180 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 30 August 2026
← The incident library
HD-INC-127
Welfare and social security · Denmark · 2024 · Fraud-detection profiling of welfare claimants

Amnesty International reported that Denmark's automated welfare fraud system ran up to sixty algorithms over millions of people, using data like citizenship to flag migrants, disabled and low-income claimants for investigation

By Ellie Harris · Filed Amnesty study published 12 November 2024

Alleged: Udbetaling Danmark; ATP (administrator) developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

Amnesty International reported that Denmark's automated welfare fraud system ran up to sixty algorithms over millions of people, using data like citizenship to flag migrants, disabled and low-income claimants for investigation

What happened

It was reported that in November 2024 Amnesty International published a study, Coded Injustice, into how Denmark hunts for welfare fraud. It found that the agency Udbetaling Danmark, and ATP, the body that runs its systems, used up to sixty algorithms to score claimants and decide who to investigate, drawing on a wide sweep of data that included residency status, citizenship, place of birth, family relationships, travel, and health. Amnesty, which was given partial access to four of the models, said the sheer reach of the data amounted to mass surveillance.

It was reported that Amnesty found the system risked treating some groups unfairly. One model, it said, flagged people for having what the system judged to be strong ties to countries outside the European Economic Area, using citizenship as a factor; another flagged living arrangements it treated as unusual, which Amnesty said could catch disabled couples who lived apart or migrant households of several generations. Amnesty said the effect fell hardest on migrants, disabled people and those on low incomes. Udbetaling Danmark and ATP disputed the findings, said that using citizenship in this way was not the processing of sensitive data, and declined Amnesty’s request to open the models to a full independent audit. Amnesty said the people the system scored were not told why they had been flagged.

What an auditable version would have shown

A person pulled in for a fraud investigation could not see what had singled them out: which data, which rule, which score. An auditable version keeps a record for each person it flags, showing what triggered the flag and why, and makes it reachable to the person and to an oversight body, so a flag can be explained and, if it is wrong or unlawful, challenged. It would also show the bigger pattern: how the flags land across groups like citizenship, disability and income, so a system that falls hardest on one group is something the agency can measure and a regulator can read, not something an outside body has to guess at from partial access.

Where the gap was

An automated system decided who among millions of claimants to investigate, on data that reached into citizenship, family and health, and neither the flagged person nor an independent auditor could see how it worked. A ConductRecord keeps each flag with the data and rule behind it, so a person can be told why they were investigated and can contest it. A MetricRecord counts how the flags land across groups, so a system that targets migrants or disabled people shows up as a number that can be checked, not a claim the agency can simply deny. When it refused a full audit, this is what the agency held back: a record open enough for someone outside to test whether the system was fair and lawful.

What governance should have looked like

Where a state runs an automated system to decide who to investigate, the data and rules it uses have to be lawful, a person has to be able to see why they were flagged and to challenge it, and the effect across groups has to be measured and open to an independent check. Best practice would be for the agency to record each flag and its basis, to give a flagged person a real explanation, and to publish, or open to an auditor, how the flags fall across citizenship, disability and income. Amnesty could only estimate the effect because it was given partial access to four models out of sixty. A record built to be checked would let that question be answered in full, not in part.

Failure Pattern: an automated fraud system scored welfare claimants on data including citizenship and living arrangements to decide who to investigate, and the people it flagged could not see or challenge how they had been judged.

Governance Principle: where an automated system flags people for investigation, the data and rules it uses must be lawful and open to challenge, a person must be able to see why they were flagged, and the system’s effect across groups must be measured and independently checkable.

The reference implementation of ConductRecord and MetricRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Udbetaling Danmark; ATP (administrator) could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →