Page 67 - Reforming Benefits Decision-Making -(updated - August 2021)

as the above shows, automated systems do not always produce fair or
reasonable outcomes, particularly where they encounter situations which have
not been thought about in their programming. Automated systems can also
contain inbuilt direct and indirect discrimination within their programming.
186
Without transparency regarding the development and operation of automated
systems accountability and remedy for these errors and discrimination is not
possible.

2.87 As Richard Pope has pointed out, currently, the UC system is very one sided–
it collects large amounts of personal data and has a detailed, real-time view of
how the public are using the service. However, those wishing to hold the
government to account for its actions have little information to go on. Being
transparent about how DWP’s systems work and how they change will enable
effective scrutiny of UC, which will help contribute to public support and
confidence in the system.
187

2.88 A number of consultees we spoke to raised questions or concerns about what
the Intelligent Automation Garage is doing. There is very little information
about it in the public domain. The DWP told us that to date it has built and
deployed 50 automations. These are applied to mundane processes and

v Post Office Limited [2021] EWCA Crim 577 (see also K. Peachy ‘Post Office scandal: What the
Horizon saga is all about’ (BBC, 23 April 2021). More than 20,000 parents were falsely accused of
childcare benefit fraud by an automated fraud detection system in the Netherlands (‘A benefits scandal
sinks the Dutch government’ (The Economist, 21 January 2021). In Ontario, Canada a predictive
analytics programme which supplemented case workers’ assessments of eligibility and benefits level,
was found to have had at the time of its launch 2,400 serious defects, which impacted clients’ eligibility
for benefits and the payments they received. Ministry of Community and Social Science, ‘SAMS –
Social Assistance Management System’ in 2015 Annual Report of the Office of the Auditor General of
Ontario (2015).
186 For example, the risk factors that the algorithm used in a childcare benefits-fraud detection system
in the Netherlands, included parents with dual-nationality as a fraud risk, which amounted to ethnic
profiling. See Dutch Parliamentary Inquiry Committee, Unprecedent Injustice (December 2020). The
data used to train an AI system may lead to implicit biases, or the parameters which have limited
adaptability to individual situations can result in indirect discrimination. This concern was raised in
relation to an authentication algorithm piloted in California, since the questions risked creating
additional obstacles for marginalised and migrant populations, for example by assuming established

residential ties. See Coalition of California Welfare Rights Organisations, Inc., ‘Advocate Response to
DSS Options for Replacing the Statewide Fingerprint Imaging System (SFIS)” (2017).
187 R. Pope, Universal Credit: Digital Welfare (see n. 56 above) p. 100-101.

58