Public benefits & eligibility

Automated unemployment-fraud adjudication with no human review — the canonical high-error, zero-correction structure.

Unlawful income-averaging debt assessment at national scale, with the burden of proof reversed onto recipients.

Privatized, automated benefits-eligibility processing that produced mass denials and ended in contract collapse and litigation.

A machine-learning fraud-risk score whose audited bias led a city to suspend it — and a rare look inside a live scoring system.

Algorithmic allocation of home-care hours that cut care for many recipients and lost repeatedly on due-process grounds.

A modest fraud-risk model coupled to an uncorrectable 270,000-person blacklist and an all-or-nothing recovery regime — the institutional amplifier that toppled a government.

A secret cross-database welfare-fraud risk-profiling system struck down by a court on privacy and transparency grounds — before its individual harms were ever counted.

A national score that quietly rates about 32 million people for benefit-fraud suspicion each month — and whose own arithmetic pushes low income, disability, and single parenthood toward the highest-risk, most-invasive-control bracket.

A secret in-house model that picked women, migrants, low earners, and the less-educated for benefit-fraud investigation far more often — audited from outside but never opened, disputed by the agency, and retired before any regulator or court ruled.

A national welfare agency links about ten registers over roughly 2.4 million recipients and runs up to sixty fraud models -- standing surveillance where most people it flags did nothing wrong, the demographic harm cannot be measured, and no court has ruled.

The secret software behind Spain's electricity social bonus denied the benefit to eligible people and gave no reasons — until a Supreme Court transparency ruling ordered its source code disclosed to the foundation that had sued, a control the government was still resisting eight months on.

A statutory cross-registry data match that cut the social-assistance caseload by a reported range of tens of thousands — hitting the poorest, and Roma especially, while constitutional, World Bank, and UN scrutiny all stayed pending.

A live benefits-fraud scoring model whose own published fairness assessment shows it refers older claimants and non-UK nationals far more often — and, for older groups, less accurately — while the agency calls continued use proportionate and scales up.

A private facial-recognition identity check became a de facto eligibility gate for pandemic unemployment benefits in 25-plus states — with multi-hour verification queues, demographically skewed completion rates, and no federal count of how many eligible workers it blocked, because a verification failure is never recorded as a denial.

A federally required, protective renewal automation carried one wrong setting — it evaluated eligibility household-by-household instead of person-by-person — and dropped roughly 500,000 enrollees across 30 states, disproportionately children who were still eligible. The engine was accurate; the error was in the specification, and no accuracy metric would have surfaced it. What caught it was a well-instrumented federal monitor-and-respond loop, but only after mass terminations.

A 2024 TCU plenary audit found INSS benefit denials nonconforming above the acceptable limit in both channels it sampled — 10.94% of automatically analyzed denials and 13.20% of manually analyzed ones — and named the cause: server productivity is measured by the number of processes analyzed, not decision quality, so denial is the fastest disposition. Automation used to drain a benefit backlog inherits that metric, and a fast denial is corrected only through a judicial channel that runs for years. There is no risk-scoring model in the record; the harm is the measurement, not a fraud score.

An entity-resolution system that decides ration eligibility by matching people across thirty-plus databases — silently cancelling entitlements when a similarly-named stranger's assets are attributed to the applicant.

A predominantly automated welfare-compliance engine unlawfully cancelled at least 1,009 jobseekers' subsistence payments because a 2022 law requiring a discretionary reasonable-excuse decision before cancellation was never built into the system, and its mandated safeguard was never finished. Built after Robodebt and meant to have learned its lessons, the framework still automated unlawful cuts at national scale; the corrective loop that existed this time — an Ombudsman, an assurance review — fired about ten months after the unlawfulness was flagged, with long-lived IT bugs running undetected underneath.

A public-facing government generative adviser that confidently told business owners to break the law — on workers' tips, on housing-voucher discrimination, on cashless-business rules — and was kept online for roughly two years after the errors were published, then shut down as a budget cut rather than an accuracy fix.

A generative-AI tool built to draft the ruling — and the written decision itself — on unemployment-benefit appeals for a human referee to sign, under a 90% self-assessed accuracy floor. The clearest live test of generative adjudication and automation-deference in a due-process entitlement, and as of early 2026 still in delayed pre-deployment testing, not confirmed live.

A court-adjudicated modern eligibility system: automated Medicaid determinations whose defective, misrouted notices severed the right to appeal — held unlawful in operation.