Why the argument against AI is aimed at the wrong surface — and what a governed boundary actually changes.
The standard case against putting AI near anything that matters goes like this: the model is a probabilistic system, it guesses, and you cannot certify a guess. Regulated institutions have heard this so many times it now reads as settled. It is not settled. It is aimed at the wrong surface.
The problem was never that a model reasons probabilistically. Human underwriters reason probabilistically. Actuarial tables are probability. A trading desk, a triage nurse, a claims adjuster — all of them run an inference under uncertainty and then act. We do not demand that they be deterministic. We demand that their actions be bounded, attributable, and reversible where the stakes require it. We govern the boundary, not the neurons.
So the interesting question is not "can we make the model deterministic?" We can't, and we shouldn't want to — determinism inside is just a worse model. The question is: what intercepts the probabilistic output before it becomes an effect in the world?
For most deployed AI, the answer is nothing. The guess is the action.
The unbounded surface
An AI wrapper takes a probabilistic output and wires it straight to an effect — a database write, a payment, an email to a customer, a filing. There is no layer between the inference and the consequence. When people say AI is "unpredictable," this is what they have actually observed: not that the model is uncertain, but that its uncertainty reaches all the way to the outcome with nothing in between.
That is an architecture problem, and architecture problems have architecture answers.
The failure mode is not a probabilistic system. It is an unbounded probabilistic surface — one that nothing intercepts. Close the surface and the same model becomes something you can stand behind.
Probabilistic inside, deterministic at the boundary
Put a control layer between the model and the world. Every action the model wants to take passes through it. The layer is deterministic: given the same authority, the same policy, and the same request, it admits or denies the same way every time. It does not reason. It checks.
- Is this action inside the scope of authority that was actually granted?
- Does the authority trace back to a human who accepted responsibility for it?
- Is the effect reversible, or does it need a higher bar before it lands?
- Is every step of this recorded such that it can be replayed later?
The model can be as creative, as probabilistic, as large as you like. Its cleverness stays on the inside. What crosses the boundary is only ever an action that a deterministic checker has already found admissible. The guess proposes; the boundary disposes.
This is the whole move. It is not a filter bolted onto the output — a filter is still probabilistic, still guessing about whether the guess was safe. It is a structural boundary that the model cannot route around, because the boundary owns the effect surface and the model does not.
What this buys the person who signs
Institutions adopt AI because the demo was impressive. Of course they do — presuming rational actors is a mistake we leave to economics. But the demo only has to survive the meeting. The signature has to survive the incident. Somewhere above the enthusiasm is someone senior whose name goes on the deployment, and when the questions arrive, the demo will not be in the room. A bounded boundary is what makes that signature defensible:
- The action space is enumerable. You can state, in advance, the set of things the system is permitted to do — not hope, state.
- Every effect is attributable. It traces to a grant of authority, and the grant traces to a person.
- The record is replayable. When something is questioned months later, you re-run the decision from the evidence instead of reconstructing it from memory.
None of that requires the model to stop guessing. It requires the guess to land somewhere that a human already fenced.
We are not in the business of making the model more accurate — other people do that well, and we send them our respect and our traffic. We are in the business of making a probabilistic system's results bounded: creative on the inside, accountable at the edge. That is the surface the argument should have been about all along.
Reference invariants: ZTG-1 (Mechanistic Boundary), ZTG-3 (Governed Effect Surface).