What happened
The Department for Work and Pensions (DWP) reported that it has used a machine learning model since May 2022 to check Universal Credit advance requests for possible fraud. An advance gives someone part of their benefit early, which they repay from later payments. The DWP says the model flags requests it considers high risk. A caseworker then checks the request, and the claimant has to provide more evidence.
In July 2025, the DWP published a fairness assessment covering April 2024 to March 2025. It found that claimants aged 45 and over, and non-UK nationals, were sent for extra checks more often. But those referrals were no more likely, or were less likely, to prove correct.
The DWP wrote that “the model is not working as effectively as we would expect”. It decided to keep using the model while retraining it. The National Audit Office (NAO) set this out in October 2025.
The DWP’s July 2026 assessment reported that the retrained model was still being tested. It also found that non-UK nationals were 3.35 times as likely to be referred as UK nationals, compared with 2.27 a year earlier.
What an auditable version would have shown
The DWP reported these differences in its own assessments. It compared how often the model flagged people in each group with how often a caseworker went on to refuse their advance.
The DWP also reported limits to what it could measure. In 2025, age was the only protected characteristic it could analyse because too few referred claimants answered its equality questions. It looked at nationality separately, without then treating it as a protected characteristic.
In 2026, the DWP analysed both age and nationality, treating nationality as part of race. It did not report separate analyses of disability, sex or ethnicity.
The DWP reported that claimants aged 66 and over were 49.24 times as likely to be referred in 2025. It said this figure should be “treated with caution” because the group made up just 0.1 per cent of the data.
The sources reviewed do not count how many legitimate claimants were referred or refused.
Where the gap was
The DWP had found that its model sent some groups for extra checks more often, without those checks proving correct more often. It said there was “minimal concern” about discrimination, unfair treatment or detrimental impact on claimants. The department kept using the model while testing a retrained version.
A MetricRecord is designed to hold signed aggregate figures, such as referral and correct-referral rates by group, bound to the underlying decisions so they can be checked independently.
A ConductRecord keeps each referral, the caseworker’s decision and the outcome.
MetricRecord and ConductRecord are Headlights designs.
What governance should have looked like
The DWP says the model does not decide whether someone gets an advance. A caseworker makes that decision without being told the model flagged the request. The department reported that people whose advances were approved after extra checks waited a median of one additional day for payment. It also said that someone refused an advance can apply again.
The DWP found the model about three times more effective than random selection in 2025 and 2.5 times in 2026.
In August 2026, the Independent Monitoring Authority for the Citizens’ Rights Agreements, which oversees the rights of EU and EEA EFTA citizens in the UK, said the DWP had assured it that “nationality, or similar, is not a risk factor in the model”. The authority took no further action.
When a public body finds that its model sends some groups for extra checks more often without those checks proving correct more often, best practice would be to set a deadline for improving it. The body should publish how it will measure the results and report whether the differences between groups have narrowed.
Failure Pattern: the DWP’s own analysis found its fraud model referred older and non-UK claimants for extra checks more often, without those referrals proving correct more often, and the department kept using it while a retrained model was tested.
Governance Principle: a public body using a risk model should be able to show, by group, how often the model refers people, how often those referrals prove correct, and what it did when the two diverged.
The reference implementation of MetricRecord and ConductRecord 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
- UC Advances model fairness assessment, July 2025 (Department for Work and Pensions)
- Effectiveness Assessment of Universal Credit Advances Model (Department for Work and Pensions, 9 July 2026)
- Tackling benefit overpayments due to fraud and error (National Audit Office, 22 October 2025)
- DWP’s fraud investigations, risk reviews and suspensions of Universal Credit (Independent Monitoring Authority for the Citizens’ Rights Agreements, 6 August 2026)