NAO report confirms DWP machine learning trial for Universal Credit fraud detection

The Department for Work and Pensions has been trialling a machine learning algorithm to detect fraud in Universal Credit claims and expects soon to use the model to stop payments before they are made, according to the National Audit Office report on the department's 2021-22 accounts, published on Thursday 7 July¹.

The Public Law Project, which reported the confirmation, says the algorithm analyses historical data to predict which cases are likely to be fraudulent in the future, without being explicitly programmed by a human¹. The information appears at paragraphs 48 and 49 of the NAO report¹. The Public Law Project says it believes this is the first confirmation that the department is using machine learning to detect fraud in Universal Credit, although there were already indications that the DWP used automated systems to assess benefit entitlement or flag cases for investigation¹.

The NAO report notes that the DWP intends to monitor the model for unintended bias and is aware that if groups with protected characteristics are disproportionately impacted, the model could obstruct fair access to benefits¹.

Ariane Adam, Legal Director of the Public Law Project, said:

She added that "without transparency there can be no evaluation, and without evaluation it is not possible to tell if a system works reliably, lawfully or fairly", and that "in the midst cost-of-living crisis, people could have benefits stopped before they are even paid out because a computer algorithm said 'no'"¹. She also said departments across government need to commit to more than being "aware" of the risks, and that the presumption should be that detailed information about how automated decision-making tools work is made available, with any exemptions justified as necessary and proportionate rather than the default¹.

Why it matters for households

Universal Credit is a means-tested benefit paid to people on low incomes or out of work, and the housing element within it helps with rent for many claimants. The confirmation concerns a fraud detection model, not the calculation of awards, but the stated intention is to use it to stop payments before they are made¹. That means a claim could be halted at the point of payment rather than after a review or investigation, on the basis of a prediction drawn from historical data¹.

The Public Law Project raises the possibility that historic data may be inaccurate or may carry human bias that the machine exacerbates, and that this could unfairly penalise or discriminate against marginalised or vulnerable groups¹. The NAO report records that the DWP intends to monitor for unintended bias and is aware the model could obstruct fair access to benefits if groups with protected characteristics are disproportionately impacted¹. No detail has been reported on which claimants the model applies to, what data feeds it, how a decision to stop a payment would be communicated, or what challenge or appeal route would exist. The department has previously refused Freedom of Information requests about its use of automation in assessing Universal Credit applications¹.

What happens next

The Public Law Project says the DWP "expects soon to use the model to stop payments before they are even made"¹. No date for that change has been reported, and the department has not published the trial data or the detail of how the model works¹. The Public Law Project is calling for government departments to be transparent about their use of algorithms and to publish data and analysis gathered from trials without charities having to make repeated Freedom of Information requests¹.