DWP publishes fairness analysis of machine learning model for the first time

The Department for Work and Pensions (DWP) published detailed information on its fairness analysis of a machine learning model for the first time in July 2025, according to the National Audit Office¹. The model has been used since May 2022 to flag potentially fraudulent claims for Universal Credit advances¹. The NAO's report on benefit overpayments, published on 22 October 2025, states that older claimants and non-UK nationals are over-referred for review under the model¹.

The DWP has deployed one machine learning model, with four others in development and testing¹. The NAO reports that the department paid £290.8 billion in benefits including State Pension in 2024-25 and spent £7.3 billion on running costs, making welfare payments to more than 23 million people across Great Britain¹.

The fairness analysis was published as part of wider work on fraud and error. The NAO found that the estimated proportion of benefit expenditure overpaid fell from 3.6% (£9.7 billion) in 2023-24 to 3.3% (£9.5 billion) in 2024-25¹. Universal Credit accounted for 67% of overpayments by value in 2024-25, and its estimated overpayment rate fell from 12.4% to 9.7% over the same period¹. For the first time since Universal Credit was rolled out nationally in 2018, it did not have the highest overpayment rate across all benefit lines; Pension Credit had the highest estimated rate at 10.3%¹.

The NAO reports that DWP rated as "red" the risk that its plans to reduce fraud and error are not successfully executed or cannot mitigate the increased propensity for fraud in society¹. Successive Comptroller and Auditor Generals have qualified their audit opinions on the regularity of DWP's accounts, excluding State Pension, for 37 years because of material fraud and error¹.

The government introduced the Public Authorities (Fraud, Error and Recovery) Bill to Parliament in January 2025¹. DWP's refreshed fraud and error strategy was approved in November 2024, setting five strategic objectives focused on preventing inaccurate payments¹. The programme began in February 2022 with seven agents and by April 2024 involved 3,100 DWP staff¹.

DWP has been awarded £6.7 billion of dedicated funding for fraud and error activity covering 2020-21 to 2028-29, with 52% of the total (£3.5 billion) due in the three years from 2026-27¹. It saved an estimated £4.5 billion in Annually Managed Expenditure through counter-fraud activities from April 2022 to March 2025, achieving £1.35 billion in 2023-24 against a £1.3 billion target and £2.0 billion in 2024-25 against a £1.7 billion target¹.

Why it matters for households

People claiming benefits can be referred for review under the machine learning model, which flags potentially fraudulent Universal Credit advance claims. The NAO's finding that older claimants and non-UK nationals are over-referred means these groups are more likely to face checks than others, though the report does not set out what happens to a claimant once referred or how long a review takes.

The overall overpayment rate fell to 3.3% in 2024-25, but remains above the 2.4% recorded in 2019-20 using the National Statistic, or 3.1% using DWP's cross-welfare rate¹. The Spring Statement 2025 forecast that overpayments would fall to the pre-pandemic level of 3.1% by 2028-29¹. The NAO notes that DWP's IT systems are not fully integrated and do not allow staff to view all the information held about a claimant, which it says makes it less likely that incorrect payments will be prevented or detected¹.

What happens next

The Public Authorities (Fraud, Error and Recovery) Bill is before Parliament, having been introduced in January 2025¹. DWP had started to develop implementation and evaluation plans to support delivery of its strategy and measure its success at the time of the NAO's work¹. The NAO reports that DWP is in the early stages of assessing its strategic controls framework, with a view to evaluating cost-effectiveness and strengthening controls where necessary¹. The NAO's report does not state whether further fairness analyses will be published.