Discussion

DWP is testing machine learning to select Universal Credit cases for review

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#14563

The Department for Work and Pensions is testing a machine-learning model to help staff select Universal Credit claims for targeted reviews. The trial puts AI-assisted case selection into a high-stakes public service, while leaving review and benefit decisions with DWP staff, according to the department’s permanent secretary.

Watch Desk analysis

What happened

The Independent reports that the DWP is running a “limited live test” of the model for its Targeted Case Reviews programme, which reassesses more than one million claims a year. Permanent secretary Dame Sarah Healey told Parliament’s Public Accounts Committee that the model supports staff in deciding which cases to target. She said it does not decide whether a review proceeds or change a claimant’s entitlement, payments or claim status.

The DWP says the trial follows its usual safeguards and equality standards, and that it will consider alternatives if the model does not deliver as expected. It has kept the criteria used to flag cases confidential, saying disclosure could affect the integrity of the benefit system. A DWP spokesperson told the Independent that staff make the selection decisions and that the department supports customers, particularly those with vulnerabilities.

Why it matters

Being selected for review is not the same as being found to have been paid incorrectly. But selection can still bring scrutiny and stress, and a model’s choices matter even when a person makes the final decision. The Independent reports that research by the Administrative Fairness Lab found only 21 per cent of reviews identified an incorrect payment, and that some claimants faced repeated reviews. That figure is the researchers’ reported finding, not a measure of this model’s performance.

The department’s assurances leave a practical question: how can people assess whether this trial is fair if its selection criteria remain confidential? Keeping details secret may protect the system from abuse, but it also makes meaningful scrutiny harder. Human involvement is important; it is not, on its own, evidence that a selection process is accurate or unbiased.

Our read

This is a concrete use of machine learning in decisions that can shape people’s experience of the benefits system, not just another government AI promise. The distinction between selecting a case for review and deciding someone’s entitlement deserves to stay clear. So does the need to explain how fairness is checked, what errors are found, and whether repeated or unnecessary reviews are reduced.

What to watch

  • Whether the DWP publishes results from the live test, including error and fairness measures.
  • How it will assess outcomes for claimants with different characteristics.
  • Whether the department changes or ends the trial if it fails to deliver as expected.

Discussion spark: Should the DWP disclose enough about its model’s case-selection criteria for the public to scrutinise fairness, even if it says secrecy helps protect the benefits system?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.