Economy

DWP is now using AI to pick whose Universal Credit gets checked

DWP targeted case reviews now use an AI model to help pick which Universal Credit claims are checked. What the trial does, who is affected and your rights.

The Department for Work and Pensions has confirmed it is using artificial intelligence to help decide which Universal Credit claimants face a targeted case review, the in-depth check of whether someone is being paid the right amount. The admission, made in a letter from the department’s permanent secretary to MPs and reported on Wednesday 18 November, explains why “DWP targeted case reviews AI” has been climbing the search charts. The short version for claimants: a machine learning model is now helping to choose who gets reviewed, but the department insists a human still makes every decision that affects your money.

Benefits letter and laptop on a kitchen table, illustrating DWP targeted case reviews using AI for Universal Credit

That reassurance is doing a good deal of work, and it is worth unpacking what is known, what is not, and what anyone opening a review letter this winter should do.

What is a DWP targeted case review?

Targeted case reviews, known to claimants as Universal Credit claim reviews, were introduced in 2022 to “identify and address incorrectness across Universal Credit claims”, according to PublicTechnology. They are not fraud investigations as such. A review can uncover an honest mistake by a claimant, an error by the department, or a change of circumstances nobody reported. The result can be an award going down, staying the same or, occasionally, going up.

The scale is considerable. Around 6,000 staff now assess more than one million claims a year. The department says the programme had saved roughly £1.7 billion by March 2026 by preventing overpayments or stopping ineligible claims, and the wider anti-fraud and error push is expected to save £13 billion by the end of the decade, The Independent reported.

What exactly is the AI doing?

In a letter to the Public Accounts Committee, Dame Sarah Healey, the DWP’s permanent secretary, described a “limited live test” of a machine learning model designed to help staff prioritise which claims to look at. In plain terms, the model flags claims that look more likely to be incorrect, and caseworkers then decide which to pick up.

Dame Sarah was explicit about the limits. The model, she wrote, does not decide whether a review goes ahead, what someone is entitled to, whether payments change or are suspended, whether a claim is closed, or the outcome of the review. “Human consideration remains central” to the process, she said. A DWP spokesperson added that “decisions on cases selected for targeted case review will always be made by a DWP member of staff”.

What the department will not say is how the model reaches its view. Dame Sarah said the risk indicators and features of the model must stay confidential to protect the “integrity of the benefit system”, on the reasonable grounds that publishing them would help fraudsters game the system. The DWP has also not disclosed how widely the model is being used or how many claims it will score.

Why are critics worried?

The concern is less about a computer sorting a queue than about who ends up at the front of it. Mark Bennett of the University of Liverpool and the Administrative Fairness Lab called the move a “potential step change” that “comes with clear risks”, and said there were “serious questions to be asked” about the reliability of the data feeding the model. He pointed to an earlier DWP tool that screened Universal Credit advance payments and, he said, disproportionately flagged older, disabled and non-UK claimants for fraud checks.

Research by the same group earlier this year found that only 21 per cent of reviews uncover any incorrect payment, meaning most people put through the process turn out to be entitled to what they were getting. Participants described anxiety and “feelings of criminalisation”, and some reported being reviewed repeatedly, months apart. Privacy campaigners at Big Brother Watch have long argued the department is not transparent enough about how people are flagged.

Who is affected?

Anyone on Universal Credit can be selected. Being chosen does not mean you are suspected of fraud, and the DWP says the trial does not change how it supports vulnerable claimants or those with complex needs. Existing safeguards, including taking individual circumstances into account and making reasonable adjustments, continue to apply.

There is also a change behind the scenes. Reviews are currently handled by a mix of DWP staff and staff from the contractor TP, formerly Teleperformance. The department has decided to end that contract “for convenience” and bring the work fully in-house, stressing that this is not because of any breach or performance problem.

What should claimants do if they are selected?

Welfare rights advisers, including the Child Poverty Action Group, have published guidance on the process. The practical points are straightforward:

  • Do not ignore the message. Reviews usually begin with a notification in your online journal. Failing to engage can lead to payments being suspended and, eventually, the claim being closed.
  • Gather evidence early. Expect to be asked about housing costs, savings, earnings, childcare, who lives with you and any health conditions.
  • Ask for adjustments if you need more time, a different format or help from someone else.
  • Challenge decisions you disagree with. You can ask for a mandatory reconsideration, and then appeal to an independent tribunal.
  • Get free advice from Citizens Advice or a local welfare rights service before you sign anything you are unsure about.

What happens next?

The DWP says it will monitor and evaluate the trial, may add safeguards or adjust it, and will “consider alternative options” if the model does not deliver. The Public Accounts Committee, which received the letter, is likely to press for evidence that the tool is accurate and fair across different groups of claimants.

For a government trying to save billions while keeping public trust, the tension is obvious: secrecy protects the system from abuse, but it also makes it harder for anyone outside the department to check the machine is not repeating old biases at scale. It is a familiar question this week, as the row over NHS staff accessing patient records also turns on who can see our data and why. For now, claimants should treat a review letter as an administrative process rather than an accusation, and answer it promptly.