The three probation organizations—Reclassering Nederland, Stichting Verslavingsreclassering GGZ, and Leger des Heils Jeugdbescherming & Reclassering—decide whether to follow OxRec’s advice.
The inspectorate warned that staff may rely too heavily on the system rather than their judgment. In an internal podcast, employees were reportedly told that their own assessment without OxRec is “as reliable as flipping a coin.”
"We believe that the Inspectorate has overstated the impact of the reported underestimation. Using simulated data for 100,000 individuals, we assessed the potential effect and found the average difference in predicted risk was small, at approximately 1 percent," said Fazel.
"While the risk tool includes mental illness, the Dutch version did not specifically include severe mental illness. However, this is unlikely to have more than a minimal impact, as no more than 3 percent of comparable samples present with such conditions, which require a specialist psychiatric diagnosis."
