A caseworker reviews application folders beside a laptop with an abstract AI network.

DWP Is Trialling AI to Pick Universal Credit Claims for Targeted Case Reviews

Key Takeaways

  • The UK's Department for Work and Pensions (DWP) has confirmed it is running a limited live trial of a machine learning model to help select Universal Credit claims for Targeted Case Reviews — the programme that reassesses more than 1 million benefit claims a year.
  • DWP Permanent Secretary Dame Sarah Healey confirmed the trial in a letter to Parliament's Public Accounts Committee. She stressed that "human consideration remains central": the model does not decide whether a review goes ahead, and cannot change entitlement, suspend payments or close a claim.
  • The debate is sharp because the existing programme is already contested: research by the Administrative Fairness Lab found that only 21% of targeted reviews identify an incorrect payment, meaning roughly four in five people reviewed are put through the process unnecessarily.
  • The model's selection criteria will remain confidential, the DWP says, to protect the integrity of the benefit system. Critics warn that secrecy, combined with the DWP's record on algorithmic bias, leaves claimants unable to challenge how they were chosen.
  • Searches for "dwp targeted case reviews ai" rose 2,900% in the past seven days, according to Google Trends — the signal that brought this story to the top of the week's AI news.

What Targeted Case Reviews are

Targeted Case Reviews, usually shortened to TCRs, are the DWP's programme for reassessing benefit claims to identify fraud and payment errors. The programme reviews more than one million claims every year, making it one of the largest administrative review operations in the British welfare system. A review is an intrusive process for the person on the receiving end: the DWP examines the claim in detail, asks for evidence, and can revisit entitlement.

The programme has been running for some time, and the DWP says it works. A department spokesperson said the targeted case review programme has already saved the taxpayer £1.7 billion by blocking incorrect Universal Credit payments. The wider anti-fraud initiative, of which TCRs are a central part, is expected to generate £13 billion in savings by the end of the decade (AF.net).

What is new is the use of machine learning to choose which claims get reviewed. Until now, cases were selected by staff using existing methods. The trial adds a machine learning model to the selection stage — a system that scores or ranks claims and helps staff decide which ones to investigate. The DWP is conducting what it calls a "limited live test" of the model, meaning it is running on real claims, with real people, at a limited scale, while the department assesses how it performs.

The confirmation came from the top of the department. Dame Sarah Healey, the DWP's Permanent Secretary, confirmed the trial in a letter to Parliament's Public Accounts Committee (PAC), the committee of MPs that scrutinises government spending. The letter followed the committee's examination of the DWP's fraud and error work, and of a National Audit Office report on the department's systems.

What the DWP says the model does — and does not do

Healey's letter is careful, and the care is the point. The model, she said, supports staff in deciding which cases to target for review. It does not determine whether a review proceeds. It does not determine benefit entitlement. It does not determine payment adjustments, suspensions or claim closures. It cannot independently initiate a review.

"Human consideration remains central" to the process, Healey wrote. The model assists selection; staff retain control over all decisions. The department says the trial operates under its existing safeguards and equality standards, and Healey said the DWP would "consider alternative options" if the technology fails to perform as expected.

There is one thing the DWP will not do, and it said so plainly: it will not disclose the model's selection criteria or other technical features. The criteria must remain confidential, Healey confirmed, to protect the "integrity of the benefit system" — the argument being that if the criteria were public, people intent on fraud could game them. That is a standard enforcement argument, and it is also the argument that makes the system hardest to scrutinise from the outside.

The DWP's framing is that this is decision support, not automated decision-making. The distinction matters legally and politically, because a system that makes decisions about people's benefits would face a different level of scrutiny — and different legal obligations — than a system that merely helps a human choose which file to open. The critics' argument is that the distinction is thinner than the DWP suggests, because a model that shapes which claims are investigated shapes who gets put through an intrusive process, even if a human signs off at the end.

Why the 21% figure matters

The trial would be controversial on its own. It is more controversial because of what is already known about the programme it feeds into. Research published by the Administrative Fairness Lab earlier in 2026, as reported by Beinsure, found that only 21% of targeted reviews identify an incorrect payment. That means roughly four out of five people selected for a review are found to have nothing wrong with their claim — yet they have still been through the process.

The researchers warned that hundreds of thousands of people are subjected to the process unnecessarily. The same research found that many claimants chosen for TCRs are repeatedly subject to multiple reviews — often only months apart — causing "feelings of criminalisation" and exacerbating anxiety. Being reviewed once is stressful; being reviewed repeatedly, with no finding of wrongdoing, is experienced by many claimants as being treated as a suspect by default.

That is the baseline the machine learning model is being added to. If the model improves the hit rate — if it picks claims more accurately than the current methods — then fewer innocent people would be reviewed, and the programme would become both fairer and more efficient. That is the DWP's case for the trial: advanced analytics can direct scarce civil service resources towards high-probability cases while leaving compliant claimants undisturbed.

But the 21% figure also sets the test the model has to pass. If the model's selections do not produce a meaningfully better hit rate than the current 21%, then the trial is adding an opaque, algorithmic layer to a programme that already puts most of the people it reviews through an unnecessary process. And if the model inherits the patterns of the historical data it was trained on, it may reproduce the same selection biases that produced the current figures — more efficiently, and harder to see.

The DWP has not published a hit rate for the model's selections, and the trial is at a limited scale. Whether the model improves on 21% is, at the time of writing, an open question rather than an answered one.

The bias question, and the DWP's record

The sharpest criticism of the trial comes from researchers who have studied the DWP's earlier use of AI. Mark Bennett, a law lecturer at the University of Liverpool who works with the Administrative Fairness Lab, warned that the deployment of machine learning marks a "potential step change" that "comes with clear risks".

Bennett points to a separate AI tool previously used by the DWP to screen Universal Credit advance payments, which disproportionately flagged older, disabled and non-UK claimants for fraud checks, as Newswav reports. "The use of machine learning as a fraud and error prevention tool in other areas of the welfare system was previously found to have shown bias in selecting people on the basis of age, disability, marital status, and nationality when identifying claims to investigate for possible fraud," he said. "There are serious questions to be asked, therefore, about the reliability of data used within the machine learning model and how, in using such technology, the DWP might avoid repeating previous mistakes."

That history is not a footnote. It is the reason the trial is being watched so closely. An earlier DWP predictive risk-profiling system was found to disproportionately flag certain claimant groups, and the concern is that a machine learning model trained on historical case data will learn the same patterns — including any patterns that reflect past practice rather than actual fraud risk.

The concern is not abstract. In evidence to the Public Accounts Committee, DWP officials have previously acknowledged that an AI system being developed to detect fraud and error was still in the learning stages of its development and was exhibiting bias against some claimants. The department's position is that the system was not ready for deployment, and that the bias was being addressed during development. The critics' position is that a system which exhibits bias in development should be treated with particular caution when it moves to a live trial on real claims.

There is a structural reason the bias question is hard to resolve. The DWP's argument for confidentiality — that the criteria must stay secret so fraudsters cannot game them — also means that independent researchers cannot examine the model's features, weights or training data. The department says the trial operates under existing safeguards and equality standards, but those safeguards are, by design, not open to external inspection. The result is a debate conducted partly in the dark: the DWP asks for trust, and the researchers ask for evidence.

What the trial means for claimants

For the roughly six million people on Universal Credit, the practical question is simple: does this change anything for me? The honest answer, at the time of writing, is that it changes the odds of being selected for a review, but no one outside the DWP can say how, or by how much.

What is known is the process. A targeted case review involves the DWP reassessing a claim, which can mean requests for documents, questions about circumstances, and a period of uncertainty while the review is carried out. The review can result in a finding of overpayment or underpayment, or in no change at all — and in roughly four out of five cases, according to the Administrative Fairness Lab, it results in no finding of an incorrect payment. The process itself is the burden, even when the outcome is clean.

The trial does not change what a review is. It changes how people are chosen for one. For a claimant, that distinction may feel thin: being chosen by a model you cannot see, on criteria you cannot know, for a review that will most likely find nothing wrong, is not obviously better than being chosen by a human using visible rules. The DWP's answer is that the model will, over time, choose better — fewer innocent people reviewed, more actual fraud caught. That is a testable claim, and the test is the trial.

There is also a question about what happens if the trial fails. Healey said the department would "consider alternative options" if the technology fails to perform as expected. That is a commitment to evaluate, not a commitment to stop, and it leaves open what "fails to perform" means — a lower hit rate, a bias finding, a public controversy, or something else. The Public Accounts Committee, which received the letter, is the body most likely to press for an answer, and its scrutiny will be the main public check on the trial while the model's criteria remain confidential.

AI in government: the wider pattern

The DWP trial is part of a wider pattern of AI entering British public services, and it is worth placing it in that context rather than treating it as an isolated story. The same week, our coverage of AI cameras detecting driver violations in the UK documented another system in which AI is being used to identify people for enforcement action — in that case, drivers. The two systems are different in almost every respect, but they share the same structure: an AI system flags a person, a human reviews the flag, and the person flagged has limited visibility into why they were chosen.

That structure is becoming the default for AI in government, and the DWP trial is one of its most consequential examples — because the people it affects are among the most vulnerable, because the consequences of being flagged are intrusive, and because the stakes of getting selection wrong are measured in anxiety and criminalisation, not just inconvenience. The contrast with other government AI uses is instructive: an AI system that helps doctors read scans or helps planners model traffic has a different risk profile from one that selects welfare claimants for investigation, because the cost of a false positive falls on a person who cannot opt out.

The DWP's investment in this area is substantial. The National Audit Office reported that the department is investing £70 million over the next two years in advanced analytics to deal with rising fraud and error, and the department believes the investment will generate savings to the taxpayer of around £1.6 billion by 2031. That figure sits alongside the wider £13 billion anti-fraud savings target for the end of the decade, and the two should not be confused: the £1.6 billion relates to the analytics investment as reported by the NAO, while the £13 billion is the broader initiative. Both figures are departmental estimates, and neither has been independently verified as an achieved outcome.

What to watch next

Several things will determine how this story develops, and they are worth watching in order.

First, the Public Accounts Committee. The committee has the letter, and it has the history: it previously questioned DWP officials about the AI system's bias and about whether the targeted case review programme was value for money, given that it was pulling up similar percentages of incorrect claims to random sampling. The committee's next session on the DWP's fraud and error work is the most likely venue for the trial to be examined in public.

Second, the hit rate. The single most informative number the DWP could publish is the proportion of the model's selections that identify an incorrect payment, compared with the 21% baseline. If the model beats that figure clearly, the DWP's case strengthens considerably. If it does not, the case for continuing the trial weakens. The department has not committed to publishing that figure, and the trial's "limited" scale means any early number will be small and noisy.

Third, the bias evidence. The Administrative Fairness Lab and other researchers will be looking for any sign that the model's selections disproportionately affect older, disabled or non-UK claimants, as the earlier tool did. That evidence, if it emerges, will be harder for the department to address while the criteria remain confidential — and it will be the strongest test of whether the safeguards Healey cited are working.

Fourth, the scale decision. A limited live test is, by definition, a step towards a decision about full deployment. The DWP has not said when that decision will be made, or what criteria it will use. The trial's existence means the decision is now on the agenda, and the public debate about it is already under way — as the 2,900% rise in searches for the topic suggests.

For a fuller picture of how AI is being deployed in UK public enforcement, see our piece on AI cameras detecting driver violations in the UK, and for the context of AI in public infrastructure, our explainer on AI in air traffic control.

What We Could Not Verify

  • How many claimants the live test covers. The DWP describes the trial as "limited", but has not published the number of claims involved, and this article does not estimate it.
  • Which features the model uses. The selection criteria and technical features are confidential by the DWP's decision. No claim about the model's inputs is made in this article.
  • The model's hit rate. No figure for the proportion of the model's selections that identify incorrect payments has been published. The 21% figure refers to the existing programme as a whole, not to the trial.
  • Whether the earlier bias findings apply to this model. The bias findings concern a separate DWP tool used to screen advance payments. This article does not claim the new model has the same bias, and the DWP says the trial operates under equality standards.
  • When a full rollout decision will be made. The DWP has not announced a timetable, and this article does not predict one.
  • The savings figures. The £1.7 billion saved, the £13 billion target and the £1.6 billion NAO figure are departmental estimates for different scopes. They are reported as stated, not independently verified here.

Sources: The Independent via Beinsure; Newswav; AF.net; Administrative Fairness Lab research (2026) via Beinsure and Newswav; DWP Permanent Secretary Dame Sarah Healey's letter to the Public Accounts Committee as reported by Beinsure, Newswav and AF.net; National Audit Office report via Tech Monitor and the AI Incident Database (Report 3409); Mark Bennett, University of Liverpool, as quoted by Newswav; Google Trends, Explore, past 7 days.

Similar Posts