AI in UK Public Services: Beyond Administrative Efficiency by Camille Chichet
This guest blog was authored by a DISC Virtual Visiting Fellow Camille Chichet, a public policy professional with experience of working within the legislative and executive branches of the French government. Her blog examines how AI in UK public administration is reshaping what the state notices, prioritises, and governs—and why transparency alone is not enough.
Public administrations are no longer merely digitising services. They are redesigning the architecture of administrative attention: which cases are seen first, which risks are made actionable, which users receive guidance, and which decisions become easier — or harder — to contest. Algorithmic systems now support core public functions: identifying fraud risk, prioritising applications, assisting officials, and mediating access to government information. Their appeal is evident. Public bodies face fiscal pressure, rising demand, and intense expectations of speed and consistency. Yet the central question is not whether AI makes administration more efficient. It is how it changes the way citizens become visible to the state.
The United Kingdom offers an unusually instructive case. Its Algorithmic Transparency Recording Standard requires central government departments, and certain public bodies delivering frontline services, to publish information on algorithmic tools that significantly influence decisions with public effect or directly interact with the public. That is a serious institutional development: visibility is a precondition for democratic scrutiny. But transparency alone is not governance. Once an algorithm enters an administrative process, the relevant question is not only whether the public knows it exists. It is what the system makes salient, what it leaves aside, and how it reallocates administrative attention.
The Department for Work and Pensions’ Universal Credit Advances Model illustrates the point. The model is a supervised machine-learning classifier used to assess the fraud risk of some Universal Credit advance applications before payment. In 2024–2025, the DWP issued 1.4 million Universal Credit advances, worth £0.8 billion, with estimated fraud and error on those advances ranging from £0 to £60 million. The department presents the model as a way to focus checks on higher-risk applications while limiting unnecessary intervention for legitimate claimants; it also states that payment decisions remain with human officials and that fairness assessments are conducted.
The case is important precisely because it is not an obvious caricature of automated government. Targeted review may be preferable to blanket control. The difficulty lies in the administrative translation: an application for urgent support becomes a calculable risk object. Features of a claimant’s administrative life are converted into signals, and those signals shape which cases are escalated, checked, delayed, or treated as routine. Even where human decision-making remains formally intact, the first ordering of attention is already partly structured by the system.
GOV.UK Chat shows a different form of the same transformation. Here, AI is not used to assess risk, but to mediate access to public information. Public pilots involved more than 10,000 users asking 26,000 questions about government services, including tax, benefits and visas; reported answer accuracy increased from 76% in early benchmarks to 90% in later testing.
The tool does not decide entitlement, and it directs users back to official sources. Still, public information is never merely informational. It shapes whether someone applies, delays, abandons a procedure, misunderstands an obligation, or identifies a route of redress. A conversational interface may reduce friction; it may also become a new layer through which citizens encounter the law.
High-stakes domains make these concerns sharper. Privacy International’s 2025 complaint about Home Office tools used in immigration operations — including Identify and Prioritise Immigration Cases and the Electronic Monitoring Review Tool — raises concerns about migrants’ personal data, limited information, and the adequacy of human involvement in contexts linked to enforcement and electronic monitoring.
These are not marginal questions of interface design. They concern the distribution of procedural power.
The debate on public-sector AI should therefore move beyond the familiar language of bias. Bias remains important, but it is only one expression of a broader institutional transformation. Algorithmic systems shape what government notices, how it sequences work, which cases become administratively urgent, and which citizens must bear the cost of correction.
A serious governance framework should rest on four requirements. First, visibility: the public should know which systems are used, by whom, and for what purpose. Second, contestability: where a system affects access to a service, benefit, procedure or enforcement action, citizens must be able to understand and challenge its role. Third, social evaluation: audits should assess not only accuracy, but also administrative burden, non-take-up, discrimination, trust and access to rights. Fourth, institutional responsibility: human oversight must be trained, documented, resourced and capable of departing from system outputs. The former CDEI already argued for mandatory transparency where algorithms significantly affect decisions concerning individuals.
The issue, then, is not the adoption of AI in itself, but the conditions under which it is integrated into public services. A system that accelerates case processing while making explanation, appeal or correction harder to access cannot be treated as a routine administrative improvement. In the welfare state, efficiency has political value only when it also strengthens citizens’ capacity to claim and defend their rights.
This blog is part of a series of articles shared with us by the Digital Society Research Group (DISC) written by Virtual Fellows engaging in critical approaches to AI policy.
