Benefit Overpayments From Fraud and Error

Paying welfare claimants the amount the law entitles them to, instead of losing money to undeclared earnings, savings, and household changes, without wrongly accusing or withholding payments from legitimate claimants.

AI is suitable with conditions

A benefits agency holds the labelled history a model needs, from past case reviews and investigations, and the UK Department for Work and Pensions reports that its machine learning model for Universal Credit advances finds risky requests around two and a half to three times as often as a random control sample and contributed to more than halving the estimated scale of advances fraud. AI has a real role because two of the three largest causes of Universal Credit fraud, undeclared self-employed earnings and undeclared partners, are circumstances no authoritative record verifies, so data matching cannot reach them and caseworker review has to be targeted somehow. It is conditional rather than strong because the fairness record is poor: DWP's own model over-referred older claimants and non-UK nationals, Rotterdam's risk model was reported to discriminate by ethnicity, age, gender, and parenthood, and Amsterdam stopped its pilot after the model wrongly flagged groups unevenly and found no more than caseworkers did. The model's measured savings, £4.4 million since 2022 on advances worth about £0.8 billion a year, are also small next to the £581 million DWP attributes to targeted case review, so the larger gains so far have come from measurement, matching, and people. A risk score belongs only where no record can settle the question, ranking work for a caseworker who decides, with a random control group, a referral blind to the score, and fairness measured on outcomes, and never deciding on its own.

Problem Statement

A benefits agency pays millions of people every few weeks, and each payment depends on circumstances that keep changing.

Some of those payments are too high, because a claimant’s earnings rose, a partner moved in, or savings grew, and the change was not reported or not acted on.

Some are too high because the agency itself made a mistake.

The agency has to find and prevent those overpayments without delaying, refusing, or accusing the far larger number of people who are being paid correctly.

Business Context

In Great Britain, the Department for Work and Pensions (DWP) estimated in May 2026 that it overpaid 3.2% of benefit expenditure in 2025-26, about £9.9 billion, and that after recoveries the net loss was 2.8%.

Universal Credit’s estimated overpayment rate was 8.5%, about £6.7 billion, down from a pandemic peak, which the National Audit Office (NAO) attributes to restoring controls suspended during the pandemic and scaling up case reviews.

DWP’s statistics give the three main causes of fraud overpayments as undeclared earnings, mostly from self-employment, failing to declare living with a partner, and undeclared capital, which together account for over £6 in every £10 overpaid to fraud.

Employee earnings are reported by employers and feed Universal Credit directly, but self-employed earnings and living arrangements are recorded nowhere the agency can check.

In the United States, the Government Accountability Office estimated that fraud in unemployment insurance during the pandemic was likely between $100 billion and $135 billion, or 11 to 15% of the benefits paid, an estimate the Department of Labor said was likely overstated.

The pressure to use data and AI comes from numbers on this scale.

The evidence about what happens when it is done badly comes from the same sector.

Why This Is Hard

The first difficulty is that most of what matters is a circumstance, not a transaction.

Whether a couple live together, whether a self-employed claimant earned what they declared, and whether savings are above a limit are facts about a person’s life that no single data set records.

The second difficulty is that the labels are biased by whoever chose what to look at.

An agency knows the outcome only for claims it reviewed or investigated, and if those were chosen by caseworker suspicion or by an earlier model, a new model learns the same choices.

Only a random sample gives an unbiased picture, and it is small.

The third difficulty is fairness, which is a legal duty here and not an aspiration.

A model that refers one group more often than its confirmed fraud rate justifies is putting extra burden on that group, even when every final decision is made by a person.

The Rotterdam welfare risk model, as reported by Lighthouse Reports, used 315 inputs including age, gender, language skills, and neighbourhood, and the reporting found that it discriminated by ethnicity, age, gender, and parenthood, with young single mothers among those most likely to be flagged.

Amnesty International found that the Dutch tax authority’s risk model for childcare benefits used nationality as a risk factor, which it describes as discrimination and racial profiling.

The fourth difficulty is that a human in the loop is not automatically a safeguard.

Experiments with Dutch civil servants and members of the public found that people did not over-rely on algorithmic advice in general, but that, outside the civil servant sample, they followed advice more readily when it matched a stereotype about a group, whether the advice came from an algorithm or a human expert.

How The Solutions Fit Together

A random sample of claims, reviewed in depth every year, is the measurement everything else depends on.

It estimates the overpayment rate and its causes, and it provides the only labels that are not shaped by who was suspected.

Data matching against authoritative records is the defence wherever such a record exists, because it checks employee earnings, deaths, and other facts directly rather than estimating them.

A same-period figure from an authoritative source can adjust the award going forward with notice and a way to dispute it, but anything retrospective, inferred, or suspected is a question for the claimant and a caseworker, never a debt.

A risk score that ranks claims or requests for a caseworker to review supports both, for the causes no record covers, such as self-employed earnings and living together.

Fully automated overpayment and fraud decisions are described here as a solution not to use, because the clearest evidence in this field is about what they do when nobody checks them.

Implementation Considerations

Measure before building, using the random sample, so effort goes to the causes that cost the most.

The NAO reports that DWP’s targeted case reviews identified unreported changes of circumstances as a key source of error, and that DWP is introducing periodic redeclaration of Universal Credit claims in response, which is a change to the claim process rather than a model.

Hold each claim as it was assessed, so a review can tell what the agency knew at the time and whether an overpayment was claimant error or official error.

For any risk score, keep a random control group referred alongside the high-risk cases.

DWP’s fairness assessment for its Universal Credit advances model describes exactly this, and also that the caseworker receives no risk rating and is not told that the model made the referral.

Measure each group’s referral rate and the rate at which its referrals are confirmed, and act when they diverge.

DWP’s July 2025 assessment found that older claimants were referred more often while their referrals were less often correct, and that non-UK nationals were referred more often although their referred requests were declined at the same rate as UK nationals’ referred requests, and said it would retrain the model.

Its transparency record of December 2025 says the model has been retrained but needs further testing before the retrained version goes live.

It could measure only one protected characteristic, age, because it lacked data on the others, which is a limit any agency should expect.

ROI Considerations

The return should be counted per route, against the cost of reviewing the claims each route selects.

The NAO reports that DWP’s targeted case reviews of 1.15 million Universal Credit claims generated estimated savings of £581 million by March 2025, and that its advances model saved an estimated £4.4 million since May 2022.

Those figures measure different things at different scales, and neither is a full return on investment, but they show where the larger savings came from.

Savings should be reported net of decisions overturned on reconsideration or appeal, and alongside the delay imposed on legitimate claimants.

DWP reports that approved advance requests the model referred were paid a median of one day later than those approved automatically.

The costs of getting this wrong dwarf the savings.

The Australian Government agreed to pay back more than $1,000 million to more than 400,000 people after a court ruled the Robodebt scheme unlawful, and Michigan agreed a $20 million settlement with around 3,000 people its automated system falsely accused of unemployment fraud.

Security, Privacy, and Regulation

Every data set in this problem is personal data about people in financial need, and some of it, such as health and disability, is special category data.

Under the EU AI Act, a system used by or for a public authority to evaluate eligibility for essential public assistance benefits, or to grant, reduce, revoke, or reclaim them, is high-risk under Annex III, with obligations for risk management, data governance, human oversight, and logging.

Regulation (EU) 2026/1744, the Digital Omnibus on AI, set those obligations for Annex III systems to apply from 2 December 2027.

A high-risk system already in use by a public authority before then has until 2 August 2030 to comply, unless its design changes significantly in the meantime.

Keeping a caseworker in the loop does not take a system out of that category: an Annex III system that profiles people is always treated as high-risk, so a model that ranks claimants for review is high-risk even though a person decides.

The same Act prohibits social scoring that leads to unjustified or disproportionate detrimental treatment.

In Australia, the Government accepted or accepted in principle all 56 recommendations of the Royal Commission into the Robodebt Scheme, and accepted in full the recommendation to consider legislation for a consistent framework for automated decision-making, with a clear path to review, plain-language public explanation, and business rules and algorithms open to independent expert scrutiny.

In the United Kingdom, the Public Authorities (Fraud, Error and Recovery) Act 2025 lets DWP require banks to identify accounts receiving certain benefits that meet set eligibility indicators, and its code of practice, published in May 2026, says no decision about entitlement will be made automatically on that information alone and that a person makes any final decision affecting entitlement.

DWP publishes an algorithmic transparency record for its advances model, and the NAO notes the tension between that transparency and the risk of helping fraudsters adapt.

Where The Evidence Disagrees

DWP judges its advances model reasonable and proportionate, citing safeguards and a hit rate around three times that of a random control group, while the NAO notes that the published analysis shows older claimants and non-UK nationals being over-referred.

The size of that advantage also differs between DWP’s own documents, around three times in its July 2025 fairness assessment and two and a half times in the later transparency record.

Targeted case review found an incorrect payment in 20% of the Universal Credit claims it reviewed in 2024-25, below DWP’s expectation of 24% and below the 24% found in the random sample.

DWP told the NAO that differences in objectives and method limit how far the two can be compared, since the random sample counts errors as small as 10p while case review looks for higher-value errors, so it remains unproven from public figures that targeting beats chance by count, even if it may by value.

Amsterdam’s model wrongly flagged Dutch and non-Dutch applicants at the same rate in testing after its training data was reweighted, but in the live pilot it was more likely to wrongly flag Dutch nationals and women, and no better than caseworkers at finding the applications that warranted scrutiny, which suggests that fairness measured before deployment does not settle the question.

Symptoms

Root causes

Business impact

Traditional approaches

Solution approaches

Each of these is written for this problem specifically, not borrowed from a generic catalogue.

Recommendation

Start by measuring the overpayment rate and its causes from a random sample of claims, and use the results to direct effort. Fix the largest causes with data matching against authoritative records, periodic redeclaration, and better claim design, treating every mismatch as a question for the claimant and a caseworker, never as a debt. Add a risk score only to rank claims or requests for a caseworker to review, with a random control group, a referral that does not show the caseworker the score, and published fairness analysis by group and outcome. A caseworker decides every refusal, suspension, overpayment, and fraud finding, and the claimant can always ask for the decision to be reconsidered and appeal it.

Data requirements

Claim and award records high sensitivity
Each claim with its declared circumstances, household members, assessed entitlement, payments made, and every change reported, kept as it was at each assessment so a later review can see what the agency knew at the time.
Third-party data matches high sensitivity
Employer-reported earnings, tax records, pension and benefit records from other agencies, and death, prison, and immigration status records, with the source, date, and period each figure covers.
Random sample review results high sensitivity
The findings of in-depth reviews of a random sample of claims, with the overpayment found, its cause, and whether it was claimant fraud, claimant error, or official error, used to estimate the rate and to provide unbiased labels.
Case review and investigation outcomes high sensitivity
The result of every targeted review and investigation, including the referral source and those that found the claim correct, needed to measure hit rates and to avoid training on only the cases somebody chose to examine.
Protected characteristics for fairness monitoring high sensitivity
Age, sex, nationality, disability, and where lawful ethnicity, held separately from any model's inputs and used only to measure whether referrals and outcomes differ between groups. Often incomplete, which limits what can be measured.

Evaluation metrics

Human oversight

A caseworker decides every refusal, suspension, change to an award, and overpayment, and a counter-fraud investigator and the prosecuting authority decide any fraud allegation. A data match or a risk score only selects a claim for a caseworker to look at, and it is never itself the evidence. The claimant is told what the agency believes has changed and can respond before a debt is raised. The department's senior leaders own the decision to deploy any risk model, review its fairness analysis at least yearly, and publish what it does.

Common pitfalls

When not to use AI

Sources

Practitioners

Research

Case studies

Regulation and guidance

Agent context

A distilled brief for an AI agent or an engineer scoping this work. It carries the constraints and pitfalls that change an implementation decision, not the whole page.

Get the agent context as Markdown

Objective

Reduce benefit overpayments from fraud and error by measuring their causes, verifying circumstances against authoritative records, and directing caseworker review to the riskiest claims, without delaying, refusing, or accusing legitimate claimants and without any automated decision that takes money from a claimant.

Constraints

Inputs

Outputs

Success criteria

Last updated 2026-10-03