Accountable AI Is an Engineering Requirement, Not a Compliance Document

The framing problem

Most organisations encounter responsible artificial intelligence as a policy document. It is written by a governance function, approved by a committee, filed, and then has almost no contact with the engineering team building the system it describes. When something goes wrong, the document is produced as evidence of intent and everybody notices it changed nothing.

Ivan Teh's position has consistently been that transparency, fairness and accountability belong in the system specification. They are properties the software either has or does not have, testable in the same way latency and uptime are testable, and owned by a named person in the same way any other requirement is owned.

Three properties with owners

Explainability. For any decision the system influences, somebody must be able to state which inputs drove the outcome, in language the affected party would understand. This is a design constraint that shapes model selection. If a marginally more accurate model cannot be explained to a customer whose loan it declined, the accuracy gain may not be worth having.

Fairness. Outcomes must be measured across the groups the system acts on, not assumed to be neutral because no protected attribute was passed as an input. Proxy variables do the work whether or not anyone intended it. This requires a monitoring apparatus and a defined threshold at which someone intervenes.

Auditability. The system must retain enough of its own history that a past decision can be reconstructed. That means versioned models, versioned training data and logged inputs. Retrofitting this after a regulator asks is expensive and frequently impossible.

What this looks like in production

Concretely, a deployment built this way has a few visible features. There is a model card that states what the system does, what it does not do and where it is known to be weak. There is a monitoring dashboard that tracks outcome distribution across segments, not only aggregate accuracy. There is a defined retraining trigger, because a model left alone decays quietly as the world moves underneath it. And there is a named owner for each of those, who is a person rather than a committee.

None of this is exotic. It resembles the operational discipline any organisation already applies to systems that handle money. The novelty is only that machine learning systems were, for a period, exempted from it on the grounds that they were experimental.

The commercial argument

The ethical case for accountable systems is well rehearsed. The commercial case gets less attention and is more likely to move a board.

An unexplainable model is a model the risk function will not sign off, which means it does not reach production, which means the investment produced nothing. A model that cannot be audited is a liability the moment a regulator takes an interest. And a model whose behaviour cannot be described to customers erodes the trust that makes the rest of the relationship work.

Teh has framed this as a condition of durability rather than a constraint on it. Systems built to be accountable survive scrutiny. Systems built for accuracy alone tend to meet scrutiny once and not recover.