Rotterdam's Welfare Fraud Risk Model — a 315-Variable Algorithm Suspended After Audit and Investigative Scrutiny

Rotterdam's city government used a machine-learning model, built with the consultancy Accenture, to give each of its roughly 30,000 welfare recipients a fraud-risk score between 0 and 1. Caseworkers then investigated the highest-scoring people. According to Follow the Money's analysis of the code, the model took 315 input variables, including age, gender, marital status, about 20 Dutch-language variables and subjective caseworker assessments. It was trained on about 12,700 recipients who had been investigated before.
In 2023 Lighthouse Reports and VPRO obtained the code and technical documentation, which the city had voluntarily disclosed in part; the training data was not disclosed. Their analysis found that people with poor Dutch scored about twice as often in the highest-risk group, women scored higher merely for being female, and single mothers were flagged 43% more often than childless single women. Lighthouse reports that hundreds of recipients had benefits terminated after investigations. In Follow the Money's reading, the model found fraud at about 39 per 100 checks against 25 for random selection.
Rotterdam suspended the system in late 2021 after its municipal audit office (Rekenkamer Rotterdam) raised ethical risks about discriminatory grounds and a government-commissioned ethics review was critical. The city said it would develop a revised version. The case is included as a cautionary example: independent scrutiny stopped the model, but only after years of operation, and the public register model was not yet in place.

Read the full analysis: https://www.ftm.eu/articles/algorithm-rotterdam-dissected