Prevention & enforcement: SWAN algorithm

Social Insurance Bank

With this model, we predict how likely it is that someone living alone might live together. The development of this model has been stopped.

Last change on 28th of October 2024, at 12:14 (CET) | Publication Standard 1.0

• Publication category: High-Risk AI-system
• Impact assessment: IAMA
• Status: In development

General information

Theme

Social Security

Begin date

Field not filled in.

Contact information

[e-mail]

Link to source registration

https://www.svb.nl/nl/over-de-svb/hoe-werken-we/algoritmes-preventie-en-handhaving

Responsible use

Goal and impact

The SWAN model allows us to focus on reducing risks. The aim is to identify and adjust situations where customers may be receiving too much state pension more quickly. This keeps the amount you have to repay low. We also ensure that supervisors are put to good use. The model helps us save money on benefits and on our work. To check if someone is cohabiting, our Prevention & Enforcement Department can visit you at home. This can have a big impact. That is why we use the SWAN model to select clients with a higher risk of illegality.

Considerations

This is a self-learning algorithm that can select customers at higher risk of cohabitation. The law states what data we are allowed to use. It also looks at what is fair and proportionate. Because checks can be more targeted, the algorithm prevents customers from receiving unnecessary home visits.

Human intervention

The algorithm selects customers with an increased likelihood of cohabitation. An employee conducts further investigation and advises whether there is illegality; not the algorithm.

Risk management

We always test our models for compliance. These are rules about ethics and verifiability. Rules we use are from the General Court of Auditors and the rules we set up ourselves.

We also recognise standard three-lines-of-defence (3LoD model). The first line is supported through various tools in quality thinking and acting and continuity risk management. The second line consists of departments such as Operational control, Risk management, Quality Assurance and Information Security, which perform and report on quality and performance controls. The third line consists of the audit serves that perform scheduled and unscheduled audits.

Legal basis

Enforcement duties within the General Old Age Pensions Act (AOW)

Elaboration on impact assessments

As this is a high-risk algorithm, an IAMA has been prepared for this and has been worked within the existing DPIA. When development of this algorithm resumes, a new DPIA will be conducted.

Impact assessment

Impact Assessment Mensenrechten en Algoritmes (IAMA)

Operations

Data

Our main sources are internal data available about the customer. For example:

• information about the address where the customer currently lives or has lived in the past
• living situation
• age

Technical design

Self-learning. A person living together receives less AOW than a single person. A regression model is trained based on historical data of AOW irregularities. This model then assigns a risk score to customers within the population. This risk score says something about how likely a single person is to live together anyway.