Amsterdam Wanted To Use AI To Make Social Assistance Fairer and More Efficient. Things Turned Out Differently.

This story was originally published in Trouw, a Dutch newspaper. To read the complete report, click here.

The Dutch government has a bad reputation when it comes to deploying algorithms in the social welfare system. While trying to combat benefit fraud, innocent citizens have been victimized by biased algorithms. So when the municipality of Amsterdam decided to develop a computer model to make its welfare system more efficient, it made sure to avoid such disasters.

The progressive capital took a different approach. No derailed manhunt for alleged fraudsters. No secrecy. No biased algorithms that disadvantage vulnerable groups. This time, artificial intelligence had to serve the citizens. The municipality tried to develop a fair and ethical machine-learning algorithm, which would help to determine which benefit applications should be scrutinized more closely. The aim was not to catch as many fraudsters as possible, but to prevent vulnerable citizens from getting into debt problems. After all, if applicants receive more benefits than they are entitled to, they are required to pay it back.

But despite these noble intentions, the municipality failed to develop an algorithm that was fair beyond doubt. While the computer model passed the test phase after some tweaking, the algorithm wrongly labelled certain population groups as potential fraud risks at a disproportionate scale once released in the real world. The project was killed.

---

As a nonprofit journalism organization, we depend on your support to fund more than 170 reporting projects every year on critical global and local issues. Donate any amount today to become a Pulitzer Center Champion and receive exclusive benefits!

---