DUO's Student-Grant Fraud-Check Algorithm — Dutch Risk Profiling Found Indirectly Discriminatory

The Dutch Executive Agency for Education (DUO) used a simple rule-based risk-profiling algorithm between 2012 and 2023 to select students receiving a supplementary grant for home visits that checked whether they really lived away from their parents. The selection criteria were type of education (vocational/MBO scored higher), distance between the student's address and the parents' address (shorter scored higher) and age (younger scored higher). Nobody tested whether these proxies were justified or whether they produced unequal outcomes.

Reporting and a subsequent external review (PwC, published 1 March 2024) found that the system led indirectly to discrimination against students with a non-European migration background. Almost 27,000 students were visited for residency checks and nearly 10,000 were accused of fraud; of 376 cases reviewed by lawyers, 97% involved students with ethnic-minority roots. The Education Minister apologised and, in April 2026, announced compensation of €500 for about 12,000 visited students and €2,000 for about 10,000 students whose grants were cut or who were otherwise sanctioned.

An independent academic audit (Holstege et al., arXiv 2502.01713) applied an unsupervised bias-detection tool, built to work without demographic data, to risk scores for over 250,000 students from 2012–2023 and highlighted the same known disparities. The tool was released as open source.

The case is a cautionary practice: a transparent, even simplistic model can discriminate when proxies correlate with ethnicity and no bias testing, documentation or contestability exists. Its lessons are pre-deployment bias audits, a documented justification for every selection criterion, and ongoing monitoring, and the open-source audit tool is a reusable outcome.

Read the full analysis: https://www.dutchnews.nl/2024/03/student-finance-group-duo-did-discriminate-in-fraud-probes/