Why public-sector AI has to be deterministic — and leave nothing to chance
If the previous piece was about the danger of letting an AI improvise, this is about the alternative. It is not “a better model.” It is determinism: the same input producing the same, correct, explainable outcome — every single time.
Determinism is an unglamorous word for a simple promise: nothing important is left to chance. In a well-built public-sector AI, the parts that matter — what gets promised, what gets dispatched, how something is escalated, what a fee is — are not generated fresh each time by a model in a slightly different mood. They are produced by rules the organisation owns, can read, can test, and can stand behind in front of an auditor or a journalist.
One decision, one owner
The principle I keep coming back to is that every load-bearing decision should have exactly one owner, and that owner should be deterministic code, not a language model. Who decides whether this is an emergency? One component, with one rule. Who decides what the caller is promised? One owner. When two parts of a system can each answer the same question differently, that ambiguity is the bug — and in a public service it is the bug that eventually pages the wrong officer or promises the wrong thing. Collapsing every important decision to a single, inspectable owner is what makes the whole system predictable.
The model is a suggestion, never a contract
This does not mean throwing the language model away — it is extraordinary at understanding messy human speech and responding with warmth. It means putting it in its proper place. The model proposes; the deterministic system disposes. Its wording is a suggestion the system is free to accept, edit or reject against hard rules — no banned promises, no invented facts, no unsafe actions. The caller experiences a natural conversation; the council gets an outcome that was never actually left to the model to choose.
Determinism is what makes accountability possible
Government decisions get audited, questioned and debated in public, and they should. An AI operating in that environment has to be reconstructable: what was said, what was decided, and why. Deterministic systems give you that for free, because the same inputs always explain the same outputs. A system that improvises cannot truly be audited — you can only ever explain what it did this time, not what it will do next. Byte-for-byte predictability is not a technical nicety; it is the thing that lets a council put its name to an automated interaction with a straight face.
The irony is that determinism is often dismissed as the boring, old-fashioned choice next to the magic of generative AI. It is exactly the reverse. The magic is easy and cheap now. The discipline to make sure the magic never touches the decisions a resident depends on — that is the hard part, and it is the whole job.