DWP’s fraud model for Universal Credit advances

Universal Credit claimants who need money before their first payment can apply for an advance. The Department for Work and Pensions (DWP) uses a machine learning model to help decide which advance claims should be checked for possible fraud.¹

In December 2024, The Guardian reported on internal DWP documents released under the Freedom of Information Act, including a fairness analysis of the system carried out in February 2024.¹

What the analysis found

According to the documents, the analysis identified a “statistically significant outcome disparity”. The system was incorrectly selecting people from some groups more often than others when recommending who should be investigated. The groups were defined by age, disability, marital status and nationality.¹

The documents also showed what had not been examined. As reported, no fairness analysis had yet been carried out in respect of race, sex, sexual orientation, religion, or pregnancy, maternity and gender reassignment.¹

The department’s position

The DWP said the system “does not present any immediate concerns of discrimination, unfair treatment or detrimental impact on customers”, and stressed that a human still makes the final decision about whether a person receives a payment.¹

Campaigners took a different view. They accused the government of a “hurt first, fix later” approach and called on ministers to be more open about which groups were likely to be wrongly flagged.¹

Why it is contested

This case does not show deliberate discrimination, and the DWP deserves some credit: it carried out a fairness analysis, and the documents were released. Fraud and error in the benefits system are real problems that the department is expected to tackle.

The difficulty is what happens next. Being selected for investigation is not neutral. It can delay payments to people who, by definition, are short of money. If a model selects some groups wrongly more often than others, a human making the “final decision” does not undo the burden of being investigated. And an analysis that has not yet looked at race or sex cannot say whether those groups are affected too.

Lessons for leaders

1. Treat selection for scrutiny as a decision with consequences. “A human makes the final decision” is reassuring only if the earlier, automated step is fair as well.
2. Test fairness across all protected characteristics, and before scaling. Partial testing leaves leaders unable to answer the question the public will ask.
3. Publish what you find and what you are doing about it. Disclosure through freedom of information is not the same as transparency. Explaining your findings and response builds more trust than having them discovered.

Questions to ask in your organisation

1. Where do we use AI to decide who gets extra scrutiny, and what does being selected cost the people affected?
2. Which protected characteristics have we tested our models against, and which have we not?
3. If our fairness testing found a disparity, who would decide whether it was acceptable, and would we publish that decision?