Försäkringskassan's discriminatory welfare-fraud AI (Sweden)

Försäkringskassan (the Swedish Social Insurance Agency) deployed a machine learning algorithm from at least 2013 to assign risk scores to social benefit applicants - primarily those claiming temporary child support - in order to prioritise cases for fraud investigation. The system operated without public awareness and was described internally as the agency's 'best-kept secret.'

In November 2024, Lighthouse Reports and Svenska Dagbladet published the investigation 'Sweden's Suspicion Machine,' revealing that the algorithm disproportionately flagged women, people with migrant backgrounds, low-income earners, and those without university degrees. Testing against six standard statistical fairness metrics confirmed systemic bias, echoing a 2018 report by Sweden's Inspectorate for Social Insurance (ISF) that found the algorithm 'does not meet equal treatment' - findings Försäkringskassan rejected at the time while continuing to operate the system. Sweden's Data Protection Authority (IMY) launched an investigation following the Lighthouse Reports revelations, and by June 2025 Försäkringskassan suspended the model, later conceding 'shortcomings in transparency and risk management' and adding temporary manual oversight.

Sweden's Integrity Committee had raised concerns as early as 2016 about risks to citizens' personal integrity, and the agency rejected virtually all freedom-of-information requests about the system over multiple years; Amnesty International called on Swedish authorities to discontinue the system in November 2024. This case parallels the Netherlands' SyRI and is a cautionary example of how ML risk-scoring in social security can embed structural inequality when deployed without adequate fairness auditing, meaningful redress mechanisms, or genuine regulatory scrutiny.