What happened

Dutch justice inspectors found that Reclassering Nederland (the probation service) had used an error‑ridden algorithmic system to  inform judges and prosecutors about individuals’ risk of reoffend. The system was used approximately 44,000 times a year. 

In one key tool, formulas were mixed up so the model never correctly predicted recidivism risk, leading to wrongly classified risk levels in around a quarter of probation reports used by judges and parole boards. 

These risk scores fed into advice on sentencing and early release, leading the tool consistently to underestimate the risk of reoffending, particularly among individuals with substance abuse issues or serious mental health conditions. 

Others may have received unnecessarily harsh or restrictive conditions. 

The inspection also found that the OXREC risk model used neighbourhood “poverty scores” and income data, variables known to correlate with migration background, which created a risk of indirect discrimination against poorer and racialised groups. 

The Inspectorate also found that staff were pressured to ignore their professional judgment. An internal podcast even told employees their own intuition was "as reliable as flipping a coin," leading to automation bias, where the machine's errors were rarely questioned.

Following public reporting and the inspectorate’s conclusions, the probation service temporarily halted use of the systems and acknowledged serious shortcomings in how the algorithms had been designed and managed. 

Why it happened

The core issue appears to stem from design and implementation flaws: the algorithm was not properly calibrated for the Dutch probation population and used outdated or mismatched data, while local deployment introduced errors such as applying models intended for one group to another. 

These technical shortcomings, coupled with a tendency for staff to over-rely on the algorithm’s outputs rather than professional judgment, amplified the impact of inaccurate scores in judicial contexts. 

Despite signals as early as 2018 about limited predictive value and a 2020 academic warning about possible discriminatory effects, the service continued using the model and did not conduct robust human‑rights or privacy impact assessments. 

The algorithm included income and neighbourhood “deprivation” scores even though the Dutch Human Rights Institute had previously warned that such variables are effectively prohibited unless strong safeguards and justification are in place, which were missing here. 

More broadly, the case reflects a familiar pattern in Dutch public sector algorithm use, where pressure to automate risk assessment outpaced transparency, documentation, and accountability mechanisms, similar to earlier welfare and tax‑benefit profiling scandals.