AI in Government Benefits Administration

Benefits agencies handle huge caseloads with limited staff, so automation is attractive. Systems range from rules engines that apply eligibility formulas, to data-matching programs that compare records across agencies, to machine-learning models that score cases for fraud risk. Australia's Robodebt is one of the most cited failures, and it was not sophisticated AI. From 2016 the scheme matched annual tax-office income with fortnightly welfare records and averaged the annual figure evenly across fortnights. People with irregular earnings, such as casual workers, appeared overpaid, and debt notices went out automatically, with recipients left to disprove them. The government later conceded that averaging alone was not a lawful basis for debts, settled a class action, and a Royal Commission that reported in 2023 strongly condemned the scheme. In the Netherlands, the childcare benefits scandal saw the tax authority wrongly treat tens of thousands of parents as fraudsters and demand large repayments. A risk-classification model that used nationality as a risk factor was part of the problem, alongside harsh all-or-nothing rules. The cabinet resigned in January 2021. Separately, a Dutch court in 2020 halted SyRI, a welfare-fraud data-linking system, on privacy and human-rights grounds. In the United States, Michigan's MiDAS system issued automated fraud findings that a state review found were mostly erroneous, and Arkansas faced litigation after an algorithm cut home-care hours for disabled Medicaid recipients. The lessons repeat: automated debts reverse the burden of proof; fraud models can encode proxies for ethnicity or poverty; and scale turns small error rates into mass harm. Safeguards include human review before adverse decisions, notices that explain the reason, accessible appeals, bias testing and public disclosure of systems in use. The EU AI Act classifies AI used by public authorities to assess eligibility for essential public benefits as high-risk.