Only one AI fraud model works at scale in the UK benefits system
Key Points
- The Universal Credit Advances Model is the DWP’s only fraud model at scale
- The model spots fraud risk 2.5 times better than random checks
- The DWP has seven algorithmic tools on the government transparency register
- Legacy IT confines each AI tool to a single system
- The DWP shelved at least six generative AI prototypes
The Universal Credit Advances Model is the only fraud and error machine-learning model the DWP runs at scale. The Centre for Social Justice (CSJ) set out the finding in State of Dependency, the first report from its Welfare 2030 inquiry.
The model risk-assesses Universal Credit advances before the DWP pays them out. This lets the department focus its checks on the advances most likely to involve fraud.
DWP performance figures show the model is around 2.5 times more effective at identifying fraud risk than a random sample of claims.
The DWP has recorded seven algorithmic tools on the government’s Algorithmic Transparency Recording Standard, including:
- The Universal Credit Advances Model, which scores advance requests for fraud risk
- The Child Maintenance Service’s Predictive Analytics Compliance tool, which helps caseworkers stop payment arrangements from breaking down
- The Urgent Journal Messages tool, which flags Universal Credit messages that may indicate a risk of harm
- The Whitemail Insights and Vulnerability Scanner, which reads around 25,000 scanned documents and letters each day
- The Conversational Platform, a phone assistant that uses speech recognition to route callers
- The CMG Return Letters tool, which pulls addresses and reference numbers from returned post
- The ESA Medical Matching tool, which maps conditions typed into a claim form against the DWP’s recognised list
The Conversational Platform is the only one of the seven tools that deals directly with claimants.
The CSJ said the department’s fragmented legacy IT confines each tool to the single system it runs on. The ESA Medical Matching tool only works on the ESA computer systems, and the Child Maintenance Service tool stays within that service’s own systems.
“Ultimately, there is little to gain from automating inefficiency,” the CSJ said.
Generative AI prototypes shelved
The DWP went to market in July 2024 for suppliers to help scale up its generative AI work. By then its AI and Innovation Directorate had already run a series of proof-of-concept experiments.
Documents released under freedom of information legislation six months later showed that the DWP had shelved at least six generative AI prototypes. These included tools to improve staff training and jobcentres, speed up disability payments and modernise communication systems.
Two of the shelved tools, Aigent and A-cubed, had appeared in the DWP’s 2023–24 annual report as examples of what the technology could achieve.
Officials said the main obstacle to moving from early testing to full use was making the products scalable, reliable and thoroughly tested.
The DWP’s Lighthouse Programme continues to explore how generative AI could support work coaches, software engineering and staff productivity. The programme builds small prototypes quickly and moves them into larger tests if early trials show savings.
The Tony Blair Institute estimated that AI could free up as much as 40% of DWP employees’ time. The institute put the productivity gain at close to £1 billion.
