Image Source: GOV.UK, Open Government License v3.0
The National Audit Office (NAO) has highlighted possible risks when the Department of Work and Pensions (DWP) uses machine learning to identify fraud and error in benefit claims.
The ministry raised the issue in its latest report on accounting, which includes a section on how the technology is used.
This is part of a three-year £70million spend on advanced analytics, which aims to save around £6.1bn by 2030-31.
According to the report, DWP will begin using a machine learning model in 2021-2022, with the algorithm trained on historical claimant data and fraud referrals for universal credit advances to eliminate fraud. It says it can flag new charges that may contain errors or errors. Similar models are being piloted for other features of universal credit, such as people living together, self-employment, capital and housing.
Claims identified as potentially fraudulent or erroneous will be referred to a caseworker, who will conduct a manual review. There is no automated decision-making and there is some evidence of bias against older claimants, so fairness analysis is done weekly.
inherent risk
NAO notes that machine learning has an inherent risk of biasing algorithms toward selecting claims from groups of people with protected characteristics, which may be due to the design of the model or the data used. points out that there is
DWP says it faces the challenge of balancing transparency about its use of technology and not tipping off fraudsters. But you need to be able to provide assurance that no group of customers has been wronged.
The ministry also acknowledges that it has taken safeguards to assess the impact of model use on various claimants and has established strict governance and controls. However, the ability to test undue influence across protected properties is currently limited. One reason is that petitioners do not always answer arbitrary demographic questions.
Additionally, personal data is segregated in our analytics platform for security reasons, which will be incorporated soon.
There are also plans to make the fairness analysis more comprehensive.
