NetherlandsPublic sector2017–2021City of Rotterdam
Machine-learning risk scoring of welfare recipients
Rotterdam put its welfare fraud model on hold in 2021 after finding bias it could not remove
Problem
Rotterdam wanted to pick which of its roughly 30,000 welfare recipients to investigate for fraud each year.
Approach
Since 2017 the city used a machine learning model that gave each recipient a score from 0 to 1 using 315 variables. About 1,000 people a year were referred for investigation. Lighthouse Reports analysed the model in 2023 and found it biased against some groups.
Results
Takeaway
Per Lighthouse, the city's own evaluations found the model only about 50 percent more accurate than random selection. The city suspended it in mid-2021 after assessments found bias it could not eliminate.
Categories
- Size
- Large enterprise
- Industry
- Public sector
- Business function
- Finance, risk and legalProduction and quality
- Innovation type
- Process innovation
- AI technology
- Predictive analytics and ML
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
Lighthouse Reports describes how it obtained and tested Rotterdam's risk-scoring model, including its inputs and score thresholds. It reports the city's internal accuracy finding and the 2021 suspension.
Full details are in the original sources.
