Digital Welfare State edition 013
July 2026
——————————————————————————
Data for early intervention
The UK government (via the now defunct department DSIT) has released a set of data standards and frameworks designed to aid the identification of vulnerable people. It will be used by data and information systems to identify people who are in need of help but may not be in contact with relevant services, and to intervene early rather than waiting for people to reach crisis point. It draws on work by the SAVVI programme to define a set of public sector data standards to enable easier data sharing between services.
The description mentions multi-agency data sharing; this is important in things like risk of homelessness where there is no one single agency responsible for all the factors which might contribute to someone becoming homeless. By bringing together multiple data sources, the theory is that people who are likely to need support can be identified more easily.
Early intervention is widely regarded as a good thing, preventing serious harm and saving money. Finding the people and households who need it is obviously a challenge: if they are not (yet) on the radar of relevant agencies because their needs don’t (yet) meet the threshold for support, how do we know that they are on a trajectory to need help?
However, how these data standards are used, and by what sort of software or other tools is critical. They raise issues of privacy and consent: how are people informed about their data being used in this way, are they able to provide informed consent, or opt out?
They also might be used in systems which have a high risk of producing, or reproducing, bias, putting people and families in contact with services in ways which create harm, rather than avoiding it. This article that I shared back in January reveals how a predictive system, using data from multiple sources, was used to predict children and young people’s risk of things like gang involvement.
The Think Family database, in use in Bristol for over 10 years, is now only used to predict children who are at risk of becoming NEET, but it was previously used to predict the risk of criminal exploitation, using dozens of datasets from a range of public bodies. Children and young people, often from racialised backgrounds, were put into contact with the police and branded as potential criminals when they had done nothing wrong. Rather than providing access to supportive early intervention where it was really needed, the article describes the negative impact of being targeted.
So while identifying potentially vulnerable people and providing support before they reach crisis is no doubt a valid objective, these data standards and frameworks need to be used incredibly carefully, acknowledging the inherent risks, and providing transparency and autonomy to people whose data is being processed.
Things to read
An Effectiveness Assessment of the predictive model which screens Universal Credit claims for potential fraud has been published. You may be aware that the machine learning model had been shown to display bias against some cohorts, including non-UK nationals, and some age groups. The assessment acknowledges that the model is ‘not working as effectively’ as they would expect for some, and that it has been retrained to try to reduce the rate at which it wrongly flags these groups. It is being tested and evaluated to see if the retraining is effective. So, watch this space.
——————————————————————————-
Lots of digital bits and pieces in this update on the UK’s Get Britain Working programme, including the customer-facing AI chatbot I played with recently, the DWP Ask tool for staff which I mentioned a while ago, and lots around better use of data and data sharing to support jobseekers. DWP Ask is set to be rolled out to all work coaches by the end of this month.
——————————————————————————-
Interesting academic article on the underlying politics which shape digital welfare surveillance systems. Using Sweden as a case study, it analyses how assumptions and characterisations of lazy, dishonest and undeserving welfare claimants have shaped the way that digital welfare systems are designed. This is something I’ve been writing about in one of the book chapters I’ve been working on. Hopefully I can share more news about that in the next newsletter!
——————————————————————————-
An intriguing story to finish. The Argentinian government are setting up a digital twin system to test the impact of different social policies. It will be used, among other things, to trial policies relating to social welfare, employment and social security.
The system will combine data and other information from multiple sources to create a virtual ‘twin’ of Argentinian society, in which policy ideas can be deployed and the effects observed. The idea is that the impacts of policy decisions can be tested and understood in the safe ‘twin’ environment before the policy is actually rolled out in real life, allowing changes to be made before actual deployment.
Commentators, opposition politicians and civil society organisations quickly raised concerns about how the system will be put together and governed - where will the data come from? How will it be governed and managed? Are there privacy concerns for Argentinian citizens? Who will set up and run the system (there are rumours that Palantir are involved)? Ironically, rather a lot of policy questions seem not to have been answered before this policy was announced.
This article discusses the challenges around transparency, trust and governance in more detail.