The conclusions of the Justice and Security Inspectorate are clear. Algorithms that the probation service has used for years to calculate the risk of someone committing a criminal offense again are full of errors. Jessica Westerik, director at the Netherlands Probation Service, responds News hour on the damning report.
She calls the conclusions “very confrontational” and says that the probation service has “some catching up” to do when it comes to the correct use of algorithms.
“We are very concerned about this and I want to take responsibility for this,” says the director. “I wouldn’t be worth a damn if I didn’t do that. At the probation service we try to teach people every day that you have to take responsibility for your mistakes and try to fix them.”
About 48,000 times a year, the probation service advises judges about the risk that someone will make a mistake again, the so-called risk of recidivism. It could be a suspect on trial or a prisoner about to be released. This advice carries a lot of weight and can mean the difference between releasing or detaining.
Outdated data from Sweden
The inspection report shows that the calculation models used by the probation service were wrong in approximately one in five cases. One of those algorithms, called Oxrec, does not meet the standards for government algorithms in almost any respect, the inspectorate concludes. Crucial formulas had been mixed up in the algorithm, drug problems were not properly included in the model and serious mental disorders were not included.
Moreover, the model ran on outdated data from Swedish detainees. “It is very questionable whether a dataset that makes predictions for the Swedish population can also be applied to the Netherlands,” says Cynthia Liem, associate professor at TU Delft and specialized in the responsible use of AI and algorithms.
Liem calls it “embarrassing” that the model has never been applied correctly and that it has never been checked in recent years. She attributes this, among other things, to the strong trust that organizations often place in computer models. “AI and algorithms make it very attractive for us to think less, question less and simply accept what is offered to us. You see this happening in many technological applications.”
This is one of the most painful reports we have seen in recent years.
The probation service has now paused the use of the models and has set up a program to scrutinize the use of these so-called risk assessment instruments.
According to the inspectorate, the risk of recidivism was estimated too low for most of the errors. A possible consequence is that society is insufficiently protected. If the estimate is too high, this could be detrimental to the suspect or convicted person, for example because he or she will then receive a higher sentence.
When asked whether people may have been released too early or punished more severely due to an incorrect assessment, director Westerik answers that she cannot rule that out. “But I think the chance is very small.”
Discrimination
“This is one of the most painful reports we have seen in recent years,” says Sven Stevenson of the Dutch Data Protection Authority. “As a citizen, you must be able to trust that such an algorithm is fair and does not discriminate. Both are not in order with several of these algorithms.”
It is extra painful that the probation service was already closed in 2020 warned for discriminative elements in the Oxrec algorithm. In that model, the characteristics of zip code and income were taken into account, and this can indirectly lead to ethnic profiling. Westerik says that the probation service took this signal “very seriously” at the time. “But at the time there seemed to be enough scientific evidence that neighborhood and income were separate factors.”
But a year later, the Netherlands Institute for Human Rights also said that it is prohibited to take these characteristics into account without properly substantiating this. And that substantiation was missing. “We haven’t done enough of that, that’s right,” says Westerik.
The director acknowledges that there was insufficient knowledge within her organization about algorithms and their maintenance. “We have to improve that and we will do so. But I also want to emphasize that probation is a human job. All our training material states that the professional judgment of our employees is leading.”
date:2026-02-12 23:03:00