NetherlandsPublic sector2019–2023

City of Amsterdam

Smart Check risk model for welfare applications

Amsterdam stopped its welfare-fraud risk model in 2023 after a pilot of about 1,600 applications flagged more people and showed new bias

StoppedPress reporting

Problem

Welfare officers checked applications for errors and fraud, and more than half of the investigations found no wrongdoing, according to city figures. The city wanted a fairer and more accurate way to pick cases.

Approach

Work began in 2019. The model was trained on 3,400 past investigations, used 15 characteristics and left out gender, nationality and age. Reweighting of the training data evened out error rates between Dutch and non-Dutch applicants in testing. A pilot started in 2023, and the city dropped it in November 2023 because it could not rule out unwanted discrimination.

Results

~1,600welfare applications run through the model in the live pilot, per MIT Technology Review's reportingPress reporting[1]

~€500,000city's own estimate of internal costs, plus a €35,000 contract with DeloitteCompany's own claim[1]

20% better than caseworkers (internal test)accuracy gain the city's team saw in internal testing before the pilot; in the pilot the model was no better than caseworkers, per the reportersCompany's own claim[1]

€2.4 million per year (expected)projected annual savings the city cited before the pilot; not achievedCompany's own claim[1]

Takeaway

MIT Technology Review reported renewed bias in the live pilot. The city reported fewer false positives than the old method but called its bias analysis inconclusive; development ended in November 2023.

Sources

Inside Amsterdam's high-stakes experiment to create fair welfare AI
MIT Technology Review11 Jun 2025News article
Slimme Check Onderzoeksproject (Smart Check research project)
City of Amsterdam, Open Research24 Jun 2024Government publication

Source summaries

[1]MIT Technology Review11 Jun 2025

MIT Technology Review, Lighthouse Reports and Trouw investigated the model using city records and their own bias tests. They found bias in 2022 testing and renewed bias after mitigation in a pilot of about 1,600 applications, which flagged more applicants. The city estimated internal costs of about €500,000, plus a €35,000 Deloitte contract, and had expected annual savings of €2.4 million.

[2]City of Amsterdam, Open Research24 Jun 2024

The city said the pilot ran from April to August 2023 after about three years of development. The final evaluation found fewer false positives than the old method, but the bias analysis was inconclusive. The Participation Council advised stopping in 2022, and the alderman decided to stop in November 2023 with the council's agreement.

Full details are in the original sources.

Added 25 Oct 2026 · Updated 25 Oct 2026