Artificial intelligence (AI) is often presented as a way to make public services faster, cheaper and more accurate. However, new research from LSE argues that AI does not simply help governments deliver existing policies more efficiently, it can also reshape how governments define social problems, decide who is deserving of support, and justify those decisions to the public.
The authors conducted an in-depth case study of Peru’s welfare-targeting system, SISFOH (Sistema de Focalización de Hogares—Household Targeting System in English). This system is used to assess household poverty and determine eligibility for social support.
The authors examined how this system collects and organises administrative data, turns that data into classifications, and uses algorithmic processes to support decisions about welfare eligibility. By focusing on a real public sector system, the paper shows how AI-enabled government works in practice, not just in theory.
Traditional public administration depends on balancing public benefit, political legitimacy and the capacity of government organisations to act. Drs Cordella and Gualdi argue that AI alters each part of this balance.
First, complex social realities must be translated into data. In the SISFOH case, poverty is represented through measurable indicators such as household assets, income records, electricity use, housing quality, and vehicle ownership. While these variables may be useful, they cannot capture every form of hardship. As a result, people whose circumstances do not fit the data model risk becoming less visible to the system.
Second, the system aggregates this information into categories such as extreme poor, poor and not poor. These classifications do more than measure poverty: they help define what poverty means for policy purposes. Once embedded in a database and algorithm, these definitions can become difficult to challenge, even if they leave out important dimensions of vulnerability.
Third, algorithmic computation changes the meaning of government capacity. Instead of relying on officials’ discretion and contextual judgement, decisions become increasingly shaped by automated rules and standardised processes.
The main finding is that AI can create what the authors call “algorithmic public value lock-in”. This happens when value judgements are built into data categories, indicators, and algorithms, making them appear neutral or technical while becoming harder to revise through democratic debate. SISFOH brought clear benefits, including faster verification, lower costs, and more consistent targeting of resources. However, it also embedded a narrow, measurable view of poverty that could overlook informal workers, urban hardship, and multidimensional forms of need.
The significance of these shortcomings became especially clear during the COVID-19 pandemic. SISFOH was designed around pre-pandemic assumptions about relatively stable socioeconomic conditions. When the crisis rapidly pushed many people into vulnerability, the system struggled to identify those newly in need, particularly people working informally. The Peruvian government had to move beyond the system’s classifications and introduce broader support measures. In this sense, the very features that made the system efficient in normal times made it less adaptable in an emergency.
This research challenges a narrow view of AI as a purely technical tool for improving public administration. It shows that decisions about data, categories, and algorithms are also decisions about values. For governments, the lesson is not that AI should be avoided, but that AI systems need to remain open to scrutiny, revision and democratic challenge. If they do not, they may make public services technically more efficient while also making them less responsive to the changing realities of people’s lives.
The research quoted in this article is from: Cordella, A., and F. Gualdi. 2026. “AI Public Value Creation: Data Encoding, Aggregation, and Algorithmic Computation.” Public Administration 1–21. https://doi.org/10.1111/padm.70071.
Wednesday 9 September 2026