Denmark's Udbetaling Danmark runs over 60 algorithmic models to flag suspected benefits fraud. Amnesty's 2024 'Coded Injustice' report found several disproportionately target disabled, low-income, migrant and refugee claimants, prompting a Parliamentary Ombudsman inquiry.
>60
Algorithmic/ML fraud-detection models in use
4 of 60
Models for which documentation was obtained
Details
Maturity
Established
Promoter
Udbetaling Danmark (ATP) / NNIT
Period
2012–2025
Keywords
social welfare, fraud detection, algorithmic risk scoring
Context
Udbetaling Danmark (UDK) has centralised most Danish welfare payments since 2012 in partnership with pension fund ATP.
Activities
UDK uses more than 60 algorithmic and machine-learning models, built by IT contractor NNIT, to flag applicants and claimants for suspected benefits-fraud investigations.
Results
Amnesty International's November 2024 report 'Coded Injustice' obtained partial documentation on only 4 of the 60 models but found that named models such as 'Really Single' (flagging 'atypical' cohabitation/living patterns) and 'Model Abroad' (flagging 'medium to high' ties to non-EEA countries) systematically increase scrutiny of people with disabilities, low-income individuals, migrants, refugees and people in non-traditional households.
Conclusions
Amnesty argues the fraud-control system likely falls under the EU AI Act's 'social scoring' prohibition (effective 1 August 2024); the Danish governing coalition declined to suspend the models, promising only 'ongoing internal monitoring', and the Parliamentary Ombudsman opened a formal inquiry in June 2024 into flagship models including the 'Gladsaxe Model'.
Implementation
Indicative cost
High (€500k–€5M)
Time to results
Long (> 3 years)
Staffing & skills
Udbetaling Danmark (ATP) caseworkers, NNIT (IT contractor that built the models)
Conditions for success
Centralised welfare-payment administration (since 2012) gives UDK the cross-programme data needed to run risk-scoring models
Common failure modes
Amnesty's 2024 investigation found several named models systematically increase scrutiny of disabled, low-income, migrant, refugee and non-traditional-household claimants, based on documentation obtained for only 4 of 60 models
Danish Parliamentary Ombudsman opened a formal inquiry (June 2024) into flagship models including the 'Gladsaxe Model'; the governing coalition declined to suspend the models
Commonly funded by
National / regional programmes
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
Do you run this practice?
Claim it —
verified implementers get a public contact pathway and can propose corrections.
Data sources
Where this practice's information was retrieved from, and when.
Amsterdam spent about five years and roughly €535,000 building Smart Check, a welfare-fraud screening algorithm engineered to avoid discrimination. A …
