When a system decides: minimum safeguards for artificial intelligence in public administration
We surveyed 148 algorithmic systems in use across 24 jurisdictions. Fewer than one in five document how they were trained, and only eleven let the affected person learn the criteria behind the decision.
Key findings
- 01Only 27 of the 148 systems surveyed have accessible technical documentation on training data and decision criteria.
- 0262% were procured through direct award, with no auditability requirements in the tender documents.
- 03Eleven systems offer the affected person an individual explanation of the decision; none provides human review with a defined deadline.
- 04We propose a mandatory public register, auditability clauses in procurement, and effective human review for decisions affecting rights.
Debate about artificial intelligence in government has been organised around general principles — transparency, fairness, accountability — that almost nobody disputes and that almost never translate into concrete obligations. This brief proposes the reverse move: start from the systems already making decisions and ask which minimum safeguards are missing.
Between September 2025 and June 2026 we surveyed national and provincial agencies through freedom-of-information requests, analysis of procurement documents and technical interviews. We identified 148 systems in operation that intervene in administrative decisions: benefit allocation, fraud detection, health triage, traffic management and public service desks.
Figure 01
Algorithmic systems in use across public administration
The problem is not the model, it is the file
Most of the systems surveyed are not frontier models: they are rules, decision trees and relatively simple statistical models. What makes them problematic is not their sophistication but their administrative opacity. An administrative act must be reasoned; when the reasoning refers to a score produced by a system whose operation nobody in the agency can explain, the guarantee of reasoned decision is hollowed out.
“We are not asking the state to explain the mathematics. We are asking it to be able to explain the decision.”
Four minimum safeguards
- 01A mandatory public register of every system intervening in administrative decisions, stating purpose, responsible agency, data used and date of last evaluation.
- 02Auditability clauses in procurement: access for the agency and an external auditor to the model, the training data and disaggregated performance metrics.
- 03Accessible individual explanation: the affected person must be able to learn the factors that determined the decision in their case.
- 04Effective human review, with a maximum deadline, defined competence and genuine capacity to depart from the system's output.
Methodology
The survey covers 24 subnational jurisdictions and 11 national agencies. Office tools, document management systems without a decision component and general-purpose commercial software were excluded. The response rate to information requests was 71%.
References
- Ferrari, L. and Linares, S. (2026). “When a system decides”. Policy Brief No. 19.
Related research
Buying technology: tenders, lock-in and the buyer's technical capacity
Facial recognition in public space: three years of case files
Competitive hiring in the state: evidence on recruitment in the national administration
