The Delegation
State.
When a public decision is made by a system, who answers for it?
The most dangerous sentence in public administration is not “the computer said no.” It is “nobody decided.”
Governments are under pressure to do more with less. They must process applications, detect fraud, allocate scarce resources, answer citizens and identify risk at a scale no human team can handle alone. AI can help. In many cases, it should.
But public power is different from private convenience. A music app can recommend the wrong song. A public system can delay a benefit, flag a family, prioritise an inspection, influence a border decision or shape who receives attention from the state. Efficiency is not enough when the outcome changes a person’s rights, livelihood or dignity.
The Quiet Expansion of Administrative Power
Algorithmic systems already support public functions across tax, finance, migration administration, fraud detection, employment screening and resource allocation. The OECD’s 2025 review of 200 government AI cases found use across core functions ranging from public services and justice to corruption and civil-service reform.
That is not evidence of a secret machine government. It is evidence that the delegation has already begun — usually in the unglamorous places where a queue is long, a budget is tight and a caseworker is asked to make too many judgments too quickly.
The mistake is to see this only as a technology purchase. It is a transfer of administrative power. And every transfer of power needs a matching transfer of accountability.
The Human in the Loop Is Not a Chair
“Human in the loop” has become a comfort phrase. It often means a person is technically present, but lacks time, authority, training or access to the information needed to challenge the system. That is not oversight. It is a rubber stamp with a job title.
Real human oversight has conditions. The reviewer must understand what the system is doing well enough to notice a failure. They must be allowed to depart from its recommendation. They must document why. And the citizen affected by the outcome must have a meaningful route to contest it.
No explanation, no authority to challenge, no practical route for appeal.
They can inspect, override, explain and be reviewed in turn.
The Right to Be Heard
Law has always understood that an official decision needs more than speed. The person affected must know the case against them, have a chance to respond and be able to seek review. These are not administrative luxuries. They are the mechanisms that keep power from becoming arbitrary.
Automated systems strain each of them. A score can be difficult to explain. A model can learn from a historical pattern without making its assumptions visible. A citizen can receive an outcome without knowing that a system shaped it, much less what evidence would change it.
European data-protection law already places safeguards around solely automated decisions with legal or similarly significant effects. The principle should travel further than the letter of one article: the more consequential the decision, the less acceptable it is for responsibility to be automated away.
The Registry Test
Before a public authority deploys an AI system that affects people, the public should be able to find a plain-language answer to five questions: what is the system for; who provides it; what data does it use; what decision can it influence; and how can a person challenge the result?
This is not a demand to publish source code for every tool. It is a demand to make public power legible. The OECD notes that public registries can give citizens a single place to see which AI systems are in use, their purpose and the jurisdictions they affect.
Scope: what can it recommend, rank or decide?
Accountability: which named authority owns the result?
Evidence: how was it tested for error, bias and failure?
Redress: how does a citizen challenge an outcome?
A State Is Not a Startup
A startup can launch, learn and pivot. A state has duties that do not disappear when an experiment goes badly. It holds records people cannot opt out of, exercises powers people cannot meaningfully refuse and makes mistakes that can follow someone for years.
That does not mean government should move slowly for the performance of caution. It means government should move deliberately where the cost of error is carried by the citizen rather than the vendor. A pilot needs a sunset clause. A model needs an audit trail. A high-impact decision needs an accountable human being who can be found.
Small states have an advantage here if they choose to use it. They can build systems close enough to the public for feedback to reach the people who make decisions. The point is not to become a frictionless laboratory. It is to become a state that can explain itself.
The Decision Cannot Be Delegated
AI can summarise a file. It can flag an anomaly. It can show a caseworker where to look. Those are real public benefits. But when the state exercises coercive power, someone must remain capable of giving reasons that a citizen can understand and contest.
“The system recommended it” is information. It is not a justification.