UWV has screened unemployment-benefit claims with a risk-scoring algorithm since 2022. The Netherlands Court of Audit's 2025 review found flagged cases yield three times more confirmed fraud than random checks, but flagged unresolved IT-management risks.
~70 %
Cases selected by the algorithm
~30 %
Cases selected at random as control sample
3x
Confirmed culpable-unemployment rate, flagged cases vs. random sample (2024 review, published May 2025)
social security, unemployment insurance, fraud detection, benefits administration
Context
UWV, the Dutch employee-insurance agency administering unemployment (WW) benefits, has since 2022 applied an algorithmic 'Risicoscan Verwijtbare Werkloosheid' (culpable-unemployment risk score) to flag applications where a claimant may bear responsibility for their own job loss for closer manual review by a caseworker.
Activities
Around 70% of reviewed cases are selected by the algorithm; the remaining 30% are chosen at random specifically to test the algorithm's own accuracy and guard against bias — a genuine randomised control sample built into the operating design.
Results
The Netherlands Court of Audit (Algemene Rekenkamer) evaluated the algorithm as part of its 2024 annual review of government algorithms, publishing results in May 2025. It found that cases flagged by the risk score yield three times as many confirmed instances of culpable unemployment as the random control sample. The same review found UWV had mitigated most privacy and bias risks but flagged outstanding general IT-control weaknesses requiring remediation in 2025.
Conclusions
Context: the government's SyRI welfare-fraud detection system was struck down by Dutch courts in 2020 over rights violations, and the Tax Administration's own algorithms were separately flagged for GDPR non-compliance in the same audit cycle — making UWV's comparatively better-managed, but still imperfect, oversight a useful reference point.
Implementation
Indicative cost
Medium (€50k–€500k)
Time to results
Long (> 3 years)
Staffing & skills
UWV caseworkers conducting manual review of flagged and randomly-sampled cases, UWV algorithm/IT operations team, Netherlands Court of Audit (Algemene Rekenkamer) as independent auditor
Conditions for success
Built-in randomised control sample (30% of cases) enabling ongoing measurement of algorithm accuracy and bias
Independent government audit cycle providing external oversight
Mitigation of privacy and bias risks per the 2024/25 audit
Common failure modes
Audit flagged outstanding general IT-control weaknesses requiring remediation in 2025
Sector context includes prior failures (SyRI system struck down by Dutch courts in 2020; Tax Administration algorithms flagged for GDPR non-compliance in the same audit cycle), underscoring the risk profile of algorithmic welfare-fraud detection generally
Commonly funded by
National / regional programmes
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
Do you run this practice?
Claim it —
verified implementers get a public contact pathway and can propose corrections.
Data sources
Where this practice's information was retrieved from, and when.
Michigan's MiDAS system auto-adjudicated unemployment fraud claims with no human review (2013-2015), wrongly accusing 34,000+ people at an ~85% error …