UK DWP Universal Credit Advances fraud model

The Universal Credit Advances model is a machine-learning model run by the UK Department for Work and Pensions (DWP). It scores every request for an advance, an up-front loan against a new claimant's first payment, for fraud risk. It refers the highest-risk requests to a caseworker, and lower-risk requests are processed automatically.

How it is used

The model has been in production since May 2022. From features of the claim, it outputs a probability that an advance is fraudulent. Its training target is drawn from historic Advances outcomes.

DWP says the model is around three times more effective at identifying fraud risk than a randomised control sample, a set of requests picked at random.

The National Audit Office, which audits government spending, reports realised savings of roughly £4.4m over 2022-23 to 2024-25. It sets them against about 1.4 million advances worth £0.8bn paid in 2024/25. It also gives an estimated band of £0m to £60m for fraud and error on advances.

Big Brother Watch, an advocacy group, estimates the model profiles on the order of a million people a year. It reports that DWP went to court to withhold information about the model's data risks.

Who decides

DWP says a human caseworker always makes the final approve-or-decline decision, with no automated decision-making. By design, the caseworker is not shown the risk score and not told the referral came from the model. High-risk cases are mixed into the caseworker's queue with control-group cases.

A declined advance does not bar future applications or affect the claimant's underlying entitlement. Advances that are referred and then approved see about a day's extra delay.

What DWP's own assessment found

A fairness analysis carried out in February 2024 found statistically significant differences between groups in referrals and outcomes. DWP summarised it as presenting "no immediate concerns". The detail came out only when the Public Law Project, an advocacy group, obtained it under freedom-of-information law. The Guardian reported it on 6 December 2024.

On 17 July 2025, DWP published a fuller fairness assessment, covering 1 April 2024 to 31 March 2025. It compares how likely each age group is to be referred with claimants aged 35 to 44.

Claimants aged 16 to 24 were 1.67 times as likely to be referred. Those aged 45 to 54 were 1.35 times as likely, and those aged 55 to 65 were 2.80 times. Claimants aged 66 and over were 49.24 times as likely. DWP says that figure rests on 0.1 percent of the cases analysed, and should be treated with caution.

The assessment also documents an accuracy inversion. Older claimants are referred more often, yet their referrals are less likely to be correct. The sources read for this case do not say how DWP judges a referral correct. A referral was 0.58 times as likely to be correct at 55 to 65. At 66 and over it was 0.23 times as likely, a figure that also rests on a small sample.

Non-UK nationals were 2.27 times as likely to be referred as UK nationals. Their referrals were about as likely to be correct, at 0.97.

These are ratios DWP itself reported, comparing groups. They are not error rates audited by anyone outside DWP.

What each side says

DWP judges continued operation "reasonable and proportionate", resting on the safeguard that a human always makes the final decision. It has committed to retrain the model to address the age and nationality disparities.

The Public Law Project says that only age was fully assessed among protected characteristics, the traits equality law protects. The data on the others fell short of a completeness standard.

It argues that the "no immediate concerns" conclusion rested on safeguards against later harm, not on showing that the model does not discriminate.

It says race, sex, sexual orientation, religion or belief, pregnancy or maternity, and gender reassignment were never assessed. The assessment did report referral figures by nationality. So, it argues, proxy discrimination cannot be ruled out. That is unequal treatment through other data that stands in for a protected characteristic.

Who else has examined it

The National Audit Office examined DWP's wider counter-fraud programme in two 2025 value-for-money reports. Parliament's Public Accounts Committee raised concern about the impact of machine learning on vulnerable claimants. It pressed DWP to report to Parliament every year on the impact on protected characteristics.

Where the programme is going

The programme is growing. The Public Authorities (Fraud, Error and Recovery) Act 2025 gives DWP powers to check eligibility against bank data. The sources describe this as a distinct system, not yet live.

The government says the Office for Budget Responsibility, the UK's official forecaster, has assessed the savings it targets at £1.5bn by 2029/30. Implementation is set from 2026, with up to 3,000 additional staff.

That system is not part of this network.

What makes this case different

In most algorithmic-bias cases the case file compares, the audit had to come from outside. In Rotterdam, journalists pried a model open. In Chile, a bias review was commissioned and then withheld.

Here DWP audited its own live model, found the disparities, and published the numbers itself. Then it kept the model running, on the grounds that a human always decides.

So the question is why a successful internal audit changed so little. The case file's answer is how the pieces are connected. The assessment is real oversight, but it runs once a year and reports after the fact.

The case file finds no live link from its finding to a check on the score between reports. It describes the retrain as committed, not yet made. A finding published but never tied to a correction is documentation, not control.

Two features of the design

First, the model learns from its own past. Its training target comes from historic Advances outcomes, and caseworker decisions become the next training labels. So the skew DWP documented in its referrals can shape the next model's scores.

Second, the caseworker's blinding cuts two ways. It holds back deference, the habit of accepting a system's output without checking it. It also removes the caseworker's means to check the score or contest it.

So the finding could return only through a check on the model itself. The case file says that check is exactly what is missing.

The case file calls this its clearest argument that transparency and correction are different things. DWP published the skew, and the skew persisted. The gap matters more as the same programme grows into bank-data checks.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A pathway counts as closed once it passes on only a few mistakes.

This case has a budget of 13 units. Explore (No Targets) sets no targets.

Under Service Targets Only, the targets can be met within the budget. The cheapest way costs 2 units and uses one tool, Mark AI-written records.

Under Service and Safety Targets, the targets also include closing every failure pathway. They can be met. The cheapest way costs 10 units and uses four tools: Vet connections, Store less data, Escalate checks, and Keep prompts neutral.

Under All Governance Targets, the targets can be met too. The cheapest ways cost 12 units and use five tools each. Two use Vet connections, Understand the system, Store less data, Escalate checks, and Keep prompts neutral, with Understand the system at either setting. The third swaps the first two for Check with a second model and Mark AI-written records.

At both of these levels, every combination that meets the targets includes Store less data, Escalate checks, and Keep prompts neutral. There, Understand the system costs 4 units. While it is in use, Upgrade model, Mark AI-written records, Vet connections, and Store less data each cost 1 unit less. They cost 2 less when Understand the system is at its stronger setting, which costs 6.

Under All Governance Targets, Vet connections changes nothing on this network unless Understand the system is also in use. Without it, Store less data works at reduced strength.

More checking is not always better here. Even if the budget allowed it, all eleven tools at their standard settings cost 24 units at the two upper levels, and 36 at their strongest. Either way they close every failure pathway, yet meet the targets at none of the three levels that set them. The added checks leave the model adding too little to the work.