Dutch authorities are confronting serious flaws in algorithms used for years to assess the risk of reoffending, according to a new report from the Justice and Security Inspectorate. The findings have prompted a response from Jessica Westerik, director of Reclassering Nederland, who addressed the critical assessment in a recent interview.
Westerik described the report’s conclusions as “exceptionally confronting” and stated that the reclassifying agency needs to “catch up” regarding the correct use of algorithms.
“We take this very seriously, and I take responsibility for it,” Westerik said. “I wouldn’t want it any other way. At Reclassering Nederland, we try to teach people every day to take responsibility for their mistakes and try to correct them.”
The reclassifying agency advises judges approximately 48,000 times a year on the likelihood of an individual re-offending – a determination known as recidivism risk. This advice carries significant weight and can determine whether a suspect is released or remains in custody.
Outdated Data from Sweden
The inspection report reveals that the calculation models used in these assessments were inaccurate in roughly one in five cases. One algorithm, known as Oxrec, fails to meet standards for government algorithms in almost every respect, according to the inspection. Crucial formulas within the algorithm were mixed up, drug issues were not properly accounted for, and severe mental health conditions were not included.
the model operated on outdated data from Swedish prisoners. “It’s questionable whether a dataset predicting outcomes for the Swedish population can be applied to the Netherlands,” said Cynthia Liem, an associate professor at Delft University of Technology specializing in the responsible use of AI, and algorithms.
Liem called it “embarrassing” that the model was never properly implemented and that there was no oversight in recent years. She attributed this to a strong tendency for organizations to place excessive trust in computer models. “AI and algorithms develop it tempting to think less, question less, and simply accept what is offered. This is happening in many technological applications.”
This is one of the most painful reports we have faced in recent years.
The reclassifying agency has paused the use of the models and launched a review of its risk assessment tools.
According to the inspection, in the majority of errors, the risk of recidivism was underestimated. This could potentially leave the public vulnerable. An overestimate could negatively impact a suspect or convicted person, potentially resulting in a harsher sentence.
When asked if individuals may have been released too early or sentenced more harshly due to inaccurate assessments, Director Westerik said she could not rule it out. “But I think the chances are very small.”
Discrimination Concerns
“This is one of the most painful reports we have faced in recent years,” said Sven Stevenson of the Autoriteit Persoonsgegevens. “As citizens, you should be able to trust that an algorithm is fair and does not discriminate. Both are not in order with several of these algorithms.”
Adding to the concern, the reclassifying agency was warned in 2020 about discriminatory elements in the Oxrec algorithm. The model factored in postal code and income, which can indirectly lead to ethnic profiling. Westerik said the agency took this signal “very seriously” at the time. “But at that time, there seemed to be sufficient scientific evidence that neighborhood and income were independent factors.”
Yet, the Human Rights Council also stated a year later that it is prohibited to weigh these characteristics without proper justification. And that justification was lacking. “We did not do that sufficiently, that is correct,” Westerik said.
The director acknowledged that her organization lacked sufficient knowledge about algorithms and their maintenance. “We require to improve that, and we will. But I also want to emphasize that reclassification is people’s perform. In all our training materials, it states that the professional judgment of our employees is leading.”
